diff --git a/THIRD-PARTY-NOTICES b/THIRD-PARTY-NOTICES
index bc8f3ff47..b46924a5a 100644
--- a/THIRD-PARTY-NOTICES
+++ b/THIRD-PARTY-NOTICES
@@ -50,6 +50,13 @@ Wan 2.1 / Wan 2.2      Apache-2.0   https://github.com/Wan-Video/Wan2.1
     Wan-Video team. Model weights are downloaded from upstream at runtime
     and are not redistributed in this repository.
 
+Real-ESRGAN            BSD-3-Clause https://github.com/xinntao/Real-ESRGAN
+BasicSR                Apache-2.0   https://github.com/XPixelGroup/BasicSR
+    The optional Real-ESRGAN integration implements RRDBNet and compact SRVGG
+    architecture definitions compatible with public Real-ESRGAN checkpoints.
+    Model weights are downloaded from upstream at runtime and are not
+    redistributed in this repository.
+
 ================================================================================
 Optional integration: integrations/omnidreams
 ================================================================================
@@ -100,6 +107,12 @@ aiohttp                Apache-2.0   https://github.com/aio-libs/aiohttp
 aiortc                 BSD-3-Clause https://github.com/aiortc/aiortc
 opencv-python-headless Apache-2.0   https://github.com/opencv/opencv-python
 
+================================================================================
+Optional integration: integrations/realesrgan
+================================================================================
+
+opencv-python-headless Apache-2.0   https://github.com/opencv/opencv-python
+
 ================================================================================
 Source-level redistributions
 ================================================================================
diff --git a/feed_uplift_video.sh b/feed_uplift_video.sh
new file mode 100755
index 000000000..d2ef3cb70
--- /dev/null
+++ b/feed_uplift_video.sh
@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -euo pipefail
+
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+usage() {
+    cat <<'EOF'
+Usage:
+  ./feed_uplift_video.sh /path/to/input.mp4 [flashdreams-feed-frames args...]
+
+Environment:
+  UPLIFT_SERVER        Server address. Default: 127.0.0.1:8090
+  UPLIFT_TARGET_FPS    Feed rate. Default: 30
+
+Examples:
+  ./feed_uplift_video.sh /tmp/clip.mp4
+  UPLIFT_SERVER=galaxy-ts4-108:8090 ./feed_uplift_video.sh /tmp/clip.mp4
+  ./feed_uplift_video.sh /tmp/clip.mp4 --no_pace
+  ./feed_uplift_video.sh /tmp/clip.mp4 --max_chunks 20
+EOF
+}
+
+if [[ $# -lt 1 || "${1:-}" == "-h" || "${1:-}" == "--help" ]]; then
+    usage
+    exit 0
+fi
+
+INPUT_VIDEO="$1"
+shift
+
+SERVER="${UPLIFT_SERVER:-${FLASHVSR_UPLIFT_SERVER:-127.0.0.1:8090}}"
+TARGET_FPS="${UPLIFT_TARGET_FPS:-${FLASHVSR_TARGET_FPS:-30}}"
+
+export PYTHONPATH="integrations/flashvsr:flashdreams${PYTHONPATH:+:${PYTHONPATH}}"
+
+echo "Feeding ${INPUT_VIDEO} continuously to uplift server at ${SERVER}"
+exec uv run --package flashdreams --extra serving flashdreams-feed-frames \
+    --server "${SERVER}" \
+    --input "${INPUT_VIDEO}" \
+    --target_fps "${TARGET_FPS}" \
+    "$@"
diff --git a/flashdreams/flashdreams/scripts/cli.py b/flashdreams/flashdreams/scripts/cli.py
index e6c69ae7f..70dc2ae6e 100644
--- a/flashdreams/flashdreams/scripts/cli.py
+++ b/flashdreams/flashdreams/scripts/cli.py
@@ -28,6 +28,7 @@
     flashdreams-run wan21-i2v-14b-480p --prompt "..." --image-path frame.png
     flashdreams-run --no-instantiate template-offline # resolve config only
     flashdreams-run wan21-t2v-1.3b-480p --postprocess.preset flashvsr-v1.1-sparse-2.0
+    flashdreams-run wan21-t2v-1.3b-480p --postprocess.preset realesrgan
 
     # Multi-GPU via context-parallelism (integration transformers auto-detect
     # CP size from the launcher's WORLD group). ``--no-python`` tells
diff --git a/flashdreams/flashdreams/serving/uplift/__init__.py b/flashdreams/flashdreams/serving/uplift/__init__.py
new file mode 100644
index 000000000..559cc7ac1
--- /dev/null
+++ b/flashdreams/flashdreams/serving/uplift/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Shared gRPC uplift protocol, client, and browser-viewer utilities."""
diff --git a/integrations/flashvsr/flashvsr/grpc/protos/__init__.py b/flashdreams/flashdreams/serving/uplift/protos/__init__.py
similarity index 90%
rename from integrations/flashvsr/flashvsr/grpc/protos/__init__.py
rename to flashdreams/flashdreams/serving/uplift/protos/__init__.py
index e8ae61670..b33e22d78 100644
--- a/integrations/flashvsr/flashvsr/grpc/protos/__init__.py
+++ b/flashdreams/flashdreams/serving/uplift/protos/__init__.py
@@ -13,5 +13,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Generated protobuf modules for FlashVSR gRPC serving."""
-
+"""Generated protobuf modules for FlashDreams video uplift serving."""
diff --git a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr.proto b/flashdreams/flashdreams/serving/uplift/protos/uplift.proto
similarity index 95%
rename from integrations/flashvsr/flashvsr/grpc/protos/flashvsr.proto
rename to flashdreams/flashdreams/serving/uplift/protos/uplift.proto
index 4efab6b28..4ff2ab456 100644
--- a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr.proto
+++ b/flashdreams/flashdreams/serving/uplift/protos/uplift.proto
@@ -15,12 +15,13 @@
 
 syntax = "proto3";
 
-package flashvsr;
+package flashdreams.uplift;
 
 // ---------------------------------------------------------------------------
-// FlashVSR gRPC Service
+// FlashDreams video uplift gRPC service
 //
-// Streaming video super-resolution (2× or 4×) using a diffusion-based model.
+// Streaming video super-resolution (2× or 4×) using any server-side model
+// implementation that speaks this protocol.
 // Two usage patterns:
 //
 //   Unary chunk-by-chunk flow (best for integration into existing pipelines):
@@ -36,7 +37,7 @@ package flashvsr;
 //     get_status        — liveness/readiness check
 // ---------------------------------------------------------------------------
 
-service FlashVSR {
+service VideoUplift {
   rpc get_status    (StatusRequest)         returns (StatusResponse);
   rpc start_session (StartSessionRequest)   returns (StartSessionResponse);
   rpc end_session   (EndSessionRequest)     returns (EndSessionResponse);
@@ -53,7 +54,7 @@ message StatusRequest {}
 message StatusResponse {
   bool            ready           = 1;  // true once the model is loaded
   string          device          = 2;  // e.g. "cuda:0"
-  string          model_name      = 3;  // e.g. "FlashVSR-v1.1"
+  string          model_name      = 3;  // e.g. "FlashVSR-v1.1" or "RealESRGAN/x2plus"
   repeated string active_sessions = 4;  // currently open session IDs
 }
 
diff --git a/flashdreams/flashdreams/serving/uplift/protos/uplift_pb2.py b/flashdreams/flashdreams/serving/uplift/protos/uplift_pb2.py
new file mode 100644
index 000000000..936b5a3a9
--- /dev/null
+++ b/flashdreams/flashdreams/serving/uplift/protos/uplift_pb2.py
@@ -0,0 +1,56 @@
+# -*- coding: utf-8 -*-
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# NO CHECKED-IN PROTOBUF GENCODE
+# source: flashdreams/serving/uplift/protos/uplift.proto
+# Protobuf Python Version: 6.31.1
+"""Generated protocol buffer code."""
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import descriptor_pool as _descriptor_pool
+from google.protobuf import runtime_version as _runtime_version
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf.internal import builder as _builder
+_runtime_version.ValidateProtobufRuntimeVersion(
+    _runtime_version.Domain.PUBLIC,
+    6,
+    31,
+    1,
+    '',
+    'flashdreams/serving/uplift/protos/uplift.proto'
+)
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+
+
+DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n.flashdreams/serving/uplift/protos/uplift.proto\x12\x12\x66lashdreams.uplift\"\x0f\n\rStatusRequest\"\\\n\x0eStatusResponse\x12\r\n\x05ready\x18\x01 \x01(\x08\x12\x0e\n\x06\x64\x65vice\x18\x02 \x01(\t\x12\x12\n\nmodel_name\x18\x03 \x01(\t\x12\x17\n\x0f\x61\x63tive_sessions\x18\x04 \x03(\t\"y\n\x13StartSessionRequest\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x14\n\x0cinput_height\x18\x02 \x01(\x05\x12\x13\n\x0binput_width\x18\x03 \x01(\x05\x12\r\n\x05scale\x18\x04 \x01(\x05\x12\x14\n\x0csparse_ratio\x18\x05 \x01(\x02\"J\n\x14StartSessionResponse\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x0f\n\x07success\x18\x02 \x01(\x08\x12\r\n\x05\x65rror\x18\x03 \x01(\t\"\'\n\x11\x45ndSessionRequest\x12\x12\n\nsession_id\x18\x01 \x01(\t\"%\n\x12\x45ndSessionResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\"\xbb\x02\n\x13UpscaleChunkRequest\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x14\n\x0cinput_height\x18\x02 \x01(\x05\x12\x13\n\x0binput_width\x18\x03 \x01(\x05\x12\r\n\x05scale\x18\x04 \x01(\x05\x12\x14\n\x0csparse_ratio\x18\x05 \x01(\x02\x12\x12\n\nframes_rgb\x18\x06 \x01(\x0c\x12\x12\n\nnum_frames\x18\x07 \x01(\x05\x12\x0e\n\x06height\x18\x08 \x01(\x05\x12\r\n\x05width\x18\t \x01(\x05\x12\x13\n\x0b\x63hunk_index\x18\n \x01(\x05\x12\x39\n\x0e\x66rame_encoding\x18\x0b \x01(\x0e\x32!.flashdreams.uplift.FrameEncoding\x12\x13\n\x0b\x66rames_jpeg\x18\x0c \x03(\x0c\x12\x14\n\x0c\x64isplay_only\x18\r \x01(\x08\"\xc1\x01\n\x14UpscaleChunkResponse\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x12\n\nframes_rgb\x18\x02 \x01(\x0c\x12\x12\n\nnum_frames\x18\x03 \x01(\x05\x12\x0e\n\x06height\x18\x04 \x01(\x05\x12\r\n\x05width\x18\x05 \x01(\x05\x12\x13\n\x0b\x63hunk_index\x18\x06 \x01(\x05\x12\x12\n\nelapsed_ms\x18\x07 \x01(\x02\x12\r\n\x05\x65rror\x18\x08 \x01(\t\x12\x16\n\x0e\x66rames_omitted\x18\t \x01(\x08*D\n\rFrameEncoding\x12\x1a\n\x16\x46RAME_ENCODING_RAW_RGB\x10\x00\x12\x17\n\x13\x46RAME_ENCODING_JPEG\x10\x01\x32\xf0\x03\n\x0bVideoUplift\x12S\n\nget_status\x12!.flashdreams.uplift.StatusRequest\x1a\".flashdreams.uplift.StatusResponse\x12\x62\n\rstart_session\x12\'.flashdreams.uplift.StartSessionRequest\x1a(.flashdreams.uplift.StartSessionResponse\x12\\\n\x0b\x65nd_session\x12%.flashdreams.uplift.EndSessionRequest\x1a&.flashdreams.uplift.EndSessionResponse\x12\x62\n\rupscale_chunk\x12\'.flashdreams.uplift.UpscaleChunkRequest\x1a(.flashdreams.uplift.UpscaleChunkResponse\x12\x66\n\rupscale_video\x12\'.flashdreams.uplift.UpscaleChunkRequest\x1a(.flashdreams.uplift.UpscaleChunkResponse(\x01\x30\x01\x62\x06proto3')
+
+_globals = globals()
+_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
+_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, 'flashdreams.serving.uplift.protos.uplift_pb2', _globals)
+if not _descriptor._USE_C_DESCRIPTORS:
+  DESCRIPTOR._loaded_options = None
+  _globals['_FRAMEENCODING']._serialized_start=974
+  _globals['_FRAMEENCODING']._serialized_end=1042
+  _globals['_STATUSREQUEST']._serialized_start=70
+  _globals['_STATUSREQUEST']._serialized_end=85
+  _globals['_STATUSRESPONSE']._serialized_start=87
+  _globals['_STATUSRESPONSE']._serialized_end=179
+  _globals['_STARTSESSIONREQUEST']._serialized_start=181
+  _globals['_STARTSESSIONREQUEST']._serialized_end=302
+  _globals['_STARTSESSIONRESPONSE']._serialized_start=304
+  _globals['_STARTSESSIONRESPONSE']._serialized_end=378
+  _globals['_ENDSESSIONREQUEST']._serialized_start=380
+  _globals['_ENDSESSIONREQUEST']._serialized_end=419
+  _globals['_ENDSESSIONRESPONSE']._serialized_start=421
+  _globals['_ENDSESSIONRESPONSE']._serialized_end=458
+  _globals['_UPSCALECHUNKREQUEST']._serialized_start=461
+  _globals['_UPSCALECHUNKREQUEST']._serialized_end=776
+  _globals['_UPSCALECHUNKRESPONSE']._serialized_start=779
+  _globals['_UPSCALECHUNKRESPONSE']._serialized_end=972
+  _globals['_VIDEOUPLIFT']._serialized_start=1045
+  _globals['_VIDEOUPLIFT']._serialized_end=1541
+# @@protoc_insertion_point(module_scope)
diff --git a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2.pyi b/flashdreams/flashdreams/serving/uplift/protos/uplift_pb2.pyi
similarity index 100%
rename from integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2.pyi
rename to flashdreams/flashdreams/serving/uplift/protos/uplift_pb2.pyi
diff --git a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2_grpc.py b/flashdreams/flashdreams/serving/uplift/protos/uplift_pb2_grpc.py
similarity index 63%
rename from integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2_grpc.py
rename to flashdreams/flashdreams/serving/uplift/protos/uplift_pb2_grpc.py
index d1d1f52de..246b7777f 100644
--- a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2_grpc.py
+++ b/flashdreams/flashdreams/serving/uplift/protos/uplift_pb2_grpc.py
@@ -5,7 +5,7 @@
 import grpc
 import warnings
 
-from flashvsr.grpc.protos import flashvsr_pb2 as flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2
+from flashdreams.serving.uplift.protos import uplift_pb2 as flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2
 
 GRPC_GENERATED_VERSION = '1.80.0'
 GRPC_VERSION = grpc.__version__
@@ -20,18 +20,19 @@
 if _version_not_supported:
     raise RuntimeError(
         f'The grpc package installed is at version {GRPC_VERSION},'
-        + ' but the generated code in flashvsr/grpc/protos/flashvsr_pb2_grpc.py depends on'
+        + ' but the generated code in flashdreams/serving/uplift/protos/uplift_pb2_grpc.py depends on'
         + f' grpcio>={GRPC_GENERATED_VERSION}.'
         + f' Please upgrade your grpc module to grpcio>={GRPC_GENERATED_VERSION}'
         + f' or downgrade your generated code using grpcio-tools<={GRPC_VERSION}.'
     )
 
 
-class FlashVSRStub(object):
+class VideoUpliftStub(object):
     """---------------------------------------------------------------------------
-    FlashVSR gRPC Service
+    FlashDreams video uplift gRPC service
 
-    Streaming video super-resolution (2× or 4×) using a diffusion-based model.
+    Streaming video super-resolution (2× or 4×) using any server-side model
+    implementation that speaks this protocol.
     Two usage patterns:
 
     Unary chunk-by-chunk flow (best for integration into existing pipelines):
@@ -56,37 +57,38 @@ def __init__(self, channel):
             channel: A grpc.Channel.
         """
         self.get_status = channel.unary_unary(
-                '/flashvsr.FlashVSR/get_status',
-                request_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StatusRequest.SerializeToString,
-                response_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StatusResponse.FromString,
+                '/flashdreams.uplift.VideoUplift/get_status',
+                request_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StatusRequest.SerializeToString,
+                response_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StatusResponse.FromString,
                 _registered_method=True)
         self.start_session = channel.unary_unary(
-                '/flashvsr.FlashVSR/start_session',
-                request_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StartSessionRequest.SerializeToString,
-                response_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StartSessionResponse.FromString,
+                '/flashdreams.uplift.VideoUplift/start_session',
+                request_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StartSessionRequest.SerializeToString,
+                response_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StartSessionResponse.FromString,
                 _registered_method=True)
         self.end_session = channel.unary_unary(
-                '/flashvsr.FlashVSR/end_session',
-                request_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.EndSessionRequest.SerializeToString,
-                response_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.EndSessionResponse.FromString,
+                '/flashdreams.uplift.VideoUplift/end_session',
+                request_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.EndSessionRequest.SerializeToString,
+                response_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.EndSessionResponse.FromString,
                 _registered_method=True)
         self.upscale_chunk = channel.unary_unary(
-                '/flashvsr.FlashVSR/upscale_chunk',
-                request_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkRequest.SerializeToString,
-                response_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkResponse.FromString,
+                '/flashdreams.uplift.VideoUplift/upscale_chunk',
+                request_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkRequest.SerializeToString,
+                response_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkResponse.FromString,
                 _registered_method=True)
         self.upscale_video = channel.stream_stream(
-                '/flashvsr.FlashVSR/upscale_video',
-                request_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkRequest.SerializeToString,
-                response_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkResponse.FromString,
+                '/flashdreams.uplift.VideoUplift/upscale_video',
+                request_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkRequest.SerializeToString,
+                response_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkResponse.FromString,
                 _registered_method=True)
 
 
-class FlashVSRServicer(object):
+class VideoUpliftServicer(object):
     """---------------------------------------------------------------------------
-    FlashVSR gRPC Service
+    FlashDreams video uplift gRPC service
 
-    Streaming video super-resolution (2× or 4×) using a diffusion-based model.
+    Streaming video super-resolution (2× or 4×) using any server-side model
+    implementation that speaks this protocol.
     Two usage patterns:
 
     Unary chunk-by-chunk flow (best for integration into existing pipelines):
@@ -135,46 +137,47 @@ def upscale_video(self, request_iterator, context):
         raise NotImplementedError('Method not implemented!')
 
 
-def add_FlashVSRServicer_to_server(servicer, server):
+def add_VideoUpliftServicer_to_server(servicer, server):
     rpc_method_handlers = {
             'get_status': grpc.unary_unary_rpc_method_handler(
                     servicer.get_status,
-                    request_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StatusRequest.FromString,
-                    response_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StatusResponse.SerializeToString,
+                    request_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StatusRequest.FromString,
+                    response_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StatusResponse.SerializeToString,
             ),
             'start_session': grpc.unary_unary_rpc_method_handler(
                     servicer.start_session,
-                    request_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StartSessionRequest.FromString,
-                    response_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StartSessionResponse.SerializeToString,
+                    request_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StartSessionRequest.FromString,
+                    response_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StartSessionResponse.SerializeToString,
             ),
             'end_session': grpc.unary_unary_rpc_method_handler(
                     servicer.end_session,
-                    request_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.EndSessionRequest.FromString,
-                    response_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.EndSessionResponse.SerializeToString,
+                    request_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.EndSessionRequest.FromString,
+                    response_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.EndSessionResponse.SerializeToString,
             ),
             'upscale_chunk': grpc.unary_unary_rpc_method_handler(
                     servicer.upscale_chunk,
-                    request_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkRequest.FromString,
-                    response_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkResponse.SerializeToString,
+                    request_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkRequest.FromString,
+                    response_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkResponse.SerializeToString,
             ),
             'upscale_video': grpc.stream_stream_rpc_method_handler(
                     servicer.upscale_video,
-                    request_deserializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkRequest.FromString,
-                    response_serializer=flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkResponse.SerializeToString,
+                    request_deserializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkRequest.FromString,
+                    response_serializer=flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkResponse.SerializeToString,
             ),
     }
     generic_handler = grpc.method_handlers_generic_handler(
-            'flashvsr.FlashVSR', rpc_method_handlers)
+            'flashdreams.uplift.VideoUplift', rpc_method_handlers)
     server.add_generic_rpc_handlers((generic_handler,))
-    server.add_registered_method_handlers('flashvsr.FlashVSR', rpc_method_handlers)
+    server.add_registered_method_handlers('flashdreams.uplift.VideoUplift', rpc_method_handlers)
 
 
  # This class is part of an EXPERIMENTAL API.
-class FlashVSR(object):
+class VideoUplift(object):
     """---------------------------------------------------------------------------
-    FlashVSR gRPC Service
+    FlashDreams video uplift gRPC service
 
-    Streaming video super-resolution (2× or 4×) using a diffusion-based model.
+    Streaming video super-resolution (2× or 4×) using any server-side model
+    implementation that speaks this protocol.
     Two usage patterns:
 
     Unary chunk-by-chunk flow (best for integration into existing pipelines):
@@ -206,9 +209,9 @@ def get_status(request,
         return grpc.experimental.unary_unary(
             request,
             target,
-            '/flashvsr.FlashVSR/get_status',
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StatusRequest.SerializeToString,
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StatusResponse.FromString,
+            '/flashdreams.uplift.VideoUplift/get_status',
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StatusRequest.SerializeToString,
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StatusResponse.FromString,
             options,
             channel_credentials,
             insecure,
@@ -233,9 +236,9 @@ def start_session(request,
         return grpc.experimental.unary_unary(
             request,
             target,
-            '/flashvsr.FlashVSR/start_session',
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StartSessionRequest.SerializeToString,
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.StartSessionResponse.FromString,
+            '/flashdreams.uplift.VideoUplift/start_session',
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StartSessionRequest.SerializeToString,
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.StartSessionResponse.FromString,
             options,
             channel_credentials,
             insecure,
@@ -260,9 +263,9 @@ def end_session(request,
         return grpc.experimental.unary_unary(
             request,
             target,
-            '/flashvsr.FlashVSR/end_session',
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.EndSessionRequest.SerializeToString,
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.EndSessionResponse.FromString,
+            '/flashdreams.uplift.VideoUplift/end_session',
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.EndSessionRequest.SerializeToString,
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.EndSessionResponse.FromString,
             options,
             channel_credentials,
             insecure,
@@ -287,9 +290,9 @@ def upscale_chunk(request,
         return grpc.experimental.unary_unary(
             request,
             target,
-            '/flashvsr.FlashVSR/upscale_chunk',
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkRequest.SerializeToString,
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkResponse.FromString,
+            '/flashdreams.uplift.VideoUplift/upscale_chunk',
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkRequest.SerializeToString,
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkResponse.FromString,
             options,
             channel_credentials,
             insecure,
@@ -314,9 +317,9 @@ def upscale_video(request_iterator,
         return grpc.experimental.stream_stream(
             request_iterator,
             target,
-            '/flashvsr.FlashVSR/upscale_video',
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkRequest.SerializeToString,
-            flashvsr_dot_grpc_dot_protos_dot_flashvsr__pb2.UpscaleChunkResponse.FromString,
+            '/flashdreams.uplift.VideoUplift/upscale_video',
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkRequest.SerializeToString,
+            flashdreams_dot_serving_dot_uplift_dot_protos_dot_uplift__pb2.UpscaleChunkResponse.FromString,
             options,
             channel_credentials,
             insecure,
diff --git a/integrations/flashvsr/flashvsr/grpc/streaming_view.py b/flashdreams/flashdreams/serving/uplift/streaming_view.py
similarity index 92%
rename from integrations/flashvsr/flashvsr/grpc/streaming_view.py
rename to flashdreams/flashdreams/serving/uplift/streaming_view.py
index eec54f48b..0b4964ad7 100644
--- a/integrations/flashvsr/flashvsr/grpc/streaming_view.py
+++ b/flashdreams/flashdreams/serving/uplift/streaming_view.py
@@ -13,13 +13,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""HTTP MJPEG viewer support for the FlashVSR gRPC server."""
+"""HTTP MJPEG viewer support for video uplift servers."""
 
 import io
 import queue
 import threading
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from http import HTTPStatus
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 
@@ -31,6 +31,7 @@
 DEFAULT_VIEWER_JPEG_QUALITY = 90
 DEFAULT_VIEWER_JPEG_BACKEND = "auto"
 DEFAULT_VIEWER_MAX_FPS = 60.0
+DEFAULT_VIEWER_PLAYBACK_FPS = 30.0
 DEFAULT_VIEWER_FRAME_STRIDE = 1
 
 _VIEWER_STOP = object()
@@ -40,6 +41,7 @@
 @dataclass
 class _ViewerPlaybackChunk:
     elapsed_ms: float
+    ready_at: float = field(default_factory=time.perf_counter)
     frames: np.ndarray | None = None
     jpegs: list[bytes] | None = None
 
@@ -154,6 +156,7 @@ def __init__(
         jpeg_backend: str,
         chunk_queue_depth: int,
         max_fps: float,
+        playback_fps: float,
         frame_stride: int,
     ) -> None:
         self.host = host
@@ -162,6 +165,7 @@ def __init__(
         self.jpeg_backend = jpeg_backend
         self.chunk_queue_depth = max(1, int(chunk_queue_depth))
         self.max_fps = float(max_fps)
+        self.playback_fps = float(playback_fps)
         self.frame_stride = max(1, int(frame_stride))
         self.original_hub = _MjpegFrameHub()
         self.upscaled_hub = _MjpegFrameHub()
@@ -207,7 +211,7 @@ def _send_index(self) -> None:
 <head>
   <meta charset="utf-8">
   <meta name="viewport" content="width=device-width, initial-scale=1">
-  <title>FlashVSR Stream</title>
+  <title>FlashDreams Uplift Stream</title>
   <style>
     html, body { margin: 0; min-height: 100%; background: #111; color: #eee; font-family: system-ui, sans-serif; }
     body { display: grid; place-items: stretch; }
@@ -231,7 +235,7 @@ def _send_index(self) -> None:
 <body>
   <main>
     <header>
-      <h1>FlashVSR stream viewer</h1>
+      <h1>FlashDreams uplift stream viewer</h1>
       <button id="screenshot" type="button">Screenshot</button>
     </header>
     <section id="viewer" class="grid">
@@ -379,10 +383,11 @@ def _playback_loop(self, channel: str, hub: _MjpegFrameHub) -> None:
             if frame_count == 0:
                 continue
 
-            elapsed_s = max(item.elapsed_ms / 1000.0, 1e-3)
-            frame_interval_s = elapsed_s / max(1, frame_count)
-            if self.max_fps > 0:
-                frame_interval_s = max(frame_interval_s, 1.0 / self.max_fps)
+            frame_interval_s = self._frame_interval_s(
+                item,
+                frame_count,
+                queued_chunks=playback_queue.qsize(),
+            )
 
             now = time.perf_counter()
             if next_frame_at < now:
@@ -400,6 +405,32 @@ def _playback_loop(self, channel: str, hub: _MjpegFrameHub) -> None:
                 hub.publish(jpeg)
                 next_frame_at += frame_interval_s
 
+    def _frame_interval_s(
+        self,
+        item: _ViewerPlaybackChunk,
+        frame_count: int,
+        *,
+        queued_chunks: int,
+    ) -> float:
+        if frame_count <= 0:
+            raise ValueError("frame_count must be positive")
+
+        if self.playback_fps > 0:
+            fps = self.playback_fps
+            if self.max_fps > fps and queued_chunks > 1:
+                ramp = min(
+                    1.0,
+                    (queued_chunks - 1) / max(1, self.chunk_queue_depth - 2),
+                )
+                fps += (self.max_fps - fps) * ramp
+            interval_s = 1.0 / fps
+        else:
+            interval_s = max(item.elapsed_ms / 1000.0, 1e-3) / frame_count
+
+        if self.max_fps > 0:
+            interval_s = max(interval_s, 1.0 / self.max_fps)
+        return interval_s
+
     def enqueue_original_chunk(self, frames: np.ndarray, elapsed_ms: float) -> None:
         item = _ViewerPlaybackChunk(
             elapsed_ms=max(1.0, float(elapsed_ms)),
diff --git a/integrations/flashvsr/flashvsr/grpc/uplift_client.py b/flashdreams/flashdreams/serving/uplift/uplift_client.py
similarity index 85%
rename from integrations/flashvsr/flashvsr/grpc/uplift_client.py
rename to flashdreams/flashdreams/serving/uplift/uplift_client.py
index 59d399f8d..9bee5458e 100644
--- a/integrations/flashvsr/flashvsr/grpc/uplift_client.py
+++ b/flashdreams/flashdreams/serving/uplift/uplift_client.py
@@ -14,22 +14,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Uplift client for the FlashVSR gRPC service.
+"""Uplift client for the FlashDreams video uplift gRPC service.
 
 Usage (from repo root):
-    uv run --no-sync python -m flashvsr.grpc.uplift_client --input clip.mp4 --output out.mp4
+    uv run --no-sync python -m flashdreams.serving.uplift.uplift_client --input clip.mp4 --output out.mp4
 
     # Use unary chunk flow instead of streaming:
-    uv run --no-sync python -m flashvsr.grpc.uplift_client --input clip.mp4 --output out.mp4 --unary
+    uv run --no-sync python -m flashdreams.serving.uplift.uplift_client --input clip.mp4 --output out.mp4 --unary
 
     # Connect to a remote server:
-    uv run --no-sync python -m flashvsr.grpc.uplift_client --server my-host:50051 --input ...
+    uv run --no-sync python -m flashdreams.serving.uplift.uplift_client --server my-host:50051 --input ...
 
     # Send JPEG inputs and rely on the server browser viewer for output:
-    uv run --no-sync python -m flashvsr.grpc.uplift_client --input clip.mp4 --input_format jpeg --display_only
+    uv run --no-sync python -m flashdreams.serving.uplift.uplift_client --input clip.mp4 --input_format jpeg --display_only
 
     # Live-ingest stress mode: loop one video as 8-frame chunks at 30 fps:
-    uv run --no-sync python -m flashvsr.grpc.uplift_client --continuous --input clip.mp4
+    uv run --no-sync python -m flashdreams.serving.uplift.uplift_client --continuous --input clip.mp4
 """
 
 import argparse
@@ -44,8 +44,8 @@
 import mediapy as media
 import numpy as np
 
-from flashvsr.grpc.protos import flashvsr_pb2 as pb2
-from flashvsr.grpc.protos import flashvsr_pb2_grpc as pb2_grpc
+from flashdreams.serving.uplift.protos import uplift_pb2 as pb2
+from flashdreams.serving.uplift.protos import uplift_pb2_grpc as pb2_grpc
 
 DEFAULT_SERVER = "localhost:50051"
 DEFAULT_MAX_MESSAGE_MB = 512
@@ -179,13 +179,31 @@ def circular_chunk(frames: np.ndarray, start: int, size: int) -> np.ndarray:
     return np.ascontiguousarray(frames[indexes])
 
 
+def continuous_loop_frame_count(
+    total_frames: int,
+    *,
+    chunk_frames: int,
+    loop_mode: str,
+) -> int:
+    if total_frames <= 0:
+        raise ValueError("total_frames must be positive")
+    if chunk_frames <= 0:
+        raise ValueError("chunk_frames must be positive")
+    if loop_mode == "frame" or total_frames <= chunk_frames:
+        return total_frames
+    if loop_mode != "chunk":
+        raise ValueError(f"unsupported loop_mode: {loop_mode}")
+    aligned_frames = total_frames - (total_frames % chunk_frames)
+    return aligned_frames if aligned_frames >= chunk_frames else total_frames
+
+
 # ---------------------------------------------------------------------------
 # Streaming flow: upscale_video
 # ---------------------------------------------------------------------------
 
 
 def upsample_stream(
-    stub: pb2_grpc.FlashVSRStub,
+    stub: pb2_grpc.VideoUpliftStub,
     video_np: np.ndarray,
     scale: int,
     sparse_ratio: float,
@@ -260,7 +278,7 @@ def request_iter():
 
 
 def upsample_unary(
-    stub: pb2_grpc.FlashVSRStub,
+    stub: pb2_grpc.VideoUpliftStub,
     video_np: np.ndarray,
     scale: int,
     sparse_ratio: float,
@@ -349,7 +367,7 @@ def upsample_unary(
 
 
 def run_file_client(argv: list[str] | None = None) -> None:
-    parser = argparse.ArgumentParser(description="FlashVSR gRPC test client")
+    parser = argparse.ArgumentParser(description="FlashDreams video uplift client")
     parser.add_argument(
         "--server",
         default=DEFAULT_SERVER,
@@ -417,7 +435,7 @@ def run_file_client(argv: list[str] | None = None) -> None:
         ("grpc.max_receive_message_length", max_bytes),
     ]
     channel = grpc.insecure_channel(args.server, options=channel_options)
-    stub = pb2_grpc.FlashVSRStub(channel)
+    stub = pb2_grpc.VideoUpliftStub(channel)
 
     # Health check
     print(f"Checking server at {args.server} ...")
@@ -488,7 +506,7 @@ def run_file_client(argv: list[str] | None = None) -> None:
 
 def run_continuous_client(argv: list[str] | None = None) -> None:
     parser = argparse.ArgumentParser(
-        description="Loop one video to FlashVSR as 8-frame live-ingest chunks"
+        description="Loop one video to an uplift server as live-ingest chunks"
     )
     parser.add_argument(
         "--server",
@@ -516,8 +534,7 @@ def run_continuous_client(argv: list[str] | None = None) -> None:
         type=float,
         default=30.0,
         help=(
-            "Ingress rate to simulate. Default 30 fps models production; "
-            "local FlashVSR may only consume around 7 fps."
+            "Ingress rate to simulate. Default 30 fps models production."
         ),
     )
     parser.add_argument(
@@ -529,7 +546,27 @@ def run_continuous_client(argv: list[str] | None = None) -> None:
         "--max_chunks",
         type=int,
         default=0,
-        help="Stop after this many 8-frame chunks. Default 0 means endless.",
+        help="Stop after this many chunks. Default 0 means endless.",
+    )
+    parser.add_argument(
+        "--continuous_chunk_frames",
+        type=int,
+        default=CONTINUOUS_CHUNK_FRAMES,
+        help=(
+            "Frames per continuous live-ingest request. FlashVSR servers commonly "
+            "use 8; frame-local upsamplers such as RealESRGAN can use 1 for "
+            "smoother viewer pacing."
+        ),
+    )
+    parser.add_argument(
+        "--loop_mode",
+        choices=["frame", "chunk"],
+        default="frame",
+        help=(
+            "How endless input looping wraps the source video. frame preserves all "
+            "frames and may wrap inside a request; chunk drops a tail "
+            "remainder so the loop restart lands on a request boundary."
+        ),
     )
     parser.add_argument(
         "--input_format",
@@ -571,6 +608,8 @@ def run_continuous_client(argv: list[str] | None = None) -> None:
         parser.error("--target_fps must be positive")
     if args.max_chunks < 0:
         parser.error("--max_chunks must be non-negative")
+    if args.continuous_chunk_frames <= 0:
+        parser.error("--continuous_chunk_frames must be positive")
     if args.report_every <= 0:
         parser.error("--report_every must be positive")
     if args.ingress_window_chunks < 2:
@@ -581,12 +620,24 @@ def run_continuous_client(argv: list[str] | None = None) -> None:
     frames = read_rgb_video(args.input)
     total_frames, height, width, _ = frames.shape
     source_fps = video_fps(args.input)
+    loop_frames = continuous_loop_frame_count(
+        total_frames,
+        chunk_frames=args.continuous_chunk_frames,
+        loop_mode=args.loop_mode,
+    )
+    loop_source = frames[:loop_frames]
     print(
         f"Loaded {args.input}: {total_frames} frames, {width}x{height}, "
         f"source_fps={source_fps:.2f}"
     )
+    if loop_frames != total_frames:
+        print(
+            f"Looping first {loop_frames} frames so the restart aligns with "
+            f"{args.continuous_chunk_frames}-frame request boundaries "
+            f"(dropped tail={total_frames - loop_frames})."
+        )
     print(
-        f"Streaming {CONTINUOUS_CHUNK_FRAMES}-frame chunks to {args.server}; "
+        f"Streaming {args.continuous_chunk_frames}-frame chunks to {args.server}; "
         f"target_ingress={args.target_fps:.2f} fps; "
         f"mode={'return_frames' if args.return_frames else 'display_only'}"
     )
@@ -599,7 +650,7 @@ def run_continuous_client(argv: list[str] | None = None) -> None:
             ("grpc.max_receive_message_length", max_bytes),
         ],
     )
-    stub = pb2_grpc.FlashVSRStub(channel)
+    stub = pb2_grpc.VideoUpliftStub(channel)
 
     try:
         status = stub.get_status(pb2.StatusRequest(), timeout=10)
@@ -615,7 +666,7 @@ def run_continuous_client(argv: list[str] | None = None) -> None:
     state = {"sent": 0, "received": 0}
     send_times: deque[float] = deque(maxlen=args.ingress_window_chunks)
     receive_times: deque[float] = deque(maxlen=args.ingress_window_chunks)
-    frame_interval = CONTINUOUS_CHUNK_FRAMES / args.target_fps
+    frame_interval = args.continuous_chunk_frames / args.target_fps
 
     def requests() -> Iterator[pb2.UpscaleChunkRequest]:
         next_send_at = time.perf_counter()
@@ -630,7 +681,11 @@ def requests() -> Iterator[pb2.UpscaleChunkRequest]:
                     next_send_at = now
                 next_send_at += frame_interval
 
-            chunk = circular_chunk(frames, source_pos, CONTINUOUS_CHUNK_FRAMES)
+            chunk = circular_chunk(
+                loop_source,
+                source_pos,
+                args.continuous_chunk_frames,
+            )
             req = build_chunk_request(
                 chunk_idx=chunk_idx,
                 frame_data=chunk,
@@ -644,14 +699,16 @@ def requests() -> Iterator[pb2.UpscaleChunkRequest]:
 
             send_times.append(time.perf_counter())
             state["sent"] += 1
-            source_pos = (source_pos + CONTINUOUS_CHUNK_FRAMES) % total_frames
+            source_pos = (source_pos + args.continuous_chunk_frames) % loop_frames
             chunk_idx += 1
             if state["sent"] % args.report_every == 0:
                 elapsed = max(send_times[-1] - send_times[0], 1e-6)
-                ingress_fps = (len(send_times) - 1) * CONTINUOUS_CHUNK_FRAMES / elapsed
+                ingress_fps = (
+                    (len(send_times) - 1) * args.continuous_chunk_frames / elapsed
+                )
                 print(
                     f"sent_chunks={state['sent']} "
-                    f"sent_frames={state['sent'] * CONTINUOUS_CHUNK_FRAMES} "
+                    f"sent_frames={state['sent'] * args.continuous_chunk_frames} "
                     f"{ANSI_GREEN}observed_ingress={ingress_fps:.2f} fps"
                     f"{ANSI_RESET} window_chunks={len(send_times)}"
                 )
@@ -673,7 +730,7 @@ def requests() -> Iterator[pb2.UpscaleChunkRequest]:
             if state["received"] % args.report_every == 0:
                 elapsed = max(receive_times[-1] - receive_times[0], 1e-6)
                 receive_fps = (
-                    (len(receive_times) - 1) * CONTINUOUS_CHUNK_FRAMES / elapsed
+                    (len(receive_times) - 1) * args.continuous_chunk_frames / elapsed
                 )
                 print(
                     f"received_chunks={state['received']} "
diff --git a/flashdreams/pyproject.toml b/flashdreams/pyproject.toml
index c7553c5e9..8bf1df105 100644
--- a/flashdreams/pyproject.toml
+++ b/flashdreams/pyproject.toml
@@ -69,6 +69,8 @@ dependencies = [
 # Mirrors nerfstudio's ``ns-train`` console script shape (one hyphenated
 # binary, tyro union over runners as direct subcommands).
 flashdreams-run = "flashdreams.scripts.cli:entrypoint"
+flashdreams-uplift-client = "flashdreams.serving.uplift.uplift_client:main"
+flashdreams-feed-frames = "flashdreams.serving.uplift.uplift_client:run_continuous_client"
 
 [project.urls]
 Homepage = "https://github.com/NVIDIA/flashdreams"
@@ -111,6 +113,10 @@ runners = [
 serving = [
     "aiohttp>=3.9",
     "aiortc>=1.9",
+    "grpcio>=1.80",
+    "mediapy>=1.1",
+    "Pillow>=10",
+    "protobuf>=7.35.0,<8",
 ]
 
 [tool.setuptools.dynamic]
diff --git a/integrations/flashvsr/README.md b/integrations/flashvsr/README.md
index 9a0ea644b..12e4d2cd3 100644
--- a/integrations/flashvsr/README.md
+++ b/integrations/flashvsr/README.md
@@ -114,11 +114,12 @@ sparse presets. Use smaller inputs, `--chunk-size 8`, fewer GPUs, or override
 `--pipeline.diffusion-model.transformer.use-cuda-graph False` if the dense run
 OOMs.
 
-## gRPC server and browser viewer
+## Uplift server and browser viewer
 
-The FlashVSR integration also ships a gRPC upsampler service that keeps the
-pipeline warm, accepts incoming frame chunks, and can publish an MJPEG browser
-viewer. The service supports unary chunks (`start_session` / `upscale_chunk` /
+The FlashVSR integration also ships an uplift upsampler service that keeps the
+pipeline warm and serves the shared `flashdreams.serving.uplift` gRPC protocol.
+It accepts incoming frame chunks and can publish an MJPEG browser viewer. The
+service supports unary chunks (`start_session` / `upscale_chunk` /
 `end_session`) and bidirectional streaming (`upscale_video`). Streaming clients
 may send 8-frame chunks at live-ingest cadence; the server coalesces those into
 FlashVSR-compatible 13-frame cold-start and 16-frame steady-state model calls.
@@ -145,16 +146,12 @@ full attention otherwise. To force the dense path, pass `--attention_mode full`.
 Use the live-ingest client to loop a video into the server at 30 fps:
 
 ```bash
-PYTHONPATH=integrations/flashvsr:flashdreams \
-uv run --no-sync python -m flashvsr.grpc.uplift_client \
-    --continuous \
+uv run --package flashdreams-flashvsr flashdreams-feed-frames \
     --server localhost:50051 \
     --input /path/to/clip.mp4
 
 # finite smoke test:
-PYTHONPATH=integrations/flashvsr:flashdreams \
-uv run --no-sync python -m flashvsr.grpc.uplift_client \
-    --continuous \
+uv run --package flashdreams-flashvsr flashdreams-feed-frames \
     --server localhost:50051 \
     --input /path/to/clip.mp4 \
     --max_chunks 4
@@ -163,8 +160,7 @@ uv run --no-sync python -m flashvsr.grpc.uplift_client \
 For a save-to-disk test client, use:
 
 ```bash
-PYTHONPATH=integrations/flashvsr:flashdreams \
-uv run --no-sync python -m flashvsr.grpc.uplift_client \
+uv run --package flashdreams-flashvsr flashdreams-uplift-client \
     --server localhost:50051 \
     --input /path/to/clip.mp4 \
     --output /tmp/clip_2x.mp4
@@ -288,6 +284,12 @@ for i, (start, size) in enumerate(chunks):
   hand-rolled kernel) or `"torch"` (pure-torch wavelet + AdaIN reference).
 - `enable_sync_and_profile`: per-AR-step CUDA-event profiling. Adds one
   `cuda.synchronize()` per step.
+- Runner `profile_warmup_frames`: output frames excluded from the steady FPS
+  summary. Warmup is applied at chunk granularity. Summary rows use the same
+  names as `realesrgan-upsample`: `model_fps` is CUDA-event timing for
+  `generate`/`finalize`, `pipeline_fps` is the synchronized FlashVSR pipeline
+  chunk path, `video_loop_fps` also includes the chunk CPU copy, and
+  `end_to_end_fps` covers the video pass through output write.
 
 ## Files
 
diff --git a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2.py b/integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2.py
deleted file mode 100644
index 08884d10b..000000000
--- a/integrations/flashvsr/flashvsr/grpc/protos/flashvsr_pb2.py
+++ /dev/null
@@ -1,56 +0,0 @@
-# -*- coding: utf-8 -*-
-# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-# Generated by the protocol buffer compiler.  DO NOT EDIT!
-# NO CHECKED-IN PROTOBUF GENCODE
-# source: flashvsr/grpc/protos/flashvsr.proto
-# Protobuf Python Version: 6.31.1
-"""Generated protocol buffer code."""
-from google.protobuf import descriptor as _descriptor
-from google.protobuf import descriptor_pool as _descriptor_pool
-from google.protobuf import runtime_version as _runtime_version
-from google.protobuf import symbol_database as _symbol_database
-from google.protobuf.internal import builder as _builder
-_runtime_version.ValidateProtobufRuntimeVersion(
-    _runtime_version.Domain.PUBLIC,
-    6,
-    31,
-    1,
-    '',
-    'flashvsr/grpc/protos/flashvsr.proto'
-)
-# @@protoc_insertion_point(imports)
-
-_sym_db = _symbol_database.Default()
-
-
-
-
-DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n#flashvsr/grpc/protos/flashvsr.proto\x12\x08\x66lashvsr\"\x0f\n\rStatusRequest\"\\\n\x0eStatusResponse\x12\r\n\x05ready\x18\x01 \x01(\x08\x12\x0e\n\x06\x64\x65vice\x18\x02 \x01(\t\x12\x12\n\nmodel_name\x18\x03 \x01(\t\x12\x17\n\x0f\x61\x63tive_sessions\x18\x04 \x03(\t\"y\n\x13StartSessionRequest\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x14\n\x0cinput_height\x18\x02 \x01(\x05\x12\x13\n\x0binput_width\x18\x03 \x01(\x05\x12\r\n\x05scale\x18\x04 \x01(\x05\x12\x14\n\x0csparse_ratio\x18\x05 \x01(\x02\"J\n\x14StartSessionResponse\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x0f\n\x07success\x18\x02 \x01(\x08\x12\r\n\x05\x65rror\x18\x03 \x01(\t\"\'\n\x11\x45ndSessionRequest\x12\x12\n\nsession_id\x18\x01 \x01(\t\"%\n\x12\x45ndSessionResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\"\xb1\x02\n\x13UpscaleChunkRequest\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x14\n\x0cinput_height\x18\x02 \x01(\x05\x12\x13\n\x0binput_width\x18\x03 \x01(\x05\x12\r\n\x05scale\x18\x04 \x01(\x05\x12\x14\n\x0csparse_ratio\x18\x05 \x01(\x02\x12\x12\n\nframes_rgb\x18\x06 \x01(\x0c\x12\x12\n\nnum_frames\x18\x07 \x01(\x05\x12\x0e\n\x06height\x18\x08 \x01(\x05\x12\r\n\x05width\x18\t \x01(\x05\x12\x13\n\x0b\x63hunk_index\x18\n \x01(\x05\x12/\n\x0e\x66rame_encoding\x18\x0b \x01(\x0e\x32\x17.flashvsr.FrameEncoding\x12\x13\n\x0b\x66rames_jpeg\x18\x0c \x03(\x0c\x12\x14\n\x0c\x64isplay_only\x18\r \x01(\x08\"\xc1\x01\n\x14UpscaleChunkResponse\x12\x12\n\nsession_id\x18\x01 \x01(\t\x12\x12\n\nframes_rgb\x18\x02 \x01(\x0c\x12\x12\n\nnum_frames\x18\x03 \x01(\x05\x12\x0e\n\x06height\x18\x04 \x01(\x05\x12\r\n\x05width\x18\x05 \x01(\x05\x12\x13\n\x0b\x63hunk_index\x18\x06 \x01(\x05\x12\x12\n\nelapsed_ms\x18\x07 \x01(\x02\x12\r\n\x05\x65rror\x18\x08 \x01(\t\x12\x16\n\x0e\x66rames_omitted\x18\t \x01(\x08*D\n\rFrameEncoding\x12\x1a\n\x16\x46RAME_ENCODING_RAW_RGB\x10\x00\x12\x17\n\x13\x46RAME_ENCODING_JPEG\x10\x01\x32\x89\x03\n\x08\x46lashVSR\x12?\n\nget_status\x12\x17.flashvsr.StatusRequest\x1a\x18.flashvsr.StatusResponse\x12N\n\rstart_session\x12\x1d.flashvsr.StartSessionRequest\x1a\x1e.flashvsr.StartSessionResponse\x12H\n\x0b\x65nd_session\x12\x1b.flashvsr.EndSessionRequest\x1a\x1c.flashvsr.EndSessionResponse\x12N\n\rupscale_chunk\x12\x1d.flashvsr.UpscaleChunkRequest\x1a\x1e.flashvsr.UpscaleChunkResponse\x12R\n\rupscale_video\x12\x1d.flashvsr.UpscaleChunkRequest\x1a\x1e.flashvsr.UpscaleChunkResponse(\x01\x30\x01\x62\x06proto3')
-
-_globals = globals()
-_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
-_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, 'flashvsr.grpc.protos.flashvsr_pb2', _globals)
-if not _descriptor._USE_C_DESCRIPTORS:
-  DESCRIPTOR._loaded_options = None
-  _globals['_FRAMEENCODING']._serialized_start=943
-  _globals['_FRAMEENCODING']._serialized_end=1011
-  _globals['_STATUSREQUEST']._serialized_start=49
-  _globals['_STATUSREQUEST']._serialized_end=64
-  _globals['_STATUSRESPONSE']._serialized_start=66
-  _globals['_STATUSRESPONSE']._serialized_end=158
-  _globals['_STARTSESSIONREQUEST']._serialized_start=160
-  _globals['_STARTSESSIONREQUEST']._serialized_end=281
-  _globals['_STARTSESSIONRESPONSE']._serialized_start=283
-  _globals['_STARTSESSIONRESPONSE']._serialized_end=357
-  _globals['_ENDSESSIONREQUEST']._serialized_start=359
-  _globals['_ENDSESSIONREQUEST']._serialized_end=398
-  _globals['_ENDSESSIONRESPONSE']._serialized_start=400
-  _globals['_ENDSESSIONRESPONSE']._serialized_end=437
-  _globals['_UPSCALECHUNKREQUEST']._serialized_start=440
-  _globals['_UPSCALECHUNKREQUEST']._serialized_end=745
-  _globals['_UPSCALECHUNKRESPONSE']._serialized_start=748
-  _globals['_UPSCALECHUNKRESPONSE']._serialized_end=941
-  _globals['_FLASHVSR']._serialized_start=1014
-  _globals['_FLASHVSR']._serialized_end=1407
-# @@protoc_insertion_point(module_scope)
diff --git a/integrations/flashvsr/flashvsr/grpc/uplift_server.py b/integrations/flashvsr/flashvsr/grpc/uplift_server.py
index dca6663de..9494f7f14 100644
--- a/integrations/flashvsr/flashvsr/grpc/uplift_server.py
+++ b/integrations/flashvsr/flashvsr/grpc/uplift_server.py
@@ -51,14 +51,15 @@
 from loguru import logger
 
 from flashvsr.config import build_flashvsr_v1_1
-from flashvsr.grpc.protos import flashvsr_pb2 as pb2
-from flashvsr.grpc.protos import flashvsr_pb2_grpc as pb2_grpc
-from flashvsr.grpc.streaming_view import (
+from flashdreams.serving.uplift.protos import uplift_pb2 as pb2
+from flashdreams.serving.uplift.protos import uplift_pb2_grpc as pb2_grpc
+from flashdreams.serving.uplift.streaming_view import (
     DEFAULT_VIEWER_CHUNK_QUEUE_DEPTH,
     DEFAULT_VIEWER_FRAME_STRIDE,
     DEFAULT_VIEWER_JPEG_BACKEND,
     DEFAULT_VIEWER_JPEG_QUALITY,
     DEFAULT_VIEWER_MAX_FPS,
+    DEFAULT_VIEWER_PLAYBACK_FPS,
     StreamingViewer,
     decode_jpeg_rgb,
     encode_jpeg_cuda_tensor,
@@ -155,7 +156,7 @@ class _DecodedStreamRequest:
     decode_ms: float
 
 
-class FlashVSR(pb2_grpc.FlashVSRServicer):
+class FlashVSR(pb2_grpc.VideoUpliftServicer):
     def __init__(
         self,
         model_path: str,
@@ -1124,7 +1125,7 @@ def append_coalesced_request(item: _DecodedStreamRequest) -> bool:
 
 
 def _add_flash_vsr_servicer_to_server(servicer: FlashVSR, server: grpc.Server) -> None:
-    pb2_grpc.add_FlashVSRServicer_to_server(servicer, server)
+    pb2_grpc.add_VideoUpliftServicer_to_server(servicer, server)
 
 
 # ---------------------------------------------------------------------------
@@ -1269,8 +1270,19 @@ def main():
         type=float,
         default=DEFAULT_VIEWER_MAX_FPS,
         help=(
-            "Upper FPS cap for HTTP viewer playback. The viewer otherwise uses "
-            "the measured FlashVSR generation speed (default: %(default)s)."
+            "Upper FPS cap for HTTP viewer catch-up playback when completed "
+            "chunks are queued (default: %(default)s)."
+        ),
+    )
+    parser.add_argument(
+        "--viewer_playback_fps",
+        type=float,
+        default=DEFAULT_VIEWER_PLAYBACK_FPS,
+        help=(
+            "Steady HTTP viewer playback FPS. If completed chunks back up, "
+            "the viewer gradually catches up, capped by --viewer_max_fps. "
+            "Set to 0 to pace from model elapsed time "
+            "(default: %(default)s)."
         ),
     )
     parser.add_argument(
@@ -1296,6 +1308,8 @@ def main():
         parser.error("--viewer_jpeg_quality must be between 1 and 100")
     if args.viewer_frame_stride < 1:
         parser.error("--viewer_frame_stride must be at least 1")
+    if args.viewer_playback_fps < 0:
+        parser.error("--viewer_playback_fps must be non-negative")
     attention_mode = _resolve_attention_mode(args.attention_mode)
 
     compile_network = bool(args.compile or args.cuda_graph)
@@ -1339,6 +1353,7 @@ def main():
             jpeg_backend=args.viewer_jpeg_backend,
             chunk_queue_depth=args.viewer_chunk_queue_depth,
             max_fps=args.viewer_max_fps,
+            playback_fps=args.viewer_playback_fps,
             frame_stride=args.viewer_frame_stride,
         )
         viewer.start()
diff --git a/integrations/flashvsr/flashvsr/postprocess.py b/integrations/flashvsr/flashvsr/postprocess.py
index 159150129..51b8b38f6 100644
--- a/integrations/flashvsr/flashvsr/postprocess.py
+++ b/integrations/flashvsr/flashvsr/postprocess.py
@@ -321,14 +321,18 @@ def _next_target_size(self) -> int:
     def _run_flashvsr_chunk(self, clip: Tensor) -> Tensor:
         assert self._pipeline is not None
         assert self._cache is not None
-        output = self._pipeline.generate(
-            autoregressive_index=self._ar_idx,
-            cache=self._cache,
-            input=clip,
-        )
-        self._pipeline.finalize(autoregressive_index=self._ar_idx, cache=self._cache)
-        self._ar_idx += 1
-        return output
+        try:
+            return self._pipeline.generate(
+                autoregressive_index=self._ar_idx,
+                cache=self._cache,
+                input=clip,
+            )
+        finally:
+            self._pipeline.finalize(
+                autoregressive_index=self._ar_idx,
+                cache=self._cache,
+            )
+            self._ar_idx += 1
 
     def _consume_metadata(self, frames: int, *, source: str) -> dict[str, Any]:
         """Consume metadata spans covering ``frames`` buffered frames."""
diff --git a/integrations/flashvsr/flashvsr/runner.py b/integrations/flashvsr/flashvsr/runner.py
index fcd6be577..12e0d3bff 100644
--- a/integrations/flashvsr/flashvsr/runner.py
+++ b/integrations/flashvsr/flashvsr/runner.py
@@ -19,6 +19,7 @@
 
 import json
 import os
+import time
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Literal
@@ -226,6 +227,11 @@ class FlashVSRRunnerConfig(RunnerConfig):
     """Sparse-attention budget multiplier used to re-derive
     ``transformer.topk_ratio`` from each video's post-crop target area."""
 
+    profile_warmup_frames: int = 0
+    """Output frames to exclude from steady-state FPS summaries. Warmup is
+    applied at chunk granularity, so any chunk touched by this threshold is
+    excluded from steady metrics."""
+
 
 def _ensure_mgpu_config_supported(
     config: FlashVSRRunnerConfig, world_size: int
@@ -259,6 +265,27 @@ def _ensure_mgpu_config_supported(
     )
 
 
+def _fps(frames: int, elapsed_s: float) -> float:
+    return frames / elapsed_s if elapsed_s > 0 else 0.0
+
+
+def _duration_fps(frames: list[int], durations_s: list[float]) -> float:
+    return _fps(sum(frames), sum(durations_s))
+
+
+def _duration_ms_fps(frames: list[int], durations_ms: list[float]) -> float:
+    return _fps(sum(frames), sum(durations_ms) / 1000.0)
+
+
+def _steady_start_index(frames: list[int], warmup_frames: int) -> int:
+    remaining = max(warmup_frames, 0)
+    index = 0
+    while index < len(frames) and remaining > 0:
+        remaining -= frames[index]
+        index += 1
+    return index
+
+
 ## Runner
 
 
@@ -433,10 +460,23 @@ def run(self) -> None:
 
         cache = self._initialize_cache()
 
+        run_started_at = time.perf_counter()
         postprocess_stream = self.create_postprocess_stream(fps=fps)
         stats_history: list[dict[str, object]] = []
+        profile_frames: list[int] = []
+        video_loop_durations_s: list[float] = []
+        pipeline_durations_s: list[float] = []
+        model_durations_ms: list[float] = []
         for chunk_idx, (start, size) in enumerate(chunks):
+            video_loop_started_at = time.perf_counter()
             clip = video_t[:, :, start : start + size]
+            pipeline_started_at = time.perf_counter()
+            start_event: torch.cuda.Event | None = None
+            end_event: torch.cuda.Event | None = None
+            if self.pipeline.device.type == "cuda":
+                start_event = torch.cuda.Event(enable_timing=True)
+                end_event = torch.cuda.Event(enable_timing=True)
+                start_event.record()
             video_chunk = self.pipeline.generate(
                 autoregressive_index=chunk_idx,
                 cache=cache,
@@ -444,13 +484,25 @@ def run(self) -> None:
             )
             pipeline_frames = int(video_chunk.shape[2])
             stats = self.pipeline.finalize(autoregressive_index=chunk_idx, cache=cache)
+            if end_event is not None:
+                end_event.record()
+                torch.cuda.synchronize(self.pipeline.device)
+                assert start_event is not None
+                model_ms = float(start_event.elapsed_time(end_event))
+            else:
+                model_ms = (time.perf_counter() - pipeline_started_at) * 1000.0
+            pipeline_elapsed_s = time.perf_counter() - pipeline_started_at
+            chunk_frames = int(video_chunk.shape[2])
+            profile_frames.append(chunk_frames)
+            pipeline_durations_s.append(pipeline_elapsed_s)
+            model_durations_ms.append(model_ms)
             postprocess_stream.process(
                 video_chunk,
                 autoregressive_index=chunk_idx,
             )
-            if postprocess_stream.collect_output and stats is not None:
-                # Pipeline throughput is based on this AR step's direct output.
-                # Postprocess emission/buffering is reported separately.
+            if stats is not None:
+                # Report throughput against the visible output frames for this
+                # chunk. ``fps`` is reserved for the output video's frame rate.
                 chunk_total_ms = stats["total_ms"]
                 chunk_fps = (
                     pipeline_frames / chunk_total_ms * 1000.0
@@ -465,6 +517,7 @@ def run(self) -> None:
                         "fps": chunk_fps,
                     }
                 )
+            video_loop_durations_s.append(time.perf_counter() - video_loop_started_at)
 
         generated = postprocess_stream.finish()
         if generated is None:
@@ -482,6 +535,30 @@ def run(self) -> None:
             f"-> {video_path.resolve()}"
         )
 
+        elapsed_s = time.perf_counter() - run_started_at
+        output_frames = int(generated.shape[2])
+        warmup_start = _steady_start_index(profile_frames, config.profile_warmup_frames)
+        steady_frames = profile_frames[warmup_start:]
+        steady_video_loop_s = video_loop_durations_s[warmup_start:]
+        steady_pipeline_s = pipeline_durations_s[warmup_start:]
+        steady_model_ms = model_durations_ms[warmup_start:]
+        logger.info(
+            "profile "
+            f"total_s={elapsed_s:.3f} "
+            f"end_to_end_fps={_fps(output_frames, elapsed_s):.2f} "
+            f"video_loop_fps={_duration_fps(profile_frames, video_loop_durations_s):.2f} "
+            f"pipeline_fps={_duration_fps(profile_frames, pipeline_durations_s):.2f} "
+            f"model_fps={_duration_ms_fps(profile_frames, model_durations_ms):.2f}"
+        )
+        logger.info(
+            "profile_steady "
+            f"warmup_frames={config.profile_warmup_frames} "
+            f"steady_frames={sum(steady_frames)} "
+            f"steady_video_loop_fps={_duration_fps(steady_frames, steady_video_loop_s):.2f} "
+            f"steady_pipeline_fps={_duration_fps(steady_frames, steady_pipeline_s):.2f} "
+            f"steady_model_fps={_duration_ms_fps(steady_frames, steady_model_ms):.2f}"
+        )
+
         if stats_history:
             stats_path = config.output_dir / f"stats_{config.runner_name}.json"
             stats_path.write_text(json.dumps(stats_history, indent=2))
diff --git a/integrations/flashvsr/pyproject.toml b/integrations/flashvsr/pyproject.toml
index 6128cd57e..0c035016e 100644
--- a/integrations/flashvsr/pyproject.toml
+++ b/integrations/flashvsr/pyproject.toml
@@ -24,12 +24,9 @@ description = "FlashVSR streaming video super-resolution (LR projector + Wan 2.1
 readme = "README.md"
 requires-python = ">=3.10"
 dependencies = [
-    "flashdreams",
-    "grpcio>=1.80",
+    "flashdreams[serving]",
     "mediapy>=1.1",
     "opencv-python>=4.8",
-    "Pillow>=10",
-    "protobuf>=6.31.1",
 ]
 
 [tool.uv]
@@ -45,9 +42,9 @@ dev = [
 ]
 
 [project.scripts]
-flashvsr-grpc-server = "flashvsr.grpc.uplift_server:main"
-flashvsr-grpc-client = "flashvsr.grpc.uplift_client:main"
-flashvsr-feed-frames = "flashvsr.grpc.uplift_client:run_continuous_client"
+uplift-server = "flashvsr.grpc.uplift_server:main"
+flashvsr-grpc-client = "flashdreams.serving.uplift.uplift_client:main"
+flashvsr-feed-frames = "flashdreams.serving.uplift.uplift_client:run_continuous_client"
 
 # Each entry registers one ``runner_name`` slug with ``flashdreams-run``.
 # The discovery layer (``flashdreams.plugins.registry.discover_runners``)
diff --git a/integrations/flashvsr/tests/test_grpc_client.py b/integrations/flashvsr/tests/test_grpc_client.py
index 501db5b23..1bfeb6747 100644
--- a/integrations/flashvsr/tests/test_grpc_client.py
+++ b/integrations/flashvsr/tests/test_grpc_client.py
@@ -18,8 +18,12 @@
 import numpy as np
 import pytest
 from flashvsr.grpc import uplift_server as grpc_server
-from flashvsr.grpc.protos import flashvsr_pb2 as pb2
-from flashvsr.grpc.uplift_client import build_chunk_request, build_chunks
+from flashdreams.serving.uplift.protos import uplift_pb2 as pb2
+from flashdreams.serving.uplift.uplift_client import (
+    build_chunk_request,
+    build_chunks,
+    continuous_loop_frame_count,
+)
 
 pytestmark = pytest.mark.ci_cpu
 
@@ -32,6 +36,30 @@ def test_build_chunks_drops_trailing_partial_chunk() -> None:
     ]
 
 
+def test_continuous_chunk_loop_drops_tail_remainder() -> None:
+    assert continuous_loop_frame_count(
+        157,
+        chunk_frames=8,
+        loop_mode="chunk",
+    ) == 152
+
+
+def test_continuous_frame_loop_keeps_tail_remainder() -> None:
+    assert continuous_loop_frame_count(
+        157,
+        chunk_frames=8,
+        loop_mode="frame",
+    ) == 157
+
+
+def test_continuous_one_frame_chunk_loop_keeps_all_frames() -> None:
+    assert continuous_loop_frame_count(
+        157,
+        chunk_frames=1,
+        loop_mode="chunk",
+    ) == 157
+
+
 def test_build_chunk_request_raw_display_only() -> None:
     frames = np.zeros((8, 4, 6, 3), dtype=np.uint8)
     request = build_chunk_request(
diff --git a/integrations/flashvsr/tests/test_postprocess.py b/integrations/flashvsr/tests/test_postprocess.py
index af4a1be2e..902ea3347 100644
--- a/integrations/flashvsr/tests/test_postprocess.py
+++ b/integrations/flashvsr/tests/test_postprocess.py
@@ -157,7 +157,7 @@ def test_flashvsr_postprocess_can_drop_short_tail(
     assert created[0].inputs == []
 
 
-def test_flashvsr_postprocess_does_not_finalize_when_generate_raises(
+def test_flashvsr_postprocess_finalizes_when_generate_raises(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
     created = _install_fake_builder(monkeypatch)
@@ -181,7 +181,7 @@ def failing_generate(
         session.process(VideoChunk(tensor=video, layout="tchw"))
 
     pipeline = created[0]
-    assert pipeline.finalized == []
+    assert pipeline.finalized == [0]
     assert len(pipeline.inputs) == 1
 
 
diff --git a/integrations/flashvsr/tests/test_streaming_view.py b/integrations/flashvsr/tests/test_streaming_view.py
new file mode 100644
index 000000000..d1edfa307
--- /dev/null
+++ b/integrations/flashvsr/tests/test_streaming_view.py
@@ -0,0 +1,75 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import pytest
+
+from flashdreams.serving.uplift.streaming_view import (
+    StreamingViewer,
+    _ViewerPlaybackChunk,
+)
+
+pytestmark = pytest.mark.ci_cpu
+
+
+def _viewer(
+    *,
+    playback_fps: float = 30.0,
+    max_fps: float = 60.0,
+    chunk_queue_depth: int = 8,
+) -> StreamingViewer:
+    return StreamingViewer(
+        host="127.0.0.1",
+        port=0,
+        jpeg_quality=90,
+        jpeg_backend="pillow",
+        chunk_queue_depth=chunk_queue_depth,
+        max_fps=max_fps,
+        playback_fps=playback_fps,
+        frame_stride=1,
+    )
+
+
+def test_first_viewer_chunk_uses_playback_fps_not_infer_time() -> None:
+    viewer = _viewer(playback_fps=30.0)
+    item = _ViewerPlaybackChunk(elapsed_ms=140.0, ready_at=100.0, jpegs=[b"x"] * 8)
+
+    interval_s = viewer._frame_interval_s(item, 8, queued_chunks=0)
+
+    assert interval_s == pytest.approx(1.0 / 30.0)
+
+
+def test_viewer_ignores_short_term_ready_cadence_without_backlog() -> None:
+    viewer = _viewer(playback_fps=30.0)
+    item = _ViewerPlaybackChunk(
+        elapsed_ms=140.0,
+        ready_at=100.0 + 8.0 / 45.0,
+        jpegs=[b"x"] * 8,
+    )
+
+    interval_s = viewer._frame_interval_s(item, 8, queued_chunks=1)
+
+    assert interval_s == pytest.approx(1.0 / 30.0)
+
+
+def test_viewer_speeds_up_when_ready_queue_has_backlog() -> None:
+    viewer = _viewer(playback_fps=30.0, max_fps=60.0, chunk_queue_depth=8)
+    item = _ViewerPlaybackChunk(
+        elapsed_ms=20.0,
+        ready_at=100.0,
+        jpegs=[b"x"] * 8,
+    )
+
+    interval_s = viewer._frame_interval_s(item, 8, queued_chunks=4)
+
+    assert interval_s == pytest.approx(1.0 / 45.0)
+
+
+def test_viewer_catchup_is_capped_by_max_fps() -> None:
+    viewer = _viewer(playback_fps=30.0, max_fps=60.0, chunk_queue_depth=8)
+    item = _ViewerPlaybackChunk(elapsed_ms=20.0, ready_at=100.0, jpegs=[b"x"] * 8)
+
+    interval_s = viewer._frame_interval_s(item, 8, queued_chunks=8)
+
+    assert interval_s == pytest.approx(1.0 / 60.0)
diff --git a/integrations/realesrgan/README.md b/integrations/realesrgan/README.md
new file mode 100644
index 000000000..220af52ff
--- /dev/null
+++ b/integrations/realesrgan/README.md
@@ -0,0 +1,128 @@
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+-->
+
+# flashdreams-realesrgan
+
+Real-ESRGAN frame and video upsampling packaged as a FlashDreams workspace
+integration. It provides:
+
+- a reusable `RealESRGANUpsampler` Python API,
+- a `RealESRGANPostProcessorConfig` for FlashDreams generated RGB video chunks,
+- a `realesrgan-upsample` CLI for image and video files.
+
+The architecture definitions follow the public Real-ESRGAN / BasicSR RRDBNet
+and compact SRVGG network layouts so public checkpoints can be loaded directly.
+Weights are downloaded at first use into
+`$FLASHDREAMS_CACHE_DIR/realesrgan` or `~/.cache/flashdreams/realesrgan`.
+
+## Install
+
+From the repository root:
+
+```bash
+uv sync --package flashdreams-realesrgan --extra dev
+```
+
+The root workspace glob includes `integrations/realesrgan`, so editable
+workspace usage works without extra path setup.
+
+## CLI
+
+Upsample one image:
+
+```bash
+uv run --package flashdreams-realesrgan realesrgan-upsample \
+  --input /path/to/input.png \
+  --output /path/to/input_2x.png \
+  --scale 2
+```
+
+Upsample a video:
+
+```bash
+uv run --package flashdreams-realesrgan realesrgan-upsample \
+  --input /path/to/input.mp4 \
+  --output /path/to/input_2x.mp4 \
+  --scale 2 \
+  --tile 256 \
+  --compile
+```
+
+Use Real-ESRGAN as a `flashdreams-run` postprocessor:
+
+```bash
+uv run --package flashdreams-realesrgan flashdreams-run \
+  --postprocess.preset realesrgan \
+  wan21-t2v-1.3b-480p
+```
+
+Useful flags:
+
+| flag | description |
+| --- | --- |
+| `--scale {2,4}` | Output scale. Defaults to `2`. |
+| `--model-name` | Public checkpoint name such as `RealESRGAN_x2plus` or `RealESRGAN_x4plus`. |
+| `--model-path` | Local checkpoint path. Skips public download. |
+| `--tile` | Tile size for lower peak VRAM. `0` processes whole frames. |
+| `--fp32` | Disable fp16 CUDA inference. |
+| `--compile` | Enable `torch.compile` with `mode="reduce-overhead"`. |
+| `--compile-mode` | Override the `torch.compile` mode. |
+| `--profile-warmup-frames` | Frames excluded from steady FPS metrics. Defaults to `10`. |
+| `--max-frames` | Video-only frame cap for quick smoke tests. |
+
+For `flashdreams-run`, Real-ESRGAN is selected with the registered
+`--postprocess.preset realesrgan` preset. The preset uses the default 2x public
+checkpoint on CUDA. Use the standalone `realesrgan-upsample` CLI or configure a
+`RealESRGANPostProcessorConfig` directly when you need custom scale, tiling,
+precision, checkpoint, or compile settings.
+
+Video runs print two profile rows. `model_fps` is CUDA-event timing for the
+Real-ESRGAN model call only. `pipeline_fps` includes image conversion,
+CPU/GPU copies, padding/cropping, the model call, and output conversion.
+`video_loop_fps` adds video frame read/write around the pipeline call.
+`end_to_end_fps` is the whole video pass wall-clock throughput.
+
+## FlashDreams Postprocess API
+
+```python
+from flashdreams.infra.postprocess import VideoPostprocessChainConfig
+from realesrgan import RealESRGANPostProcessorConfig
+
+postprocess = VideoPostprocessChainConfig(
+    processors=(
+        RealESRGANPostProcessorConfig(
+            scale=2,
+            model_name="RealESRGAN_x2plus",
+            tile=256,
+            compile_model=True,
+            device="cuda",
+        ),
+    )
+)
+```
+
+The postprocessor accepts generated RGB chunks in any FlashDreams video layout,
+runs frame-local Real-ESRGAN inference, and returns `bvtchw` chunks in
+`[-1, 1]`.
+
+## Tests
+
+CPU-safe tests use random weights and do not download checkpoints:
+
+```bash
+uv run --package flashdreams-realesrgan pytest integrations/realesrgan/tests/test_realesrgan.py -m ci_cpu
+```
+
+Checkpoint-backed GPU smoke test:
+
+```bash
+uv run --package flashdreams-realesrgan python integrations/realesrgan/scripts/gpu_smoke.py \
+  --scale 2 \
+  --device cuda
+```
+
+The smoke script downloads `RealESRGAN_x2plus` once, runs one CUDA fp16
+forward pass on a `16x16` RGB frame, checks the `32x32` output, and prints
+model load time, inference time, and peak allocated VRAM.
diff --git a/integrations/realesrgan/pyproject.toml b/integrations/realesrgan/pyproject.toml
new file mode 100644
index 000000000..818ba1e96
--- /dev/null
+++ b/integrations/realesrgan/pyproject.toml
@@ -0,0 +1,59 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+[build-system]
+requires = ["setuptools>=69", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flashdreams-realesrgan"
+version = "0.1.0"
+description = "Real-ESRGAN frame and video upsampler integration for flashdreams."
+readme = "README.md"
+requires-python = ">=3.10"
+dependencies = [
+    "flashdreams[serving]",
+    "opencv-python-headless>=4.8",
+]
+
+[tool.uv]
+managed = true
+
+[tool.uv.sources]
+flashdreams = { workspace = true }
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0",
+]
+
+[project.scripts]
+realesrgan-upsample = "realesrgan.cli:main"
+uplift-server = "realesrgan.grpc.uplift_server:main"
+
+[project.entry-points."flashdreams.postprocess_presets"]
+realesrgan = "realesrgan.postprocess:POSTPROCESS_PRESET_REALESRGAN"
+
+[tool.setuptools.packages.find]
+include = ["realesrgan*"]
+exclude = ["tests"]
+
+[tool.pytest.ini_options]
+addopts = "--import-mode=importlib -p flashdreams._pytest_plugins.marker_enforcement"
+markers = [
+    "ci_cpu: CPU-safe test, runs on the CPU CI runner",
+    "ci_gpu: requires GPU or libGL (cv2), runs on the GPU CI runner",
+    "manual: heavy or environment-specific test, opt-in only",
+]
diff --git a/integrations/realesrgan/realesrgan/__init__.py b/integrations/realesrgan/realesrgan/__init__.py
new file mode 100644
index 000000000..fb64357ff
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/__init__.py
@@ -0,0 +1,38 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Real-ESRGAN upsampling integration for FlashDreams."""
+
+from realesrgan.postprocess import (
+    POSTPROCESS_PRESET_REALESRGAN,
+    RealESRGANPostProcessorConfig,
+)
+from realesrgan.upsampler import (
+    MODEL_CONFIGS,
+    MODEL_URLS,
+    RealESRGANFrameProfile,
+    RealESRGANUpsampler,
+    create_model,
+)
+
+__all__ = [
+    "MODEL_CONFIGS",
+    "MODEL_URLS",
+    "POSTPROCESS_PRESET_REALESRGAN",
+    "RealESRGANFrameProfile",
+    "RealESRGANPostProcessorConfig",
+    "RealESRGANUpsampler",
+    "create_model",
+]
diff --git a/integrations/realesrgan/realesrgan/archs.py b/integrations/realesrgan/realesrgan/archs.py
new file mode 100644
index 000000000..7b94ee1b5
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/archs.py
@@ -0,0 +1,191 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Real-ESRGAN network architectures.
+
+The RRDBNet and compact SRVGG definitions mirror the public Real-ESRGAN
+and BasicSR architectures so upstream checkpoints can be loaded directly.
+"""
+
+from __future__ import annotations
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+def pixel_unshuffle(x: torch.Tensor, scale: int) -> torch.Tensor:
+    """Return the inverse pixel-shuffle transform for ``x``."""
+    batch, channels, height, width = x.shape
+    assert height % scale == 0 and width % scale == 0
+    out_channels = channels * scale * scale
+    height_out = height // scale
+    width_out = width // scale
+    x = x.view(batch, channels, height_out, scale, width_out, scale)
+    return x.permute(0, 1, 3, 5, 2, 4).reshape(
+        batch, out_channels, height_out, width_out
+    )
+
+
+def default_init_weights(module_list: list[nn.Module], scale: float = 1.0) -> None:
+    """Initialize convolutional and linear weights in ``module_list``."""
+    for module in module_list:
+        for child in module.modules():
+            if isinstance(child, nn.Conv2d):
+                nn.init.kaiming_normal_(child.weight, a=0, mode="fan_in")
+                child.weight.data *= scale
+                if child.bias is not None:
+                    child.bias.data.zero_()
+            elif isinstance(child, nn.Linear):
+                nn.init.kaiming_normal_(child.weight, a=0, mode="fan_in")
+                child.weight.data *= scale
+                if child.bias is not None:
+                    child.bias.data.zero_()
+
+
+class ResidualDenseBlock(nn.Module):
+    """Residual dense block used by RRDBNet."""
+
+    def __init__(self, num_feat: int = 64, num_grow_ch: int = 32) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
+        self.conv2 = nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv3 = nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv4 = nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv5 = nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+        default_init_weights(
+            [self.conv1, self.conv2, self.conv3, self.conv4, self.conv5],
+            0.1,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Run the residual dense block."""
+        x1 = self.lrelu(self.conv1(x))
+        x2 = self.lrelu(self.conv2(torch.cat((x, x1), dim=1)))
+        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), dim=1)))
+        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), dim=1)))
+        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), dim=1))
+        return x5 * 0.2 + x
+
+
+class RRDB(nn.Module):
+    """Residual-in-residual dense block used by RRDBNet."""
+
+    def __init__(self, num_feat: int, num_grow_ch: int = 32) -> None:
+        super().__init__()
+        self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Run the residual-in-residual block."""
+        out = self.rdb1(x)
+        out = self.rdb2(out)
+        out = self.rdb3(out)
+        return out * 0.2 + x
+
+
+class RRDBNet(nn.Module):
+    """RRDBNet used by RealESRGAN_x2plus and RealESRGAN_x4plus."""
+
+    def __init__(
+        self,
+        num_in_ch: int,
+        num_out_ch: int,
+        scale: int = 4,
+        num_feat: int = 64,
+        num_block: int = 23,
+        num_grow_ch: int = 32,
+    ) -> None:
+        super().__init__()
+        self.scale = scale
+        if scale == 2:
+            num_in_ch *= 4
+        elif scale == 1:
+            num_in_ch *= 16
+
+        self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+        self.body = nn.Sequential(
+            *[
+                RRDB(num_feat=num_feat, num_grow_ch=num_grow_ch)
+                for _ in range(num_block)
+            ]
+        )
+        self.conv_body = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Run RRDBNet on an RGB image batch in ``[0, 1]``."""
+        if self.scale == 2:
+            feat = pixel_unshuffle(x, scale=2)
+        elif self.scale == 1:
+            feat = pixel_unshuffle(x, scale=4)
+        else:
+            feat = x
+
+        feat = self.conv_first(feat)
+        body_feat = self.conv_body(self.body(feat))
+        feat = feat + body_feat
+        feat = self.lrelu(self.conv_up1(F.interpolate(feat, scale_factor=2)))
+        feat = self.lrelu(self.conv_up2(F.interpolate(feat, scale_factor=2)))
+        return self.conv_last(self.lrelu(self.conv_hr(feat)))
+
+
+class SRVGGNetCompact(nn.Module):
+    """Compact VGG-style super-resolution network used by Real-ESRGAN v3."""
+
+    def __init__(
+        self,
+        num_in_ch: int = 3,
+        num_out_ch: int = 3,
+        num_feat: int = 64,
+        num_conv: int = 16,
+        upscale: int = 4,
+        act_type: str = "prelu",
+    ) -> None:
+        super().__init__()
+        self.upscale = upscale
+        self.body = nn.ModuleList()
+        self.body.append(nn.Conv2d(num_in_ch, num_feat, 3, 1, 1))
+        self.body.append(self._make_activation(act_type, num_feat))
+        for _ in range(num_conv):
+            self.body.append(nn.Conv2d(num_feat, num_feat, 3, 1, 1))
+            self.body.append(self._make_activation(act_type, num_feat))
+        self.body.append(nn.Conv2d(num_feat, num_out_ch * upscale * upscale, 3, 1, 1))
+        self.upsampler = nn.PixelShuffle(upscale)
+
+    @staticmethod
+    def _make_activation(act_type: str, num_feat: int) -> nn.Module:
+        if act_type == "relu":
+            return nn.ReLU(inplace=True)
+        if act_type == "prelu":
+            return nn.PReLU(num_parameters=num_feat)
+        if act_type == "leakyrelu":
+            return nn.LeakyReLU(negative_slope=0.1, inplace=True)
+        raise ValueError(f"Unknown activation type: {act_type}")
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Run SRVGG on an RGB image batch in ``[0, 1]``."""
+        out = x
+        for layer in self.body:
+            out = layer(out)
+        out = self.upsampler(out)
+        base = F.interpolate(x, scale_factor=self.upscale, mode="nearest")
+        return out + base
diff --git a/integrations/realesrgan/realesrgan/cli.py b/integrations/realesrgan/realesrgan/cli.py
new file mode 100644
index 000000000..b7394421c
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/cli.py
@@ -0,0 +1,217 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Command-line Real-ESRGAN image/video upsampler."""
+
+from __future__ import annotations
+
+import argparse
+import time
+from pathlib import Path
+from typing import Literal
+
+import cv2
+import torch
+
+from realesrgan.upsampler import (
+    RealESRGANUpsampler,
+    default_model_name,
+    write_bgr_image,
+)
+
+IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".tif", ".tiff", ".webp"}
+VIDEO_EXTENSIONS = {".mp4", ".avi", ".mov", ".mkv", ".webm", ".flv", ".wmv"}
+
+
+def main() -> None:
+    """Run the Real-ESRGAN upsampler CLI."""
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--input", "-i", type=Path, required=True)
+    parser.add_argument("--output", "-o", type=Path, required=True)
+    parser.add_argument("--scale", "-s", type=int, choices=(2, 4), default=2)
+    parser.add_argument("--model-name", type=str, default=None)
+    parser.add_argument("--model-path", type=Path, default=None)
+    parser.add_argument("--tile", type=int, default=0)
+    parser.add_argument("--tile-pad", type=int, default=10)
+    parser.add_argument("--pre-pad", type=int, default=10)
+    parser.add_argument("--fp32", action="store_true")
+    parser.add_argument(
+        "--compile",
+        action="store_true",
+        help="Compile the model with torch.compile().",
+    )
+    parser.add_argument(
+        "--compile-mode",
+        type=str,
+        default="reduce-overhead",
+        help="Mode passed to torch.compile() when --compile is used.",
+    )
+    parser.add_argument("--device", type=str, default=None)
+    parser.add_argument(
+        "--profile-warmup-frames",
+        type=int,
+        default=10,
+        help="Frames to exclude from steady FPS profiling.",
+    )
+    parser.add_argument(
+        "--max-frames",
+        type=int,
+        default=None,
+        help="Optional cap for video smoke tests.",
+    )
+    args = parser.parse_args()
+
+    scale: Literal[2, 4] = 2 if args.scale == 2 else 4
+    model_name = args.model_name or default_model_name(scale)
+    upsampler = RealESRGANUpsampler(
+        scale=scale,
+        model_name=model_name,
+        model_path=args.model_path,
+        tile=args.tile,
+        tile_pad=args.tile_pad,
+        pre_pad=args.pre_pad,
+        half=not args.fp32,
+        compile_model=args.compile,
+        compile_mode=args.compile_mode,
+        device=args.device,
+    )
+
+    suffix = args.input.suffix.lower()
+    if suffix in IMAGE_EXTENSIONS:
+        _upsample_image(args.input, args.output, upsampler)
+        return
+    if suffix in VIDEO_EXTENSIONS:
+        _upsample_video(
+            args.input,
+            args.output,
+            upsampler,
+            max_frames=args.max_frames,
+            profile_warmup_frames=args.profile_warmup_frames,
+        )
+        return
+    raise SystemExit(f"Unsupported input extension {suffix!r}.")
+
+
+def _upsample_image(
+    input_path: Path,
+    output_path: Path,
+    upsampler: RealESRGANUpsampler,
+) -> None:
+    image = cv2.imread(str(input_path), cv2.IMREAD_UNCHANGED)
+    if image is None:
+        raise RuntimeError(f"Failed to read image {input_path}")
+    output, mode = upsampler.upsample_bgr_image(image)
+    if mode == "RGBA":
+        output_path = output_path.with_suffix(".png")
+    write_bgr_image(output_path, output)
+    print(f"wrote {output.shape[1]}x{output.shape[0]} {output_path}")
+
+
+def _upsample_video(
+    input_path: Path,
+    output_path: Path,
+    upsampler: RealESRGANUpsampler,
+    *,
+    max_frames: int | None,
+    profile_warmup_frames: int,
+) -> None:
+    cap = cv2.VideoCapture(str(input_path))
+    if not cap.isOpened():
+        raise RuntimeError(f"Failed to open video {input_path}")
+
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    fps = float(cap.get(cv2.CAP_PROP_FPS) or 30.0)
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    fourcc_factory = getattr(cv2, "VideoWriter_fourcc")
+    writer = cv2.VideoWriter(
+        str(output_path),
+        fourcc_factory(*"mp4v"),
+        fps,
+        (width * upsampler.scale, height * upsampler.scale),
+    )
+    if not writer.isOpened():
+        cap.release()
+        raise RuntimeError(f"Failed to create video writer {output_path}")
+
+    start = time.perf_counter()
+    frame_idx = 0
+    process_durations: list[float] = []
+    pipeline_durations: list[float] = []
+    model_durations_ms: list[float | None] = []
+    try:
+        while True:
+            process_start = time.perf_counter()
+            ok, frame = cap.read()
+            if not ok:
+                break
+            pipeline_start = time.perf_counter()
+            output, _mode, profile = upsampler.upsample_bgr_image_profiled(frame)
+            _synchronize(upsampler)
+            pipeline_durations.append(time.perf_counter() - pipeline_start)
+            model_durations_ms.append(profile.model_ms)
+            writer.write(output)
+            process_durations.append(time.perf_counter() - process_start)
+            frame_idx += 1
+            if max_frames is not None and frame_idx >= max_frames:
+                break
+    finally:
+        cap.release()
+        writer.release()
+
+    elapsed = time.perf_counter() - start
+    total_note = f"/{total_frames}" if total_frames else ""
+    warmup_frames = min(max(profile_warmup_frames, 0), len(process_durations))
+    steady_process = process_durations[warmup_frames:]
+    steady_pipeline = pipeline_durations[warmup_frames:]
+    steady_model_ms = model_durations_ms[warmup_frames:]
+    print(f"wrote {frame_idx}{total_note} frames -> {output_path}")
+    print(
+        f"profile total_s={elapsed:.3f} "
+        f"end_to_end_fps={_fps(frame_idx, elapsed):.2f} "
+        f"video_loop_fps={_duration_fps(process_durations):.2f} "
+        f"pipeline_fps={_duration_fps(pipeline_durations):.2f} "
+        f"model_fps={_duration_ms_fps(model_durations_ms):.2f}"
+    )
+    print(
+        f"profile_steady warmup_frames={warmup_frames} "
+        f"steady_frames={len(steady_process)} "
+        f"steady_video_loop_fps={_duration_fps(steady_process):.2f} "
+        f"steady_pipeline_fps={_duration_fps(steady_pipeline):.2f} "
+        f"steady_model_fps={_duration_ms_fps(steady_model_ms):.2f}"
+    )
+
+
+def _synchronize(upsampler: RealESRGANUpsampler) -> None:
+    if upsampler.device.type == "cuda":
+        torch.cuda.synchronize(upsampler.device)
+
+
+def _duration_fps(durations: list[float]) -> float:
+    return _fps(len(durations), sum(durations))
+
+
+def _duration_ms_fps(durations_ms: list[float | None]) -> float:
+    durations = [duration for duration in durations_ms if duration is not None]
+    return _fps(len(durations), sum(durations) / 1000.0)
+
+
+def _fps(frames: int, elapsed: float) -> float:
+    return frames / elapsed if elapsed > 0 else 0.0
+
+
+if __name__ == "__main__":
+    main()
diff --git a/integrations/realesrgan/realesrgan/grpc/__init__.py b/integrations/realesrgan/realesrgan/grpc/__init__.py
new file mode 100644
index 000000000..9cce0e10a
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/grpc/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""gRPC serving utilities for the Real-ESRGAN integration."""
diff --git a/integrations/realesrgan/realesrgan/grpc/uplift_server.py b/integrations/realesrgan/realesrgan/grpc/uplift_server.py
new file mode 100644
index 000000000..e066030da
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/grpc/uplift_server.py
@@ -0,0 +1,917 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Real-ESRGAN implementation of the FlashVSR uplift gRPC protocol."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import signal
+import sys
+import threading
+import time
+import uuid
+from concurrent import futures
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Literal
+
+import grpc
+import numpy as np
+import torch
+from loguru import logger
+
+from flashdreams.serving.uplift.protos import uplift_pb2 as pb2
+from flashdreams.serving.uplift.protos import uplift_pb2_grpc as pb2_grpc
+from flashdreams.serving.uplift.streaming_view import (
+    DEFAULT_VIEWER_CHUNK_QUEUE_DEPTH,
+    DEFAULT_VIEWER_FRAME_STRIDE,
+    DEFAULT_VIEWER_JPEG_BACKEND,
+    DEFAULT_VIEWER_JPEG_QUALITY,
+    DEFAULT_VIEWER_MAX_FPS,
+    DEFAULT_VIEWER_PLAYBACK_FPS,
+    StreamingViewer,
+    decode_jpeg_rgb,
+    encode_jpeg_cuda_tensor,
+)
+from realesrgan.upsampler import RealESRGANUpsampler, default_model_name
+
+DEFAULT_PORT = 50052
+DEFAULT_MAX_MESSAGE_MB = 512
+DEFAULT_STREAM_QUEUE_DEPTH = 16
+Scale = Literal[2, 4]
+
+
+@dataclass
+class _Session:
+    scale: Scale
+    model_name: str
+    created_at: float = field(default_factory=time.time)
+
+
+@dataclass(frozen=True)
+class _RunChunkResult:
+    frames_rgb: bytes
+    frames_out: np.ndarray | None
+    num_frames: int
+    height: int
+    width: int
+    elapsed_ms: float
+
+
+@dataclass(frozen=True)
+class _FrameRunResult:
+    output: torch.Tensor
+    model_ms: float | None
+
+
+class _StaticShapeExecutor:
+    """Run Real-ESRGAN with stable CUDA input storage for one input shape."""
+
+    def __init__(self, upsampler: RealESRGANUpsampler, height: int, width: int) -> None:
+        if upsampler.device.type != "cuda":
+            raise ValueError("Static Real-ESRGAN executor requires a CUDA device")
+        if upsampler.tile > 0:
+            raise ValueError("Static Real-ESRGAN executor does not support tiling")
+        self.upsampler = upsampler
+        self.height = int(height)
+        self.width = int(width)
+        self.scale = int(upsampler.scale)
+        self.device = upsampler.device
+        self.dtype = upsampler.dtype
+        self._input = torch.empty(
+            (1, 3, self.height, self.width),
+            device=self.device,
+            dtype=self.dtype,
+        )
+        total_pad_h, total_pad_w = self._total_padding()
+        self._crop = (
+            total_pad_h * upsampler.model_scale,
+            total_pad_w * upsampler.model_scale,
+        )
+        self._model_input = (
+            torch.empty(
+                (1, 3, self.height + total_pad_h, self.width + total_pad_w),
+                device=self.device,
+                dtype=self.dtype,
+            )
+            if total_pad_h or total_pad_w
+            else self._input
+        )
+        self._total_pad_h = total_pad_h
+        self._total_pad_w = total_pad_w
+
+    @property
+    def output_height(self) -> int:
+        return self.height * self.scale
+
+    @property
+    def output_width(self) -> int:
+        return self.width * self.scale
+
+    @torch.no_grad()
+    def run_frame(self, frame_rgb: np.ndarray) -> _FrameRunResult:
+        self._copy_frame(frame_rgb)
+        self._fill_model_input()
+
+        start: torch.cuda.Event | None = None
+        end: torch.cuda.Event | None = None
+        if self.upsampler.compile_model:
+            torch.compiler.cudagraph_mark_step_begin()
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        start.record()
+        output = self.upsampler.model(self._model_input)
+        end.record()
+        output = self.upsampler._crop_output(output, self._crop)
+        return _FrameRunResult(output=output, model_ms=_elapsed_ms(start, end))
+
+    def _total_padding(self) -> tuple[int, int]:
+        pad_h = int(self.upsampler.pre_pad)
+        pad_w = int(self.upsampler.pre_pad)
+        if self.upsampler.model_scale == 2:
+            pad_h += (2 - (self.height + pad_h) % 2) % 2
+            pad_w += (2 - (self.width + pad_w) % 2) % 2
+        return pad_h, pad_w
+
+    def _copy_frame(self, frame_rgb: np.ndarray) -> None:
+        if frame_rgb.shape[:2] != (self.height, self.width):
+            raise ValueError(
+                f"Static executor expected {self.height}×{self.width}, "
+                f"got {frame_rgb.shape[0]}×{frame_rgb.shape[1]}"
+            )
+        cpu = torch.from_numpy(np.ascontiguousarray(frame_rgb))
+        self._input[0].copy_(cpu.permute(2, 0, 1), non_blocking=False)
+        self._input.div_(255.0)
+
+    def _fill_model_input(self) -> None:
+        if self._model_input is self._input:
+            return
+        h = self.height
+        w = self.width
+        pad_h = self._total_pad_h
+        pad_w = self._total_pad_w
+        self._model_input[:, :, :h, :w].copy_(self._input)
+        if pad_w:
+            self._model_input[:, :, :h, w:].copy_(
+                torch.flip(
+                    self._model_input[:, :, :h, w - pad_w - 1 : w - 1],
+                    dims=[3],
+                )
+            )
+        if pad_h:
+            self._model_input[:, :, h:, :].copy_(
+                torch.flip(
+                    self._model_input[:, :, h - pad_h - 1 : h - 1, :],
+                    dims=[2],
+                )
+            )
+
+
+def _elapsed_ms(
+    start: torch.cuda.Event | None, end: torch.cuda.Event | None
+) -> float | None:
+    if start is None or end is None:
+        return None
+    end.synchronize()
+    return float(start.elapsed_time(end))
+
+
+def _output_tensor_to_rgb(output: torch.Tensor) -> np.ndarray:
+    """Convert direct Real-ESRGAN model output [1, 3, H, W] in [0, 1] to RGB."""
+    return np.ascontiguousarray(
+        output[0]
+        .detach()
+        .float()
+        .clamp(0, 1)
+        .mul(255.0)
+        .round()
+        .to(torch.uint8)
+        .permute(1, 2, 0)
+        .contiguous()
+        .cpu()
+        .numpy()
+    )
+
+
+def _resolve_scale(scale: int) -> Scale:
+    if scale == 2:
+        return 2
+    if scale == 4:
+        return 4
+    raise ValueError(f"Real-ESRGAN scale must be 2 or 4, got {scale}")
+
+
+def _synchronize_device(device: torch.device | str) -> None:
+    device = torch.device(device)
+    if device.type == "cuda" and torch.cuda.is_available():
+        torch.cuda.synchronize(device)
+
+
+def _context_set_error(context: grpc.ServicerContext | None, details: str) -> None:
+    if context is None:
+        return
+    context.set_code(grpc.StatusCode.INTERNAL)
+    context.set_details(details)
+
+
+class RealESRGANUplift(pb2_grpc.VideoUpliftServicer):
+    """Serve Real-ESRGAN behind the existing uplift gRPC surface."""
+
+    def __init__(
+        self,
+        *,
+        default_scale: int = 2,
+        model_name: str | None = None,
+        model_path: str | Path | None = None,
+        tile: int = 0,
+        tile_pad: int = 10,
+        pre_pad: int = 10,
+        half: bool = True,
+        compile_model: bool = False,
+        compile_mode: str = "reduce-overhead",
+        device: str | torch.device | None = None,
+        stream_queue_depth: int = DEFAULT_STREAM_QUEUE_DEPTH,
+        viewer: StreamingViewer | None = None,
+        omit_grpc_frames_when_viewing: bool = False,
+        warmup_height: int = 64,
+        warmup_width: int = 64,
+        warmup_frames: int = 3,
+        warmup: bool = True,
+        load_checkpoint: bool = True,
+    ) -> None:
+        self._default_scale = _resolve_scale(default_scale)
+        self._default_model_name = model_name
+        self._model_path = model_path
+        self._tile = int(tile)
+        self._tile_pad = int(tile_pad)
+        self._pre_pad = int(pre_pad)
+        self._half = bool(half)
+        self._compile_model = bool(compile_model)
+        self._compile_mode = compile_mode
+        self._device = torch.device(
+            device
+            if device is not None
+            else ("cuda" if torch.cuda.is_available() else "cpu")
+        )
+        self._stream_queue_depth = max(1, int(stream_queue_depth))
+        self._viewer = viewer
+        self._omit_grpc_frames_when_viewing = bool(
+            omit_grpc_frames_when_viewing and viewer is not None
+        )
+        self._load_checkpoint = bool(load_checkpoint)
+
+        self._upsampler_pool: dict[tuple[Scale, str], RealESRGANUpsampler] = {}
+        self._executor_pool: dict[
+            tuple[Scale, str, int, int], _StaticShapeExecutor
+        ] = {}
+        self._pool_lock = threading.Lock()
+        self._sessions: dict[str, _Session] = {}
+        self._stream_sessions: set[str] = set()
+        self._sessions_lock = threading.Lock()
+        self._gpu_lock = threading.Lock()
+        self._warned_cuda_viewer_fallback = False
+
+        logger.info(
+            "Loading Real-ESRGAN model scale={} model={} device={} half={} "
+            "compile={} tile={} ...",
+            self._default_scale,
+            self._model_name_for_scale(self._default_scale),
+            self._device,
+            self._half,
+            self._compile_model,
+            self._tile,
+        )
+        default_upsampler = self._get_upsampler(self._default_scale)
+        if warmup:
+            self._warm_up(
+                default_upsampler,
+                self._default_scale,
+                warmup_height,
+                warmup_width,
+                warmup_frames,
+            )
+        logger.info("Real-ESRGAN uplift server ready.")
+
+    def _model_name_for_scale(self, scale: Scale) -> str:
+        return self._default_model_name or default_model_name(scale)
+
+    def _get_upsampler(self, scale: int) -> RealESRGANUpsampler:
+        resolved_scale = _resolve_scale(scale)
+        model_name = self._model_name_for_scale(resolved_scale)
+        key = (resolved_scale, model_name)
+        with self._pool_lock:
+            if key not in self._upsampler_pool:
+                logger.info(
+                    "Creating Real-ESRGAN upsampler scale={} model={} device={}",
+                    resolved_scale,
+                    model_name,
+                    self._device,
+                )
+                self._upsampler_pool[key] = RealESRGANUpsampler(
+                    scale=resolved_scale,
+                    model_name=model_name,
+                    model_path=self._model_path,
+                    tile=self._tile,
+                    tile_pad=self._tile_pad,
+                    pre_pad=self._pre_pad,
+                    half=self._half,
+                    compile_model=self._compile_model,
+                    compile_mode=self._compile_mode,
+                    device=self._device,
+                    load_checkpoint=self._load_checkpoint,
+                )
+            return self._upsampler_pool[key]
+
+    def _get_executor(
+        self,
+        upsampler: RealESRGANUpsampler,
+        scale: int,
+        height: int,
+        width: int,
+    ) -> _StaticShapeExecutor | None:
+        if self._device.type != "cuda" or self._tile > 0:
+            return None
+        if not hasattr(upsampler, "model"):
+            return None
+        resolved_scale = _resolve_scale(scale)
+        model_name = self._model_name_for_scale(resolved_scale)
+        key = (resolved_scale, model_name, int(height), int(width))
+        with self._pool_lock:
+            if key not in self._executor_pool:
+                logger.info(
+                    "Creating static Real-ESRGAN executor for {}×{} scale={} "
+                    "model={} compile_mode={}",
+                    height,
+                    width,
+                    resolved_scale,
+                    model_name,
+                    self._compile_mode if self._compile_model else "disabled",
+                )
+                self._executor_pool[key] = _StaticShapeExecutor(
+                    upsampler=upsampler,
+                    height=height,
+                    width=width,
+                )
+            return self._executor_pool[key]
+
+    def _warm_up(
+        self,
+        upsampler: RealESRGANUpsampler,
+        scale: int,
+        height: int,
+        width: int,
+        frames: int,
+    ) -> None:
+        height = max(8, int(height))
+        width = max(8, int(width))
+        frames = max(1, int(frames))
+        logger.info(
+            "Warming up Real-ESRGAN with {} {}×{} frame(s) ...",
+            frames,
+            height,
+            width,
+        )
+        frame = np.zeros((height, width, 3), dtype=np.uint8)
+        start = time.perf_counter()
+        with self._gpu_lock:
+            executor = self._get_executor(upsampler, scale, height, width)
+            for _ in range(frames):
+                if executor is not None:
+                    executor.run_frame(frame)
+                else:
+                    self._upsample_frame_rgb(upsampler, frame)
+        _synchronize_device(self._device)
+        logger.info(
+            "Real-ESRGAN warmup complete in {:.0f} ms.",
+            (time.perf_counter() - start) * 1000.0,
+        )
+
+    def _request_to_frames_rgb(self, request: pb2.UpscaleChunkRequest) -> np.ndarray:
+        total = int(request.num_frames)
+        height = int(request.height)
+        width = int(request.width)
+        if total <= 0:
+            raise ValueError("num_frames must be positive")
+        if request.frame_encoding == pb2.FRAME_ENCODING_JPEG:
+            if len(request.frames_jpeg) != total:
+                raise ValueError(
+                    f"frame_encoding=JPEG requires {total} frames_jpeg payloads; "
+                    f"got {len(request.frames_jpeg)}"
+                )
+            frames = [decode_jpeg_rgb(frame) for frame in request.frames_jpeg]
+            arr = np.stack(frames, axis=0)
+            if height and width and arr.shape[1:3] != (height, width):
+                raise ValueError(
+                    f"JPEG payload dimensions {arr.shape[1]}×{arr.shape[2]} "
+                    f"do not match request height/width {height}×{width}"
+                )
+            return np.ascontiguousarray(arr)
+
+        expected = total * height * width * 3
+        if len(request.frames_rgb) != expected:
+            raise ValueError(
+                f"RAW_RGB payload has {len(request.frames_rgb)} bytes; "
+                f"expected {expected} for {total}×{height}×{width}×3"
+            )
+        return np.ascontiguousarray(
+            np.frombuffer(request.frames_rgb, dtype=np.uint8).reshape(
+                total, height, width, 3
+            )
+        )
+
+    def _upsample_frame_rgb(
+        self, upsampler: RealESRGANUpsampler, frame: np.ndarray
+    ) -> np.ndarray:
+        tensor = torch.from_numpy(
+            np.ascontiguousarray(frame).astype(np.float32).transpose(2, 0, 1)
+        )
+        tensor = tensor / 127.5 - 1.0
+        output = upsampler.upsample_frame_tensor(tensor)
+        arr = (
+            ((output + 1.0) * 127.5)
+            .clamp(0, 255)
+            .round()
+            .to(torch.uint8)
+            .permute(1, 2, 0)
+            .contiguous()
+            .numpy()
+        )
+        return np.ascontiguousarray(arr)
+
+    def _should_omit_response_frames(self, request: pb2.UpscaleChunkRequest) -> bool:
+        return bool(request.display_only or self._omit_grpc_frames_when_viewing)
+
+    def _run_chunk(
+        self,
+        upsampler: RealESRGANUpsampler,
+        request: pb2.UpscaleChunkRequest,
+        *,
+        return_frames: bool,
+    ) -> _RunChunkResult:
+        decode_t0 = time.perf_counter()
+        frames = self._request_to_frames_rgb(request)
+        decode_ms = (time.perf_counter() - decode_t0) * 1000.0
+
+        executor = self._get_executor(
+            upsampler,
+            request.scale or self._default_scale,
+            int(frames.shape[1]),
+            int(frames.shape[2]),
+        )
+        if executor is not None:
+            return self._run_chunk_static(
+                executor,
+                request,
+                frames,
+                decode_ms=decode_ms,
+                return_frames=return_frames,
+            )
+        return self._run_chunk_dynamic(
+            upsampler,
+            request,
+            frames,
+            decode_ms=decode_ms,
+            return_frames=return_frames,
+        )
+
+    def _run_chunk_dynamic(
+        self,
+        upsampler: RealESRGANUpsampler,
+        request: pb2.UpscaleChunkRequest,
+        frames: np.ndarray,
+        *,
+        decode_ms: float,
+        return_frames: bool,
+    ) -> _RunChunkResult:
+        infer_t0 = time.perf_counter()
+        with self._gpu_lock:
+            output_frames = [
+                self._upsample_frame_rgb(upsampler, frame) for frame in frames
+            ]
+        _synchronize_device(self._device)
+        infer_ms = (time.perf_counter() - infer_t0) * 1000.0
+
+        frames_out = np.ascontiguousarray(np.stack(output_frames, axis=0))
+        num_frames, height, width, _channels = frames_out.shape
+        if self._viewer is not None:
+            self._viewer.enqueue_original_chunk(frames, elapsed_ms=infer_ms)
+            self._viewer.enqueue_upscaled_chunk(frames_out, elapsed_ms=infer_ms)
+
+        encode_t0 = time.perf_counter()
+        frames_rgb = frames_out.tobytes() if return_frames else b""
+        encode_ms = (time.perf_counter() - encode_t0) * 1000.0
+        logger.info(
+            "  chunk {} timing: decode_in={:.0f} ms infer={:.0f} ms "
+            "encode_out={:.0f} ms total={:.0f} ms (out {}×{})",
+            request.chunk_index,
+            decode_ms,
+            infer_ms,
+            encode_ms,
+            decode_ms + infer_ms + encode_ms,
+            width,
+            height,
+        )
+        return _RunChunkResult(
+            frames_rgb=frames_rgb,
+            frames_out=frames_out if return_frames else None,
+            num_frames=num_frames,
+            height=height,
+            width=width,
+            elapsed_ms=infer_ms,
+        )
+
+    def _run_chunk_static(
+        self,
+        executor: _StaticShapeExecutor,
+        request: pb2.UpscaleChunkRequest,
+        frames: np.ndarray,
+        *,
+        decode_ms: float,
+        return_frames: bool,
+    ) -> _RunChunkResult:
+        viewer = self._viewer
+        wants_viewer = viewer is not None
+        wants_cuda_jpegs = (
+            wants_viewer
+            and viewer.jpeg_backend in ("auto", "cuda")
+            and executor.device.type == "cuda"
+        )
+        output_frames: list[np.ndarray] = []
+        upscaled_jpegs: list[bytes] = []
+        model_ms_total = 0.0
+        infer_t0 = time.perf_counter()
+        with self._gpu_lock:
+            for index, frame in enumerate(frames):
+                result = executor.run_frame(frame)
+                if result.model_ms is not None:
+                    model_ms_total += result.model_ms
+                need_cpu = return_frames or (wants_viewer and not wants_cuda_jpegs)
+                if wants_cuda_jpegs:
+                    try:
+                        upscaled_jpegs.extend(
+                            self._encode_viewer_cuda_jpeg(result.output)
+                        )
+                    except Exception:
+                        if viewer is not None and viewer.jpeg_backend == "cuda":
+                            raise
+                        if not self._warned_cuda_viewer_fallback:
+                            logger.opt(exception=True).warning(
+                                "CUDA viewer JPEG encode unavailable; falling "
+                                "back to CPU/Pillow encoding"
+                            )
+                            self._warned_cuda_viewer_fallback = True
+                        wants_cuda_jpegs = False
+                        need_cpu = True
+                if need_cpu:
+                    output_frames.append(_output_tensor_to_rgb(result.output))
+                elif index == len(frames) - 1:
+                    # Keep the last model op ordered before the response when
+                    # neither gRPC nor viewer consumes the output tensor.
+                    _synchronize_device(executor.device)
+        infer_ms = (time.perf_counter() - infer_t0) * 1000.0
+        if model_ms_total > 0:
+            infer_ms = model_ms_total
+
+        frames_out: np.ndarray | None = None
+        if output_frames:
+            frames_out = np.ascontiguousarray(np.stack(output_frames, axis=0))
+        if viewer is not None:
+            viewer.enqueue_original_chunk(frames, elapsed_ms=infer_ms)
+            if upscaled_jpegs:
+                viewer.enqueue_upscaled_jpeg_chunk(
+                    upscaled_jpegs,
+                    elapsed_ms=infer_ms,
+                )
+            elif frames_out is not None:
+                viewer.enqueue_upscaled_chunk(frames_out, elapsed_ms=infer_ms)
+
+        encode_t0 = time.perf_counter()
+        frames_rgb = frames_out.tobytes() if return_frames and frames_out is not None else b""
+        encode_ms = (time.perf_counter() - encode_t0) * 1000.0
+        logger.info(
+            "  chunk {} timing: decode_in={:.0f} ms infer={:.0f} ms "
+            "encode_out={:.0f} ms total={:.0f} ms (out {}×{} static)",
+            request.chunk_index,
+            decode_ms,
+            infer_ms,
+            encode_ms,
+            decode_ms + infer_ms + encode_ms,
+            executor.output_width,
+            executor.output_height,
+        )
+        return _RunChunkResult(
+            frames_rgb=frames_rgb,
+            frames_out=frames_out if return_frames else None,
+            num_frames=int(frames.shape[0]),
+            height=executor.output_height,
+            width=executor.output_width,
+            elapsed_ms=infer_ms,
+        )
+
+    def _encode_viewer_cuda_jpeg(self, output: torch.Tensor) -> list[bytes]:
+        # encode_jpeg_cuda_tensor expects [1, 3, T, H, W] in [-1, 1].
+        video = output.mul(2.0).sub(1.0).unsqueeze(2)
+        return encode_jpeg_cuda_tensor(
+            video,
+            quality=self._viewer.jpeg_quality if self._viewer is not None else 90,
+            frame_stride=1,
+        )
+
+    def get_status(self, request, context):
+        with self._sessions_lock:
+            active = sorted([*self._sessions.keys(), *self._stream_sessions])
+        return pb2.StatusResponse(
+            ready=True,
+            device=str(self._device),
+            model_name=f"RealESRGAN/{self._model_name_for_scale(self._default_scale)}",
+            active_sessions=active,
+        )
+
+    def start_session(self, request, context):
+        session_id = request.session_id or str(uuid.uuid4())
+        scale = _resolve_scale(request.scale or self._default_scale)
+        model_name = self._model_name_for_scale(scale)
+        with self._sessions_lock:
+            if session_id in self._sessions:
+                return pb2.StartSessionResponse(
+                    session_id=session_id,
+                    success=False,
+                    error=f"session '{session_id}' already exists; call end_session first",
+                )
+        try:
+            self._get_upsampler(scale)
+            with self._sessions_lock:
+                self._sessions[session_id] = _Session(scale=scale, model_name=model_name)
+            logger.info("start_session {} (scale={} model={})", session_id, scale, model_name)
+            return pb2.StartSessionResponse(session_id=session_id, success=True)
+        except Exception as exc:
+            logger.exception("start_session failed")
+            return pb2.StartSessionResponse(success=False, error=str(exc))
+
+    def end_session(self, request, context):
+        with self._sessions_lock:
+            self._sessions.pop(request.session_id, None)
+        logger.info("end_session {}", request.session_id)
+        return pb2.EndSessionResponse(success=True)
+
+    def upscale_chunk(self, request, context):
+        with self._sessions_lock:
+            session = self._sessions.get(request.session_id)
+        if session is None:
+            message = f"session '{request.session_id}' not found; call start_session first"
+            if context is not None:
+                context.set_code(grpc.StatusCode.NOT_FOUND)
+                context.set_details(message)
+            return pb2.UpscaleChunkResponse(error=message)
+
+        try:
+            upsampler = self._get_upsampler(session.scale)
+            omit_frames = self._should_omit_response_frames(request)
+            result = self._run_chunk(
+                upsampler,
+                request,
+                return_frames=not omit_frames,
+            )
+            logger.info(
+                "upscale_chunk {} chunk={} T={} -> {:.0f} ms",
+                request.session_id,
+                request.chunk_index,
+                request.num_frames,
+                result.elapsed_ms,
+            )
+            return pb2.UpscaleChunkResponse(
+                session_id=request.session_id,
+                frames_rgb=result.frames_rgb,
+                num_frames=result.num_frames,
+                height=result.height,
+                width=result.width,
+                chunk_index=request.chunk_index,
+                elapsed_ms=result.elapsed_ms,
+                frames_omitted=omit_frames,
+            )
+        except Exception as exc:
+            logger.exception("upscale_chunk failed")
+            _context_set_error(context, str(exc))
+            return pb2.UpscaleChunkResponse(error=str(exc))
+
+    def upscale_video(self, request_iterator, context):
+        session_id = str(uuid.uuid4())
+        stream_started = False
+        scale = self._default_scale
+        upsampler: RealESRGANUpsampler | None = None
+        try:
+            for request in request_iterator:
+                if not stream_started:
+                    session_id = request.session_id or session_id
+                    scale = _resolve_scale(request.scale or self._default_scale)
+                    upsampler = self._get_upsampler(scale)
+                    with self._sessions_lock:
+                        self._stream_sessions.add(session_id)
+                    stream_started = True
+                    logger.info(
+                        "upscale_video stream {} ({}×{} scale={}) queue_depth={}",
+                        session_id,
+                        request.height or request.input_height,
+                        request.width or request.input_width,
+                        scale,
+                        self._stream_queue_depth,
+                    )
+                assert upsampler is not None
+                omit_frames = self._should_omit_response_frames(request)
+                result = self._run_chunk(
+                    upsampler,
+                    request,
+                    return_frames=not omit_frames,
+                )
+                logger.info(
+                    "upscale_video {} chunk={} T={} infer={:.0f} ms",
+                    session_id,
+                    request.chunk_index,
+                    request.num_frames,
+                    result.elapsed_ms,
+                )
+                yield pb2.UpscaleChunkResponse(
+                    session_id=session_id,
+                    frames_rgb=result.frames_rgb,
+                    num_frames=result.num_frames,
+                    height=result.height,
+                    width=result.width,
+                    chunk_index=request.chunk_index,
+                    elapsed_ms=result.elapsed_ms,
+                    frames_omitted=omit_frames,
+                )
+        except Exception as exc:
+            logger.exception("upscale_video failed")
+            yield pb2.UpscaleChunkResponse(session_id=session_id, error=str(exc))
+        finally:
+            with self._sessions_lock:
+                self._stream_sessions.discard(session_id)
+            logger.info("upscale_video stream {}: completed", session_id)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Real-ESRGAN uplift gRPC server")
+    parser.add_argument("--port", type=int, default=DEFAULT_PORT)
+    parser.add_argument("--workers", type=int, default=4, help="gRPC thread pool size")
+    parser.add_argument("--max_message_mb", type=int, default=DEFAULT_MAX_MESSAGE_MB)
+    parser.add_argument("--default_scale", type=int, choices=[2, 4], default=2)
+    parser.add_argument("--model_name", type=str, default=None)
+    parser.add_argument("--model_path", type=Path, default=None)
+    parser.add_argument("--tile", type=int, default=0)
+    parser.add_argument("--tile_pad", type=int, default=10)
+    parser.add_argument("--pre_pad", type=int, default=10)
+    parser.add_argument("--fp32", action="store_true", help="Disable fp16 on CUDA.")
+    parser.add_argument("--compile", action="store_true", help="Use torch.compile().")
+    parser.add_argument("--compile_mode", default="reduce-overhead")
+    parser.add_argument("--device", default=None)
+    parser.add_argument(
+        "--stream_queue_depth",
+        type=int,
+        default=DEFAULT_STREAM_QUEUE_DEPTH,
+        help="Reported per-stream queue depth; GPU work is serialized by a device lock.",
+    )
+    parser.add_argument("--no_warmup", action="store_true")
+    parser.add_argument("--warmup_height", type=int, default=64)
+    parser.add_argument("--warmup_width", type=int, default=64)
+    parser.add_argument(
+        "--warmup_frames",
+        type=int,
+        default=3,
+        help=(
+            "Number of same-shape frames to run during warmup. "
+            "TorchInductor reduce-overhead may need multiple calls before "
+            "the CUDA graph path is fully settled."
+        ),
+    )
+    parser.add_argument(
+        "--viewer_port",
+        type=int,
+        default=0,
+        help="Start an HTTP MJPEG viewer on this port. Disabled when 0.",
+    )
+    parser.add_argument("--viewer_host", default="0.0.0.0")
+    parser.add_argument("--viewer_jpeg_quality", type=int, default=DEFAULT_VIEWER_JPEG_QUALITY)
+    parser.add_argument(
+        "--viewer_jpeg_backend",
+        choices=["auto", "cuda", "pillow"],
+        default=DEFAULT_VIEWER_JPEG_BACKEND,
+    )
+    parser.add_argument(
+        "--viewer_chunk_queue_depth",
+        type=int,
+        default=DEFAULT_VIEWER_CHUNK_QUEUE_DEPTH,
+    )
+    parser.add_argument("--viewer_max_fps", type=float, default=DEFAULT_VIEWER_MAX_FPS)
+    parser.add_argument(
+        "--viewer_playback_fps",
+        type=float,
+        default=DEFAULT_VIEWER_PLAYBACK_FPS,
+        help=(
+            "Steady HTTP viewer playback FPS. If completed chunks back up, "
+            "the viewer gradually catches up, capped by --viewer_max_fps. "
+            "Set to 0 to pace from model elapsed time."
+        ),
+    )
+    parser.add_argument("--viewer_frame_stride", type=int, default=DEFAULT_VIEWER_FRAME_STRIDE)
+    parser.add_argument(
+        "--viewer_return_grpc_frames",
+        action="store_true",
+        help="When viewer is enabled, still include raw RGB frames in gRPC responses.",
+    )
+    args = parser.parse_args()
+
+    if not 1 <= args.viewer_jpeg_quality <= 100:
+        parser.error("--viewer_jpeg_quality must be between 1 and 100")
+    if args.viewer_frame_stride < 1:
+        parser.error("--viewer_frame_stride must be at least 1")
+    if args.viewer_playback_fps < 0:
+        parser.error("--viewer_playback_fps must be non-negative")
+    if args.warmup_frames < 1:
+        parser.error("--warmup_frames must be at least 1")
+
+    max_bytes = args.max_message_mb * 1024 * 1024
+    server = grpc.server(
+        futures.ThreadPoolExecutor(max_workers=args.workers),
+        options=[
+            ("grpc.max_send_message_length", max_bytes),
+            ("grpc.max_receive_message_length", max_bytes),
+        ],
+    )
+    addr = f"[::]:{args.port}"
+    bound_port = server.add_insecure_port(addr)
+    if bound_port == 0:
+        raise RuntimeError(
+            f"Failed to bind gRPC server to {addr}; another process is likely "
+            "already listening on that port."
+        )
+
+    viewer: StreamingViewer | None = None
+    if args.viewer_port:
+        viewer = StreamingViewer(
+            host=args.viewer_host,
+            port=args.viewer_port,
+            jpeg_quality=args.viewer_jpeg_quality,
+            jpeg_backend=args.viewer_jpeg_backend,
+            chunk_queue_depth=args.viewer_chunk_queue_depth,
+            max_fps=args.viewer_max_fps,
+            playback_fps=args.viewer_playback_fps,
+            frame_stride=args.viewer_frame_stride,
+        )
+        viewer.start()
+
+    servicer = RealESRGANUplift(
+        default_scale=args.default_scale,
+        model_name=args.model_name,
+        model_path=args.model_path,
+        tile=args.tile,
+        tile_pad=args.tile_pad,
+        pre_pad=args.pre_pad,
+        half=not args.fp32,
+        compile_model=args.compile,
+        compile_mode=args.compile_mode,
+        device=args.device,
+        stream_queue_depth=args.stream_queue_depth,
+        viewer=viewer,
+        omit_grpc_frames_when_viewing=not args.viewer_return_grpc_frames,
+        warmup_height=args.warmup_height,
+        warmup_width=args.warmup_width,
+        warmup_frames=args.warmup_frames,
+        warmup=not args.no_warmup,
+    )
+    pb2_grpc.add_VideoUpliftServicer_to_server(servicer, server)
+    server.start()
+    logger.info("Server listening on [::]:{} (max msg {} MB)", bound_port, args.max_message_mb)
+
+    stop_requested = threading.Event()
+
+    def _on_signal(signum: int, _frame: object) -> None:
+        if stop_requested.is_set():
+            logger.warning("Second signal {} received; forcing exit.", signum)
+            os._exit(130)
+        logger.info("Signal {} received; requesting shutdown.", signum)
+        stop_requested.set()
+
+    signal.signal(signal.SIGINT, _on_signal)
+    signal.signal(signal.SIGTERM, _on_signal)
+
+    try:
+        while not stop_requested.wait(timeout=0.5):
+            pass
+    finally:
+        logger.info("Shutting down ...")
+        stopped = server.stop(grace=5)
+        if not stopped.wait(timeout=10):
+            logger.warning("gRPC server.stop did not complete within 10s; exiting anyway.")
+        if viewer is not None:
+            try:
+                viewer.stop()
+            except Exception:
+                logger.exception("viewer.stop() raised; continuing shutdown")
+        sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/integrations/realesrgan/realesrgan/postprocess.py b/integrations/realesrgan/realesrgan/postprocess.py
new file mode 100644
index 000000000..c5b718bf4
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/postprocess.py
@@ -0,0 +1,149 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""FlashDreams post-processing wrapper for Real-ESRGAN."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Literal
+
+import torch
+
+from flashdreams.infra.postprocess import (
+    VideoChunk,
+    VideoPostProcessor,
+    VideoPostProcessorConfig,
+    VideoPostProcessorSession,
+    VideoSpec,
+    to_bvtchw,
+)
+from realesrgan.upsampler import RealESRGANUpsampler, default_model_name
+
+
+@dataclass(kw_only=True)
+class RealESRGANPostProcessorConfig(VideoPostProcessorConfig):
+    """Post-process generated RGB video with Real-ESRGAN."""
+
+    _target: type["RealESRGANPostProcessor"] = field(
+        default_factory=lambda: RealESRGANPostProcessor
+    )
+
+    scale: Literal[2, 4] = 2
+    """Spatial upsample factor."""
+
+    model_name: str | None = None
+    """Real-ESRGAN checkpoint name. Defaults to the general model for scale."""
+
+    model_path: str | Path | None = None
+    """Optional local checkpoint path."""
+
+    tile: int = 0
+    """Input tile size. ``0`` processes each frame as a single tensor."""
+
+    tile_pad: int = 10
+    """Input tile overlap in pixels."""
+
+    pre_pad: int = 10
+    """Reflection padding applied before model inference."""
+
+    half: bool = True
+    """Use fp16 on CUDA."""
+
+    compile_model: bool = False
+    """Compile the Real-ESRGAN model with ``torch.compile``."""
+
+    compile_mode: str = "reduce-overhead"
+    """Mode passed to ``torch.compile`` when :attr:`compile_model` is enabled."""
+
+    device: str = "cuda"
+    """Torch device used by the Real-ESRGAN model."""
+
+    def output_spec(self, input_spec: VideoSpec) -> VideoSpec:
+        """Return the spatially upsampled output specification."""
+        return VideoSpec(
+            height=input_spec.height * self.scale,
+            width=input_spec.width * self.scale,
+            fps=input_spec.fps,
+            channels=input_spec.channels,
+        )
+
+
+class RealESRGANPostProcessor(VideoPostProcessor[RealESRGANPostProcessorConfig]):
+    """Factory for Real-ESRGAN post-processing sessions."""
+
+    def start(self, spec: VideoSpec) -> VideoPostProcessorSession:
+        """Start one Real-ESRGAN post-processing stream."""
+        return _RealESRGANPostProcessorSession(self.config)
+
+
+POSTPROCESS_PRESET_REALESRGAN = RealESRGANPostProcessorConfig()
+"""Default Real-ESRGAN post-processing preset."""
+
+
+class _RealESRGANPostProcessorSession(VideoPostProcessorSession):
+    def __init__(self, config: RealESRGANPostProcessorConfig) -> None:
+        self._config = config
+        self._upsampler: RealESRGANUpsampler | None = None
+
+    @torch.no_grad()
+    def process(self, chunk: VideoChunk) -> list[VideoChunk]:
+        canonical = to_bvtchw(chunk.tensor, layout=chunk.layout)
+        batch, views, _, channels, _, _ = canonical.shape
+        if channels != 3:
+            raise ValueError(
+                f"Real-ESRGAN expects RGB chunks; got {channels} channels."
+            )
+        upsampler = self._ensure_upsampler()
+        outputs = []
+        for batch_idx in range(batch):
+            view_outputs = []
+            for view_idx in range(views):
+                view_outputs.append(
+                    upsampler.upsample_video_tensor(canonical[batch_idx, view_idx])
+                )
+            outputs.append(torch.stack(view_outputs, dim=0))
+        output = torch.stack(outputs, dim=0)
+        return [
+            VideoChunk(
+                tensor=output,
+                layout="bvtchw",
+                metadata={**chunk.metadata, "source": "realesrgan"},
+            )
+        ]
+
+    def flush(self) -> list[VideoChunk]:
+        """Real-ESRGAN is frame-local and keeps no buffered tail."""
+        return []
+
+    def _ensure_upsampler(self) -> RealESRGANUpsampler:
+        if self._upsampler is None:
+            model_name = self._config.model_name or default_model_name(
+                self._config.scale
+            )
+            self._upsampler = RealESRGANUpsampler(
+                scale=self._config.scale,
+                model_name=model_name,
+                model_path=self._config.model_path,
+                tile=self._config.tile,
+                tile_pad=self._config.tile_pad,
+                pre_pad=self._config.pre_pad,
+                half=self._config.half,
+                compile_model=self._config.compile_model,
+                compile_mode=self._config.compile_mode,
+                device=self._config.device,
+            )
+        return self._upsampler
diff --git a/integrations/realesrgan/realesrgan/upsampler.py b/integrations/realesrgan/realesrgan/upsampler.py
new file mode 100644
index 000000000..787dd899e
--- /dev/null
+++ b/integrations/realesrgan/realesrgan/upsampler.py
@@ -0,0 +1,477 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Real-ESRGAN checkpoint loading and tensor upsampling."""
+
+from __future__ import annotations
+
+import math
+import os
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Literal
+
+import cv2
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from flashdreams.core.io.download import download_to_cache
+from realesrgan.archs import RRDBNet, SRVGGNetCompact
+
+RealESRGANModelName = Literal[
+    "RealESRGAN_x4plus",
+    "RealESRNet_x4plus",
+    "RealESRGAN_x4plus_anime_6B",
+    "RealESRGAN_x2plus",
+    "realesr-animevideov3",
+    "realesr-general-x4v3",
+]
+
+MODEL_URLS: dict[str, str] = {
+    "RealESRGAN_x4plus": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth",
+    "RealESRNet_x4plus": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/RealESRNet_x4plus.pth",
+    "RealESRGAN_x4plus_anime_6B": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth",
+    "RealESRGAN_x2plus": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth",
+    "realesr-animevideov3": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth",
+    "realesr-general-x4v3": "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth",
+}
+"""Public Real-ESRGAN checkpoint URLs. Files are cached at runtime."""
+
+MODEL_CONFIGS: dict[str, dict[str, int | str]] = {
+    "RealESRGAN_x4plus": {
+        "arch": "RRDBNet",
+        "num_in_ch": 3,
+        "num_out_ch": 3,
+        "num_feat": 64,
+        "num_block": 23,
+        "num_grow_ch": 32,
+        "scale": 4,
+    },
+    "RealESRNet_x4plus": {
+        "arch": "RRDBNet",
+        "num_in_ch": 3,
+        "num_out_ch": 3,
+        "num_feat": 64,
+        "num_block": 23,
+        "num_grow_ch": 32,
+        "scale": 4,
+    },
+    "RealESRGAN_x4plus_anime_6B": {
+        "arch": "RRDBNet",
+        "num_in_ch": 3,
+        "num_out_ch": 3,
+        "num_feat": 64,
+        "num_block": 6,
+        "num_grow_ch": 32,
+        "scale": 4,
+    },
+    "RealESRGAN_x2plus": {
+        "arch": "RRDBNet",
+        "num_in_ch": 3,
+        "num_out_ch": 3,
+        "num_feat": 64,
+        "num_block": 23,
+        "num_grow_ch": 32,
+        "scale": 2,
+    },
+    "realesr-animevideov3": {
+        "arch": "SRVGGNetCompact",
+        "num_in_ch": 3,
+        "num_out_ch": 3,
+        "num_feat": 64,
+        "num_conv": 16,
+        "upscale": 4,
+        "act_type": "prelu",
+        "scale": 4,
+    },
+    "realesr-general-x4v3": {
+        "arch": "SRVGGNetCompact",
+        "num_in_ch": 3,
+        "num_out_ch": 3,
+        "num_feat": 64,
+        "num_conv": 32,
+        "upscale": 4,
+        "act_type": "prelu",
+        "scale": 4,
+    },
+}
+"""Model architecture table for public Real-ESRGAN checkpoints."""
+
+CACHE_DIR = (
+    Path(os.path.expanduser(os.getenv("FLASHDREAMS_CACHE_DIR", "~/.cache/flashdreams")))
+    / "realesrgan"
+)
+"""User-writable cache directory for Real-ESRGAN checkpoint downloads."""
+
+
+@dataclass(frozen=True, kw_only=True)
+class RealESRGANFrameProfile:
+    """Per-frame timing captured while producing a Real-ESRGAN output."""
+
+    model_ms: float | None
+    """CUDA-event time for the model/tiled-model call, excluding CPU I/O."""
+
+
+def default_model_name(scale: Literal[2, 4]) -> RealESRGANModelName:
+    """Return the default public Real-ESRGAN model for ``scale``."""
+    return "RealESRGAN_x2plus" if scale == 2 else "RealESRGAN_x4plus"
+
+
+def create_model(model_name: str) -> tuple[nn.Module, int]:
+    """Create the architecture for a named Real-ESRGAN checkpoint."""
+    if model_name not in MODEL_CONFIGS:
+        raise ValueError(
+            f"Unknown Real-ESRGAN model {model_name!r}; "
+            f"available models: {sorted(MODEL_CONFIGS)}"
+        )
+    config = MODEL_CONFIGS[model_name]
+    arch = str(config["arch"])
+    scale = int(config["scale"])
+    if arch == "RRDBNet":
+        return (
+            RRDBNet(
+                num_in_ch=int(config["num_in_ch"]),
+                num_out_ch=int(config["num_out_ch"]),
+                num_feat=int(config["num_feat"]),
+                num_block=int(config["num_block"]),
+                num_grow_ch=int(config["num_grow_ch"]),
+                scale=scale,
+            ),
+            scale,
+        )
+    if arch == "SRVGGNetCompact":
+        return (
+            SRVGGNetCompact(
+                num_in_ch=int(config["num_in_ch"]),
+                num_out_ch=int(config["num_out_ch"]),
+                num_feat=int(config["num_feat"]),
+                num_conv=int(config["num_conv"]),
+                upscale=int(config["upscale"]),
+                act_type=str(config["act_type"]),
+            ),
+            scale,
+        )
+    raise ValueError(f"Unknown Real-ESRGAN architecture {arch!r}.")
+
+
+def resolve_model_path(model_name: str, model_path: str | Path | None = None) -> Path:
+    """Return a local checkpoint path, downloading public weights if needed."""
+    if model_path is not None:
+        return Path(model_path).expanduser()
+    if model_name not in MODEL_URLS:
+        raise ValueError(f"No public checkpoint URL for model {model_name!r}.")
+    return download_to_cache(MODEL_URLS[model_name], cache_dir=CACHE_DIR)
+
+
+def load_weights(model: nn.Module, checkpoint_path: str | Path) -> None:
+    """Load Real-ESRGAN weights from ``checkpoint_path`` into ``model``."""
+    checkpoint = torch.load(
+        Path(checkpoint_path).expanduser(),
+        map_location="cpu",
+        weights_only=False,
+    )
+    if isinstance(checkpoint, dict) and "params_ema" in checkpoint:
+        state = checkpoint["params_ema"]
+    elif isinstance(checkpoint, dict) and "params" in checkpoint:
+        state = checkpoint["params"]
+    else:
+        state = checkpoint
+    model.load_state_dict(state, strict=True)
+
+
+class RealESRGANUpsampler:
+    """Real-ESRGAN RGB tensor upsampler.
+
+    Args:
+        scale: Output upsample factor.
+        model_name: Public Real-ESRGAN model name. Defaults to the general
+            x2 or x4 RRDBNet model matching ``scale``.
+        model_path: Optional local checkpoint path.
+        tile: Optional input tile size. ``0`` runs each frame as one tensor.
+        tile_pad: Input overlap on every tile edge.
+        pre_pad: Reflection padding applied before model inference.
+        half: Use fp16 on CUDA.
+        compile_model: Compile the model with ``torch.compile`` after loading.
+        compile_mode: Mode passed to ``torch.compile``.
+        device: Torch device string or object.
+    """
+
+    def __init__(
+        self,
+        *,
+        scale: Literal[2, 4] = 2,
+        model_name: str | None = None,
+        model_path: str | Path | None = None,
+        tile: int = 0,
+        tile_pad: int = 10,
+        pre_pad: int = 10,
+        half: bool = True,
+        compile_model: bool = False,
+        compile_mode: str = "reduce-overhead",
+        device: str | torch.device | None = None,
+        load_checkpoint: bool = True,
+    ) -> None:
+        self.scale = scale
+        self.model_name = model_name or default_model_name(scale)
+        self.tile = tile
+        self.tile_pad = tile_pad
+        self.pre_pad = pre_pad
+        self.half = half
+        self.compile_model = compile_model
+        self.compile_mode = compile_mode
+        self.device = torch.device(
+            device
+            if device is not None
+            else ("cuda" if torch.cuda.is_available() else "cpu")
+        )
+        self.model, self.model_scale = create_model(self.model_name)
+        if self.model_scale != scale:
+            raise ValueError(
+                f"Model {self.model_name!r} has native scale {self.model_scale}; "
+                f"requested scale {scale}."
+            )
+        if load_checkpoint:
+            load_weights(self.model, resolve_model_path(self.model_name, model_path))
+        self.model.eval().to(self.device)
+        if self.half and self.device.type == "cuda":
+            self.model.half()
+        if self.compile_model:
+            self.model = torch.compile(self.model, mode=self.compile_mode)
+        self.dtype = (
+            torch.float16 if self.half and self.device.type == "cuda" else torch.float32
+        )
+
+    @torch.no_grad()
+    def upsample_video_tensor(self, video: torch.Tensor) -> torch.Tensor:
+        """Upsample ``video`` shaped ``[T, C, H, W]`` in ``[-1, 1]``."""
+        if video.ndim != 4 or video.shape[1] != 3:
+            raise ValueError(
+                f"Expected [T, 3, H, W] video tensor; got {tuple(video.shape)}"
+            )
+        frames = [self.upsample_frame_tensor(frame) for frame in video]
+        return torch.stack(frames, dim=0)
+
+    @torch.no_grad()
+    def upsample_frame_tensor(self, frame: torch.Tensor) -> torch.Tensor:
+        """Upsample one RGB frame shaped ``[3, H, W]`` in ``[-1, 1]``."""
+        output, _profile = self.upsample_frame_tensor_profiled(frame)
+        return output
+
+    @torch.no_grad()
+    def upsample_frame_tensor_profiled(
+        self, frame: torch.Tensor
+    ) -> tuple[torch.Tensor, RealESRGANFrameProfile]:
+        """Upsample one RGB frame and return model-only timing metadata."""
+        if frame.ndim != 3 or frame.shape[0] != 3:
+            raise ValueError(
+                f"Expected [3, H, W] frame tensor; got {tuple(frame.shape)}"
+            )
+        image = ((frame.to(self.device, dtype=self.dtype) + 1.0) * 0.5).clamp(0, 1)
+        output, profile = self._process_image_profiled(image.unsqueeze(0))
+        output = (output.squeeze(0).float().cpu().clamp(0, 1) * 2.0) - 1.0
+        return output, profile
+
+    @torch.no_grad()
+    def upsample_bgr_image(self, image: np.ndarray) -> tuple[np.ndarray, str]:
+        """Upsample an OpenCV image array and return ``(image, mode)``."""
+        output, mode, _profile = self.upsample_bgr_image_profiled(image)
+        return output, mode
+
+    @torch.no_grad()
+    def upsample_bgr_image_profiled(
+        self, image: np.ndarray
+    ) -> tuple[np.ndarray, str, RealESRGANFrameProfile]:
+        """Upsample an OpenCV image array and return model timing metadata."""
+        mode = "RGB"
+        alpha: np.ndarray | None = None
+        if image.ndim == 2:
+            mode = "L"
+            image_rgb = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
+        elif image.shape[2] == 4:
+            mode = "RGBA"
+            alpha = image[:, :, 3]
+            image_rgb = cv2.cvtColor(image[:, :, :3], cv2.COLOR_BGR2RGB)
+        else:
+            image_rgb = cv2.cvtColor(image[:, :, :3], cv2.COLOR_BGR2RGB)
+
+        max_range = 65535.0 if image.dtype == np.uint16 else 255.0
+        tensor = torch.from_numpy(
+            (image_rgb.astype(np.float32) / max_range).transpose(2, 0, 1)
+        )
+        output, profile = self.upsample_frame_tensor_profiled(tensor * 2.0 - 1.0)
+        output = (output + 1.0) * 0.5
+        if mode == "L":
+            return _rgb_tensor_to_gray_image(output, image.dtype), mode, profile
+        output_bgr = _rgb_tensor_to_bgr_image(output, image.dtype)
+        if mode == "RGBA" and alpha is not None:
+            alpha_out = cv2.resize(
+                alpha,
+                (output_bgr.shape[1], output_bgr.shape[0]),
+                interpolation=cv2.INTER_LINEAR,
+            )
+            output_bgr = _bgr_to_bgra(output_bgr, alpha_out)
+
+        return output_bgr, mode, profile
+
+    def _process_image(self, image: torch.Tensor) -> torch.Tensor:
+        output, _profile = self._process_image_profiled(image)
+        return output
+
+    def _process_image_profiled(
+        self, image: torch.Tensor
+    ) -> tuple[torch.Tensor, RealESRGANFrameProfile]:
+        padded, crop = self._pad_input(image)
+        start: torch.cuda.Event | None = None
+        end: torch.cuda.Event | None = None
+        if self.device.type == "cuda":
+            if self.compile_model:
+                torch.compiler.cudagraph_mark_step_begin()
+            start = torch.cuda.Event(enable_timing=True)
+            end = torch.cuda.Event(enable_timing=True)
+            start.record()
+        output = (
+            self._process_tiled(padded)
+            if self.tile > 0
+            else self.model(padded.to(dtype=self.dtype))
+        )
+        if end is not None:
+            end.record()
+        output = self._crop_output(output, crop)
+        model_ms: float | None = None
+        if start is not None and end is not None:
+            torch.cuda.synchronize(self.device)
+            model_ms = float(start.elapsed_time(end))
+        return output, RealESRGANFrameProfile(model_ms=model_ms)
+
+    def _pad_input(self, image: torch.Tensor) -> tuple[torch.Tensor, tuple[int, int]]:
+        if self.pre_pad:
+            image = F.pad(image, (0, self.pre_pad, 0, self.pre_pad), "reflect")
+        mod_scale = 2 if self.model_scale == 2 else None
+        mod_pad_h = 0
+        mod_pad_w = 0
+        if mod_scale is not None:
+            _, _, height, width = image.shape
+            mod_pad_h = (mod_scale - height % mod_scale) % mod_scale
+            mod_pad_w = (mod_scale - width % mod_scale) % mod_scale
+            if mod_pad_h or mod_pad_w:
+                image = F.pad(image, (0, mod_pad_w, 0, mod_pad_h), "reflect")
+        return image, (
+            (self.pre_pad + mod_pad_h) * self.model_scale,
+            (self.pre_pad + mod_pad_w) * self.model_scale,
+        )
+
+    def _crop_output(self, output: torch.Tensor, crop: tuple[int, int]) -> torch.Tensor:
+        crop_h, crop_w = crop
+        if crop_h:
+            output = output[:, :, :-crop_h, :]
+        if crop_w:
+            output = output[:, :, :, :-crop_w]
+        return output
+
+    def _process_tiled(self, image: torch.Tensor) -> torch.Tensor:
+        batch, channels, height, width = image.shape
+        output = image.new_zeros(
+            batch,
+            channels,
+            height * self.model_scale,
+            width * self.model_scale,
+        )
+        tiles_x = math.ceil(width / self.tile)
+        tiles_y = math.ceil(height / self.tile)
+        for y in range(tiles_y):
+            for x in range(tiles_x):
+                in_x0 = x * self.tile
+                in_y0 = y * self.tile
+                in_x1 = min(in_x0 + self.tile, width)
+                in_y1 = min(in_y0 + self.tile, height)
+                pad_x0 = max(in_x0 - self.tile_pad, 0)
+                pad_y0 = max(in_y0 - self.tile_pad, 0)
+                pad_x1 = min(in_x1 + self.tile_pad, width)
+                pad_y1 = min(in_y1 + self.tile_pad, height)
+                tile = image[:, :, pad_y0:pad_y1, pad_x0:pad_x1]
+                output_tile = self.model(tile.to(dtype=self.dtype))
+
+                out_x0 = in_x0 * self.model_scale
+                out_y0 = in_y0 * self.model_scale
+                out_x1 = in_x1 * self.model_scale
+                out_y1 = in_y1 * self.model_scale
+                tile_x0 = (in_x0 - pad_x0) * self.model_scale
+                tile_y0 = (in_y0 - pad_y0) * self.model_scale
+                tile_x1 = tile_x0 + (in_x1 - in_x0) * self.model_scale
+                tile_y1 = tile_y0 + (in_y1 - in_y0) * self.model_scale
+                output[:, :, out_y0:out_y1, out_x0:out_x1] = output_tile[
+                    :, :, tile_y0:tile_y1, tile_x0:tile_x1
+                ]
+        return output
+
+
+def write_bgr_image(path: str | Path, image: np.ndarray) -> None:
+    """Write ``image`` to ``path`` and raise if OpenCV rejects it."""
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if not cv2.imwrite(str(path), image):
+        raise RuntimeError(f"Failed to write image to {path}")
+
+
+def _rgb_tensor_to_bgr_image(tensor: torch.Tensor, dtype: np.dtype) -> np.ndarray:
+    """Convert a CPU ``[3, H, W]`` RGB float tensor in ``[0, 1]`` to BGR image."""
+    image_dtype = _output_image_dtype(dtype)
+    scale = _dtype_max(image_dtype)
+    torch_dtype = _torch_dtype(image_dtype)
+    return (
+        tensor.clamp(0, 1)
+        .mul(scale)
+        .round()
+        .to(torch_dtype)[[2, 1, 0]]
+        .permute(1, 2, 0)
+        .contiguous()
+        .numpy()
+    )
+
+
+def _rgb_tensor_to_gray_image(tensor: torch.Tensor, dtype: np.dtype) -> np.ndarray:
+    """Convert a CPU ``[3, H, W]`` RGB float tensor in ``[0, 1]`` to grayscale."""
+    image_dtype = _output_image_dtype(dtype)
+    scale = _dtype_max(image_dtype)
+    torch_dtype = _torch_dtype(image_dtype)
+    tensor = tensor.clamp(0, 1)
+    gray = tensor[0] * 0.299 + tensor[1] * 0.587 + tensor[2] * 0.114
+    return gray.mul(scale).round().to(torch_dtype).contiguous().numpy()
+
+
+def _bgr_to_bgra(bgr: np.ndarray, alpha: np.ndarray) -> np.ndarray:
+    """Append ``alpha`` to a BGR image without routing through float color convert."""
+    output = np.empty((*bgr.shape[:2], 4), dtype=bgr.dtype)
+    output[:, :, :3] = bgr
+    output[:, :, 3] = alpha.astype(bgr.dtype, copy=False)
+    return output
+
+
+def _output_image_dtype(dtype: np.dtype) -> np.dtype:
+    return (
+        np.dtype(np.uint16)
+        if np.dtype(dtype) == np.dtype(np.uint16)
+        else np.dtype(np.uint8)
+    )
+
+
+def _dtype_max(dtype: np.dtype) -> float:
+    return 65535.0 if np.dtype(dtype) == np.dtype(np.uint16) else 255.0
+
+
+def _torch_dtype(dtype: np.dtype) -> torch.dtype:
+    return torch.uint16 if np.dtype(dtype) == np.dtype(np.uint16) else torch.uint8
diff --git a/integrations/realesrgan/scripts/feed_realesrgan_uplift_video.sh b/integrations/realesrgan/scripts/feed_realesrgan_uplift_video.sh
new file mode 100755
index 000000000..b16bf7dc6
--- /dev/null
+++ b/integrations/realesrgan/scripts/feed_realesrgan_uplift_video.sh
@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+if [[ $# -lt 1 ]]; then
+  echo "usage: $0 INPUT_VIDEO [flashdreams-feed-frames args...]" >&2
+  exit 2
+fi
+
+INPUT_VIDEO="$1"
+shift
+
+SERVER="${SERVER:-localhost:50052}"
+SCALE="${SCALE:-2}"
+MAX_CHUNKS="${MAX_CHUNKS:-20}"
+TARGET_FPS="${TARGET_FPS:-30}"
+CONTINUOUS_CHUNK_FRAMES="${CONTINUOUS_CHUNK_FRAMES:-8}"
+INPUT_FORMAT="${INPUT_FORMAT:-jpeg}"
+INPUT_JPEG_QUALITY="${INPUT_JPEG_QUALITY:-90}"
+LOOP_MODE="${LOOP_MODE:-chunk}"
+
+uv run --package flashdreams-realesrgan flashdreams-feed-frames \
+  --server "${SERVER}" \
+  --input "${INPUT_VIDEO}" \
+  --scale "${SCALE}" \
+  --max_chunks "${MAX_CHUNKS}" \
+  --target_fps "${TARGET_FPS}" \
+  --continuous_chunk_frames "${CONTINUOUS_CHUNK_FRAMES}" \
+  --loop_mode "${LOOP_MODE}" \
+  --input_format "${INPUT_FORMAT}" \
+  --input_jpeg_quality "${INPUT_JPEG_QUALITY}" \
+  "$@"
diff --git a/integrations/realesrgan/scripts/gpu_smoke.py b/integrations/realesrgan/scripts/gpu_smoke.py
new file mode 100644
index 000000000..69a32d2af
--- /dev/null
+++ b/integrations/realesrgan/scripts/gpu_smoke.py
@@ -0,0 +1,100 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Run a checkpoint-backed Real-ESRGAN GPU smoke test."""
+
+from __future__ import annotations
+
+import argparse
+import time
+from typing import Literal, cast
+
+import torch
+
+from realesrgan.upsampler import RealESRGANUpsampler, default_model_name
+
+
+def main() -> None:
+    """Load a Real-ESRGAN checkpoint and run one synthetic RGB frame."""
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--scale", type=int, choices=(2, 4), default=2)
+    parser.add_argument("--model-name", type=str, default=None)
+    parser.add_argument("--size", type=int, default=16)
+    parser.add_argument("--tile", type=int, default=0)
+    parser.add_argument("--fp32", action="store_true")
+    parser.add_argument("--device", type=str, default="cuda")
+    args = parser.parse_args()
+
+    if args.device.startswith("cuda") and not torch.cuda.is_available():
+        raise SystemExit("CUDA is not available.")
+
+    scale = cast(Literal[2, 4], args.scale)
+    model_name = args.model_name or default_model_name(scale)
+    model_start = time.perf_counter()
+    upsampler = RealESRGANUpsampler(
+        scale=scale,
+        model_name=model_name,
+        tile=args.tile,
+        pre_pad=0,
+        half=not args.fp32,
+        device=args.device,
+    )
+    _synchronize(args.device)
+    model_elapsed = time.perf_counter() - model_start
+
+    if args.device.startswith("cuda"):
+        torch.cuda.reset_peak_memory_stats(args.device)
+    frame = torch.zeros(3, args.size, args.size)
+    infer_start = time.perf_counter()
+    output = upsampler.upsample_frame_tensor(frame)
+    _synchronize(args.device)
+    infer_elapsed = time.perf_counter() - infer_start
+
+    expected_shape = (3, args.size * scale, args.size * scale)
+    if output.shape != expected_shape:
+        raise RuntimeError(
+            f"Expected output shape {expected_shape}, got {output.shape}"
+        )
+    if not torch.isfinite(output).all():
+        raise RuntimeError("Real-ESRGAN output contains non-finite values.")
+
+    peak_mb = (
+        torch.cuda.max_memory_allocated(args.device) / 1024 / 1024
+        if args.device.startswith("cuda")
+        else 0.0
+    )
+    device_name = (
+        torch.cuda.get_device_name(args.device)
+        if args.device.startswith("cuda")
+        else args.device
+    )
+    print(f"device={device_name}")
+    print(
+        f"output_shape={tuple(output.shape)} dtype={output.dtype} "
+        f"range=({float(output.min()):.4f}, {float(output.max()):.4f})"
+    )
+    print(
+        f"model_load_s={model_elapsed:.3f} "
+        f"infer_s={infer_elapsed:.3f} peak_allocated_mb={peak_mb:.1f}"
+    )
+
+
+def _synchronize(device: str) -> None:
+    if device.startswith("cuda"):
+        torch.cuda.synchronize(device)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/integrations/realesrgan/scripts/run_realesrgan_uplift_server.sh b/integrations/realesrgan/scripts/run_realesrgan_uplift_server.sh
new file mode 100755
index 000000000..73f996db9
--- /dev/null
+++ b/integrations/realesrgan/scripts/run_realesrgan_uplift_server.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+PORT="${PORT:-50052}"
+VIEWER_PORT="${VIEWER_PORT:-8081}"
+
+uv run --package flashdreams-realesrgan uplift-server \
+  --port "${PORT}" \
+  --viewer_port "${VIEWER_PORT}" \
+  --viewer_jpeg_backend auto \
+  --viewer_jpeg_quality 75 \
+  --viewer_frame_stride 1 \
+  --viewer_playback_fps 30 \
+  --compile \
+  --compile_mode reduce-overhead \
+  --warmup_height 496 \
+  --warmup_width 896 \
+  --warmup_frames 3 \
+  "$@"
diff --git a/integrations/realesrgan/tests/test_realesrgan.py b/integrations/realesrgan/tests/test_realesrgan.py
new file mode 100644
index 000000000..b352edde6
--- /dev/null
+++ b/integrations/realesrgan/tests/test_realesrgan.py
@@ -0,0 +1,230 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for the Real-ESRGAN upsampler integration."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+import numpy as np
+import pytest
+import torch
+
+import realesrgan.postprocess as postprocess_mod
+from realesrgan.postprocess import RealESRGANPostProcessorConfig
+from realesrgan.upsampler import (
+    RealESRGANFrameProfile,
+    RealESRGANUpsampler,
+    create_model,
+)
+
+from flashdreams.infra.postprocess import VideoChunk
+
+pytestmark = pytest.mark.ci_cpu
+
+
+def test_create_x2_model_matches_expected_shape() -> None:
+    model, scale = create_model("RealESRGAN_x2plus")
+    model.eval()
+    x = torch.zeros(1, 3, 8, 8)
+
+    with torch.no_grad():
+        y = model(x)
+
+    assert scale == 2
+    assert y.shape == (1, 3, 16, 16)
+
+
+def test_random_weight_upsampler_scales_tensor_without_checkpoint() -> None:
+    upsampler = RealESRGANUpsampler(
+        scale=2,
+        model_name="RealESRGAN_x2plus",
+        pre_pad=0,
+        half=False,
+        device="cpu",
+        load_checkpoint=False,
+    )
+    frame = torch.zeros(3, 8, 8)
+
+    output = upsampler.upsample_frame_tensor(frame)
+
+    assert output.shape == (3, 16, 16)
+    assert output.dtype == torch.float32
+
+
+def test_random_weight_upsampler_profiled_tensor_reports_cpu_profile() -> None:
+    upsampler = RealESRGANUpsampler(
+        scale=2,
+        model_name="RealESRGAN_x2plus",
+        pre_pad=0,
+        half=False,
+        device="cpu",
+        load_checkpoint=False,
+    )
+    frame = torch.zeros(3, 8, 8)
+
+    output, profile = upsampler.upsample_frame_tensor_profiled(frame)
+
+    assert output.shape == (3, 16, 16)
+    assert profile == RealESRGANFrameProfile(model_ms=None)
+
+
+def test_upsampler_can_compile_model(monkeypatch: pytest.MonkeyPatch) -> None:
+    compile_calls = []
+
+    def fake_compile(model: torch.nn.Module, *, mode: str) -> torch.nn.Module:
+        compile_calls.append(mode)
+        return model
+
+    monkeypatch.setattr(torch, "compile", fake_compile)
+
+    RealESRGANUpsampler(
+        scale=2,
+        model_name="RealESRGAN_x2plus",
+        pre_pad=0,
+        half=False,
+        compile_model=True,
+        compile_mode="reduce-overhead",
+        device="cpu",
+        load_checkpoint=False,
+    )
+
+    assert compile_calls == ["reduce-overhead"]
+
+
+def test_bgr_image_output_uses_opencv_channel_order(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    upsampler = RealESRGANUpsampler(
+        scale=2,
+        model_name="RealESRGAN_x2plus",
+        pre_pad=0,
+        half=False,
+        device="cpu",
+        load_checkpoint=False,
+    )
+
+    def fake_upsample_frame_tensor_profiled(
+        frame: torch.Tensor,
+    ) -> tuple[torch.Tensor, RealESRGANFrameProfile]:
+        rgb = torch.empty(3, 2, 2)
+        rgb[0].fill_(0.2)
+        rgb[1].fill_(0.4)
+        rgb[2].fill_(0.6)
+        return rgb * 2.0 - 1.0, RealESRGANFrameProfile(model_ms=None)
+
+    monkeypatch.setattr(
+        upsampler,
+        "upsample_frame_tensor_profiled",
+        fake_upsample_frame_tensor_profiled,
+    )
+
+    image = np.zeros((1, 1, 3), dtype=np.uint8)
+    output, mode = upsampler.upsample_bgr_image(image)
+
+    assert mode == "RGB"
+    assert output.dtype == np.uint8
+    assert output.shape == (2, 2, 3)
+    assert output[0, 0].tolist() == [153, 102, 51]
+
+
+def test_bgra_image_output_preserves_alpha(monkeypatch: pytest.MonkeyPatch) -> None:
+    upsampler = RealESRGANUpsampler(
+        scale=2,
+        model_name="RealESRGAN_x2plus",
+        pre_pad=0,
+        half=False,
+        device="cpu",
+        load_checkpoint=False,
+    )
+
+    def fake_upsample_frame_tensor_profiled(
+        frame: torch.Tensor,
+    ) -> tuple[torch.Tensor, RealESRGANFrameProfile]:
+        return torch.zeros(3, 2, 2), RealESRGANFrameProfile(model_ms=None)
+
+    monkeypatch.setattr(
+        upsampler,
+        "upsample_frame_tensor_profiled",
+        fake_upsample_frame_tensor_profiled,
+    )
+
+    image = np.array([[[10, 20, 30, 77]]], dtype=np.uint8)
+    output, mode = upsampler.upsample_bgr_image(image)
+
+    assert mode == "RGBA"
+    assert output.dtype == np.uint8
+    assert output.shape == (2, 2, 4)
+    assert np.all(output[:, :, 3] == 77)
+
+
+@dataclass
+class _FakeUpsampler:
+    scale: int
+    model_name: str | None = None
+    model_path: str | None = None
+    tile: int = 0
+    tile_pad: int = 10
+    pre_pad: int = 10
+    half: bool = True
+    compile_model: bool = False
+    compile_mode: str = "reduce-overhead"
+    device: str = "cuda"
+
+    def upsample_video_tensor(self, video: torch.Tensor) -> torch.Tensor:
+        return video.repeat_interleave(self.scale, dim=-2).repeat_interleave(
+            self.scale,
+            dim=-1,
+        )
+
+
+def test_postprocess_scales_batched_views(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(postprocess_mod, "RealESRGANUpsampler", _FakeUpsampler)
+    session = (
+        RealESRGANPostProcessorConfig(scale=2)
+        .setup()
+        .start(postprocess_mod.VideoSpec(height=2, width=3, channels=3))
+    )
+    chunk = VideoChunk(
+        tensor=torch.zeros(1, 2, 4, 3, 2, 3),
+        layout="bvtchw",
+    )
+
+    outputs = session.process(chunk)
+
+    assert len(outputs) == 1
+    assert outputs[0].layout == "bvtchw"
+    assert outputs[0].metadata["source"] == "realesrgan"
+    assert outputs[0].tensor.shape == (1, 2, 4, 3, 4, 6)
+
+
+def test_postprocess_reports_scaled_output_spec() -> None:
+    config = RealESRGANPostProcessorConfig(scale=4)
+
+    spec = config.output_spec(postprocess_mod.VideoSpec(height=8, width=12, fps=24))
+
+    assert spec == postprocess_mod.VideoSpec(height=32, width=48, fps=24)
+
+
+def test_realesrgan_postprocess_preset_is_registered() -> None:
+    from flashdreams.plugins.registry import resolve_postprocess_preset
+
+    processor = resolve_postprocess_preset("realesrgan")
+
+    assert isinstance(processor, RealESRGANPostProcessorConfig)
+    assert processor.scale == 2
+    assert processor.device == "cuda"
+    assert processor.half is True
diff --git a/integrations/realesrgan/tests/test_uplift_server.py b/integrations/realesrgan/tests/test_uplift_server.py
new file mode 100644
index 000000000..fec00f834
--- /dev/null
+++ b/integrations/realesrgan/tests/test_uplift_server.py
@@ -0,0 +1,127 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tests for the Real-ESRGAN uplift gRPC servicer."""
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+import torch
+
+import realesrgan.grpc.uplift_server as server_mod
+from flashdreams.serving.uplift.protos import uplift_pb2 as pb2
+
+pytestmark = pytest.mark.ci_cpu
+
+
+class _FakeUpsampler:
+    def __init__(
+        self,
+        *,
+        scale: int,
+        model_name: str,
+        model_path: object,
+        tile: int,
+        tile_pad: int,
+        pre_pad: int,
+        half: bool,
+        compile_model: bool,
+        compile_mode: str,
+        device: torch.device,
+        load_checkpoint: bool,
+    ) -> None:
+        self.scale = scale
+        self.model_name = model_name
+
+    def upsample_frame_tensor(self, frame: torch.Tensor) -> torch.Tensor:
+        return frame.repeat_interleave(self.scale, dim=1).repeat_interleave(
+            self.scale,
+            dim=2,
+        )
+
+
+@pytest.fixture
+def servicer(monkeypatch: pytest.MonkeyPatch) -> server_mod.RealESRGANUplift:
+    monkeypatch.setattr(server_mod, "RealESRGANUpsampler", _FakeUpsampler)
+    return server_mod.RealESRGANUplift(
+        default_scale=2,
+        device="cpu",
+        warmup=False,
+        load_checkpoint=False,
+    )
+
+
+def _raw_request(frames: np.ndarray, *, chunk_index: int = 0) -> pb2.UpscaleChunkRequest:
+    total, height, width, _channels = frames.shape
+    return pb2.UpscaleChunkRequest(
+        num_frames=total,
+        height=height,
+        width=width,
+        chunk_index=chunk_index,
+        frame_encoding=pb2.FRAME_ENCODING_RAW_RGB,
+        frames_rgb=np.ascontiguousarray(frames).tobytes(),
+    )
+
+
+def test_unary_upscale_chunk_returns_scaled_frames(
+    servicer: server_mod.RealESRGANUplift,
+) -> None:
+    started = servicer.start_session(
+        pb2.StartSessionRequest(session_id="s0", scale=2),
+        None,
+    )
+    assert started.success
+
+    frames = np.arange(2 * 2 * 3 * 3, dtype=np.uint8).reshape(2, 2, 3, 3)
+    request = _raw_request(frames)
+    request.session_id = "s0"
+    response = servicer.upscale_chunk(request, None)
+
+    assert not response.error
+    assert not response.frames_omitted
+    assert response.num_frames == 2
+    assert response.height == 4
+    assert response.width == 6
+    output = np.frombuffer(response.frames_rgb, dtype=np.uint8).reshape(2, 4, 6, 3)
+    assert output[0, 0, 0].tolist() == frames[0, 0, 0].tolist()
+    assert output[0, 1, 1].tolist() == frames[0, 0, 0].tolist()
+
+
+def test_display_only_omits_response_frames(
+    servicer: server_mod.RealESRGANUplift,
+) -> None:
+    started = servicer.start_session(
+        pb2.StartSessionRequest(session_id="s0", scale=2),
+        None,
+    )
+    assert started.success
+
+    frames = np.zeros((1, 2, 3, 3), dtype=np.uint8)
+    request = _raw_request(frames)
+    request.session_id = "s0"
+    request.display_only = True
+
+    response = servicer.upscale_chunk(request, None)
+
+    assert not response.error
+    assert response.frames_omitted
+    assert response.frames_rgb == b""
+    assert response.height == 4
+    assert response.width == 6
+
+
+def test_streaming_upscale_video_uses_first_request_scale(
+    servicer: server_mod.RealESRGANUplift,
+) -> None:
+    frames = np.zeros((1, 2, 3, 3), dtype=np.uint8)
+    request = _raw_request(frames)
+    request.session_id = "stream"
+    request.scale = 2
+
+    responses = list(servicer.upscale_video(iter([request]), None))
+
+    assert len(responses) == 1
+    assert responses[0].session_id == "stream"
+    assert responses[0].height == 4
+    assert responses[0].width == 6
diff --git a/pyproject.toml b/pyproject.toml
index 250048b84..40624cac5 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -51,6 +51,7 @@ extraPaths = [
     "integrations/fastvideo_causal_wan22",
     "integrations/hy_worldplay",
     "integrations/lingbot",
+    "integrations/realesrgan",
     "integrations/self_forcing",
     "integrations/wan21",
     "integrations/wan22",
@@ -74,6 +75,7 @@ extra-paths = [
     "integrations/fastvideo_causal_wan22",
     "integrations/hy_worldplay",
     "integrations/lingbot",
+    "integrations/realesrgan",
     "integrations/self_forcing",
     "integrations/wan21",
     "integrations/wan22",
diff --git a/run_uplift_server.sh b/run_uplift_server.sh
new file mode 100755
index 000000000..8eae892bc
--- /dev/null
+++ b/run_uplift_server.sh
@@ -0,0 +1,127 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -euo pipefail
+
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+usage() {
+    cat <<'EOF'
+Usage:
+  ./run_uplift_server.sh [flashvsr|realesrgan] [uplift-server args...]
+
+Environment:
+  UPSAMPLER             flashvsr or realesrgan. Default: flashvsr
+  PORT                  gRPC port. Default: 8090
+  VIEWER_HOST           HTTP viewer bind host. Default: 0.0.0.0
+  VIEWER_PORT           HTTP viewer port. Defaults: 8091
+  VIEWER_JPEG_BACKEND   auto, cuda, or pillow. Default: auto
+  VIEWER_JPEG_QUALITY   JPEG quality. Default: 75
+  VIEWER_FRAME_STRIDE   Publish every Nth viewer frame. Default: 1
+  VIEWER_PLAYBACK_FPS   Steady viewer playback FPS. Default: 30
+  WARMUP_HEIGHT         Real-ESRGAN warmup input height. Default: 704
+  WARMUP_WIDTH          Real-ESRGAN warmup input width. Default: 1280
+  COMPILE               Set to 1 to pass --compile. Default: 0
+  CUDA_GRAPH            FlashVSR only: set to 1 to pass --cuda_graph. Default: 0
+
+Examples:
+  ./run_uplift_server.sh flashvsr --attention_mode auto --sparse_ratio 1.5
+  ./run_uplift_server.sh realesrgan --default_scale 2
+  UPSAMPLER=realesrgan VIEWER_PORT=8081 ./run_uplift_server.sh --compile
+EOF
+}
+
+if [[ "${1:-}" == "-h" || "${1:-}" == "--help" ]]; then
+    usage
+    exit 0
+fi
+
+if [[ "${1:-}" == "flashvsr" || "${1:-}" == "realesrgan" || "${1:-}" == "real-esrgan" ]]; then
+    UPSAMPLER="$1"
+    shift
+else
+    UPSAMPLER="${UPSAMPLER:-flashvsr}"
+fi
+
+case "${UPSAMPLER}" in
+    flashvsr)
+        PACKAGE="flashdreams-flashvsr"
+        MODULE="flashvsr.grpc.uplift_server"
+        LABEL="FlashVSR"
+        DEFAULT_VIEWER_PORT=8081
+        ;;
+    realesrgan|real-esrgan)
+        PACKAGE="flashdreams-realesrgan"
+        MODULE="realesrgan.grpc.uplift_server"
+        LABEL="Real-ESRGAN"
+        DEFAULT_VIEWER_PORT=8081
+        ;;
+    *)
+        echo "unsupported UPSAMPLER=${UPSAMPLER}; expected flashvsr or realesrgan" >&2
+        exit 2
+        ;;
+esac
+
+PORT="${PORT:-8090}"
+VIEWER_HOST="${VIEWER_HOST:-0.0.0.0}"
+VIEWER_PORT="${VIEWER_PORT:-${DEFAULT_VIEWER_PORT}}"
+VIEWER_JPEG_BACKEND="${VIEWER_JPEG_BACKEND:-auto}"
+VIEWER_JPEG_QUALITY="${VIEWER_JPEG_QUALITY:-75}"
+VIEWER_FRAME_STRIDE="${VIEWER_FRAME_STRIDE:-1}"
+VIEWER_PLAYBACK_FPS="${VIEWER_PLAYBACK_FPS:-30}"
+WARMUP_HEIGHT="${WARMUP_HEIGHT:-704}"
+WARMUP_WIDTH="${WARMUP_WIDTH:-1280}"
+COMPILE="${COMPILE:-0}"
+CUDA_GRAPH="${CUDA_GRAPH:-0}"
+
+export TORCHINDUCTOR_CACHE_DIR="${TORCHINDUCTOR_CACHE_DIR:-${HOME}/.cache/torchinductor-${UPSAMPLER}}"
+export TRITON_CACHE_DIR="${TRITON_CACHE_DIR:-${HOME}/.cache/triton-${UPSAMPLER}}"
+
+server_args=(
+    --port "${PORT}"
+    --viewer_host "${VIEWER_HOST}"
+    --viewer_port "${VIEWER_PORT}"
+    --viewer_jpeg_backend "${VIEWER_JPEG_BACKEND}"
+    --viewer_jpeg_quality "${VIEWER_JPEG_QUALITY}"
+    --viewer_frame_stride "${VIEWER_FRAME_STRIDE}"
+    --viewer_playback_fps "${VIEWER_PLAYBACK_FPS}"
+)
+
+if [[ "${UPSAMPLER}" == "realesrgan" || "${UPSAMPLER}" == "real-esrgan" ]]; then
+    server_args+=(
+        --warmup_height "${WARMUP_HEIGHT}"
+        --warmup_width "${WARMUP_WIDTH}"
+    )
+fi
+
+if [[ "${COMPILE}" == "1" || "${COMPILE}" == "true" ]]; then
+    server_args+=(--compile)
+fi
+
+if [[ "${UPSAMPLER}" == "flashvsr" && ( "${CUDA_GRAPH}" == "1" || "${CUDA_GRAPH}" == "true" ) ]]; then
+    server_args+=(--cuda_graph)
+fi
+
+echo "Starting ${LABEL} uplift server on :${PORT}"
+if [[ "${VIEWER_PORT}" != "0" ]]; then
+    echo "Starting ${LABEL} HTTP viewer on ${VIEWER_HOST}:${VIEWER_PORT}"
+    echo "Using ${LABEL} viewer JPEG backend: ${VIEWER_JPEG_BACKEND}"
+fi
+
+# Use the module path rather than the shared ``uplift-server`` console script:
+# duplicate script names across workspace packages resolve nondeterministically
+# in one uv-managed venv.
+exec uv run --package "${PACKAGE}" python -m "${MODULE}" "${server_args[@]}" "$@"
diff --git a/uv.lock b/uv.lock
index 2b0d03215..8b2417c52 100644
--- a/uv.lock
+++ b/uv.lock
@@ -27,6 +27,7 @@ members = [
     "flashdreams-hy-worldplay",
     "flashdreams-lingbot",
     "flashdreams-omnidreams",
+    "flashdreams-realesrgan",
     "flashdreams-self-forcing",
     "flashdreams-wan21",
     "flashdreams-wan22",
@@ -1257,6 +1258,10 @@ runners = [
 serving = [
     { name = "aiohttp" },
     { name = "aiortc" },
+    { name = "grpcio" },
+    { name = "mediapy" },
+    { name = "pillow" },
+    { name = "protobuf" },
 ]
 
 [package.dev-dependencies]
@@ -1282,15 +1287,19 @@ requires-dist = [
     { name = "einops", specifier = ">=0.7" },
     { name = "filelock", specifier = ">=3" },
     { name = "ftfy", specifier = ">=6.0" },
+    { name = "grpcio", marker = "extra == 'serving'", specifier = ">=1.80" },
     { name = "huggingface-hub", specifier = ">=0.33" },
     { name = "loguru", specifier = ">=0.7" },
     { name = "mediapy", marker = "extra == 'dev'", specifier = ">=1.1" },
     { name = "mediapy", marker = "extra == 'examples'", specifier = ">=1.1" },
     { name = "mediapy", marker = "extra == 'runners'", specifier = ">=1.1" },
+    { name = "mediapy", marker = "extra == 'serving'", specifier = ">=1.1" },
     { name = "numpy", specifier = ">=1.24,<2.5" },
     { name = "nvidia-ml-py", specifier = ">=12.0" },
     { name = "opencv-python-headless", marker = "extra == 'examples'", specifier = ">=4.5" },
     { name = "opencv-python-headless", marker = "extra == 'runners'", specifier = ">=4.5" },
+    { name = "pillow", marker = "extra == 'serving'", specifier = ">=10" },
+    { name = "protobuf", marker = "extra == 'serving'", specifier = ">=7.35.0,<8" },
     { name = "pytest", marker = "extra == 'dev'", specifier = ">=8.0" },
     { name = "pytest-manual-marker", marker = "extra == 'dev'", specifier = ">=2.0" },
     { name = "safetensors", specifier = ">=0.4" },
@@ -1395,12 +1404,9 @@ name = "flashdreams-flashvsr"
 version = "0.1.0"
 source = { editable = "integrations/flashvsr" }
 dependencies = [
-    { name = "flashdreams" },
-    { name = "grpcio" },
+    { name = "flashdreams", extra = ["serving"] },
     { name = "mediapy" },
     { name = "opencv-python" },
-    { name = "pillow" },
-    { name = "protobuf" },
 ]
 
 [package.optional-dependencies]
@@ -1411,13 +1417,10 @@ dev = [
 
 [package.metadata]
 requires-dist = [
-    { name = "flashdreams", editable = "flashdreams" },
-    { name = "grpcio", specifier = ">=1.80" },
+    { name = "flashdreams", extras = ["serving"], editable = "flashdreams" },
     { name = "grpcio-tools", marker = "extra == 'dev'", specifier = ">=1.80" },
     { name = "mediapy", specifier = ">=1.1" },
     { name = "opencv-python", specifier = ">=4.8" },
-    { name = "pillow", specifier = ">=10" },
-    { name = "protobuf", specifier = ">=6.31.1" },
     { name = "pytest", marker = "extra == 'dev'", specifier = ">=8.0" },
 ]
 provides-extras = ["dev"]
@@ -1567,6 +1570,28 @@ requires-dist = [
 ]
 provides-extras = ["interactive-drive", "dev"]
 
+[[package]]
+name = "flashdreams-realesrgan"
+version = "0.1.0"
+source = { editable = "integrations/realesrgan" }
+dependencies = [
+    { name = "flashdreams", extra = ["serving"] },
+    { name = "opencv-python-headless" },
+]
+
+[package.optional-dependencies]
+dev = [
+    { name = "pytest" },
+]
+
+[package.metadata]
+requires-dist = [
+    { name = "flashdreams", extras = ["serving"], editable = "flashdreams" },
+    { name = "opencv-python-headless", specifier = ">=4.8" },
+    { name = "pytest", marker = "extra == 'dev'", specifier = ">=8.0" },
+]
+provides-extras = ["dev"]
+
 [[package]]
 name = "flashdreams-self-forcing"
 version = "0.1.0"