feat(Model Support): add Krea-2-Turbo model + LoRA support (WIP)

README.md

@@ -72,6 +72,7 @@ Invoke features an organized gallery system for easily storing, accessing, and r
 - Flux.2 Klein 9B
 - Z-Image Turbo
 - Z-Image Base
+- Krea 2 Turbo
 - Anima
 - Qwen Image
 - Qwen Image Edit

docs/src/content/docs/features/krea-2.mdx

@@ -0,0 +1,71 @@
+---
+title: Krea-2
+description: Generate images with the Krea-2 text-to-image models (Turbo and Raw), including the GGUF / single-file workflow and the conditioning enhancers.
+lastUpdated: 2026-07-13
+sidebar:
+  order: 5
+---
+
+Krea-2 is a ~12B single-stream diffusion-transformer text-to-image family. InvokeAI supports both
+published checkpoints:
+
+- **Krea-2-Turbo** — distilled for fast, low-step generation. Runs at **8 steps** with **CFG disabled**
+  (CFG Scale `1.0`). This is the recommended checkpoint for everyday use.
+- **Krea-2-Raw** — the undistilled base checkpoint. Runs at **more steps (~28)** with **CFG enabled**
+  (CFG Scale ~`5.5`, equivalent to the reference pipeline's guidance `4.5`) and supports negative prompts. It is primarily intended as a base for
+  finetuning / LoRA training, but full inference is supported.
+
+The variant is detected automatically on install, and selecting a Krea-2 model sets sensible defaults
+(steps, CFG, 1024×1024) for that variant.
+
+## Hardware
+
+Krea-2 is a large model. See the [System Requirements](../../start-here/system-requirements) table for
+details. In short, on a 24 GB card enable **FP8** in the model's Default Settings to fit 1024² (with a
+LoRA). For lower VRAM, use a **GGUF** transformer (Q4_K ≈ 12 GB total).
+
+## Installing
+
+The easiest path is the **Krea-2 launchpad bundle** in the Model Manager, which installs the models and
+their dependencies together.
+
+Krea-2 needs three components:
+
+| Component               | Diffusers install       | GGUF / single-file install           |
+| ----------------------- | ----------------------- | ------------------------------------ |
+| Transformer             | bundled in the pipeline | the `.gguf` / single-file checkpoint |
+| VAE (Qwen-Image)        | bundled                 | installed separately                 |
+| Text encoder (Qwen3-VL) | bundled                 | installed separately                 |
+
+- **Diffusers** (e.g. `krea/Krea-2-Turbo`, `krea/Krea-2-Raw`): a single ~26 GB install that bundles the
+  VAE and text encoder. Nothing else is required.
+- **GGUF / single-file**: the checkpoint ships **only the transformer**. You must also install a
+  standalone **Qwen-Image VAE** and a **Qwen3-VL encoder** (both included in the launchpad bundle).
+
+When you select a GGUF/single-file Krea-2 model, InvokeAI auto-selects an installed VAE and Qwen3-VL
+encoder if you have them. If none are installed, you'll be prompted to pick them (in the model dropdowns)
+before you can generate. Selecting a Diffusers Krea-2 model clears those standalone selections and uses
+the bundled components.
+
+:::note[Qwen3-VL encoder download]
+Single-file Qwen3-VL encoders bundle only the weights. The tokenizer/config are fetched once from
+HuggingFace (`Qwen/Qwen3-VL-4B-Instruct`) on first use, then cached for offline use. A single-file fp8
+encoder is kept resident in fp8 (roughly half the VRAM of the bf16 encoder).
+:::
+
+## Conditioning enhancers
+
+Two optional, off-by-default toggles are available under **Advanced Options** (below CFG Scale). They
+transform the text conditioning and are especially useful for the distilled Turbo checkpoint:
+
+- **Conditioning Rebalance** — per-layer weighting of the text embedding to improve prompt adherence.
+- **Seed Variance Enhancer** — injects controlled noise into the conditioning to restore per-seed
+  diversity (the distilled model otherwise produces near-identical images across seeds), trading some
+  prompt adherence for variety.
+
+Both are recorded in image metadata and can be recalled.
+
+## LoRA
+
+Krea-2 LoRAs (diffusers PEFT format) are supported and apply to both the transformer and — where the
+LoRA includes text-encoder layers — the Qwen3-VL encoder.

docs/src/content/docs/start-here/system-requirements.mdx

@@ -28,6 +28,7 @@ The requirements below are rough guidelines for best performance. GPUs with less
 | FLUX.2 Klein 4B | 1024x1024 | Nvidia 30xx+ | 12GB | 16GB | FP8 works with 8GB+; Diffusers + encoder |
 | FLUX.2 Klein 9B | 1024x1024 | Nvidia 40xx | 24GB | 32GB | FP8 works with 12GB+; Diffusers + encoder |
 | Z-Image Turbo | 1024x1024 | Nvidia 20xx+ | 8GB | 16GB | Q4_K 8GB; Q8/BF16 16GB+ |
+| Krea-2 (Turbo / Raw) | 1024x1024 | Nvidia 40xx | 24GB | 32GB | FP8 works with 16GB+; GGUF Q4_K ~12GB. Diffusers ~26GB; GGUF needs a standalone VAE + Qwen3-VL encoder |
 
 :::tip[`tmpfs` on Linux]
   If your temporary directory is mounted as a `tmpfs`, ensure it has sufficient space.

invokeai/app/api/dependencies.py

@@ -60,6 +60,7 @@
     CogView4ConditioningInfo,
     ConditioningFieldData,
     FLUXConditioningInfo,
+    Krea2ConditioningInfo,
     QwenImageConditioningInfo,
     SD3ConditioningInfo,
     SDXLConditioningInfo,
@@ -153,6 +154,7 @@ def initialize(
                     CogView4ConditioningInfo,
                     ZImageConditioningInfo,
                     QwenImageConditioningInfo,
+                    Krea2ConditioningInfo,
                     AnimaConditioningInfo,
                 ],
                 ephemeral=True,

invokeai/app/invocations/fields.py

@@ -156,6 +156,7 @@ class FieldDescriptions:
     t5_encoder = "T5 tokenizer and text encoder"
     glm_encoder = "GLM (THUDM) tokenizer and text encoder"
     qwen3_encoder = "Qwen3 tokenizer and text encoder"
+    qwen3_vl_encoder = "Qwen3-VL tokenizer and text encoder"
     clip_embed_model = "CLIP Embed loader"
     clip_g_model = "CLIP-G Embed loader"
     unet = "UNet (scheduler, LoRAs)"
@@ -172,6 +173,7 @@ class FieldDescriptions:
     sd3_model = "SD3 model (MMDiTX) to load"
     cogview4_model = "CogView4 model (Transformer) to load"
     z_image_model = "Z-Image model (Transformer) to load"
+    krea2_model = "Krea-2 model (Transformer) to load"
     qwen_image_model = "Qwen Image Edit model (Transformer) to load"
     qwen_vl_encoder = "Qwen2.5-VL tokenizer, processor and text/vision encoder"
     sdxl_main_model = "SDXL Main model (UNet, VAE, CLIP1, CLIP2) to load"
@@ -350,6 +352,12 @@ class QwenImageConditioningField(BaseModel):
     conditioning_name: str = Field(description="The name of conditioning tensor")
 
 
+class Krea2ConditioningField(BaseModel):
+    """A Krea-2 conditioning tensor primitive value"""
+
+    conditioning_name: str = Field(description="The name of conditioning tensor")
+
+
 class AnimaConditioningField(BaseModel):
     """An Anima conditioning tensor primitive value.
 

invokeai/app/invocations/krea2_conditioning_rebalance.py

@@ -0,0 +1,79 @@
+import math
+
+import torch
+
+from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
+from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField, Krea2ConditioningField
+from invokeai.app.invocations.primitives import Krea2ConditioningOutput
+from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.backend.krea2.sampling_utils import KREA2_SELECT_LAYERS
+from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
+    ConditioningFieldData,
+    Krea2ConditioningInfo,
+)
+
+_NUM_TEXT_LAYERS = len(KREA2_SELECT_LAYERS)  # 12
+
+
+@invocation(
+    "krea2_conditioning_rebalance",
+    title="Conditioning Rebalance - Krea-2",
+    tags=["conditioning", "krea2", "krea-2"],
+    category="conditioning",
+    version="1.0.0",
+    classification=Classification.Prototype,
+)
+class Krea2ConditioningRebalanceInvocation(BaseInvocation):
+    """Per-layer rebalancing of Krea-2 text conditioning (improves prompt adherence).
+
+    Krea-2 conditioning stacks 12 Qwen3-VL hidden-state layers per token. Weighting those layers
+    individually (and applying an overall multiplier) lets you push the model harder toward the prompt,
+    counteracting the quality-dilution from distillation. Ported from the ComfyUI
+    `ConditioningKrea2Rebalance` node. This is an optional pass between the text encoder and denoise.
+    """
+
+    conditioning: Krea2ConditioningField = InputField(
+        description=FieldDescriptions.cond, input=Input.Connection, title="Conditioning"
+    )
+    per_layer_weights: str = InputField(
+        default="1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0",
+        description=f"Comma-separated gains for the {_NUM_TEXT_LAYERS} tapped encoder layers (exactly "
+        f"{_NUM_TEXT_LAYERS} values).",
+    )
+    multiplier: float = InputField(
+        default=4.0,
+        allow_inf_nan=False,
+        description="Overall multiplier applied to the conditioning after per-layer weighting.",
+    )
+
+    def _parse_weights(self) -> list[float]:
+        try:
+            weights = [float(x.strip()) for x in self.per_layer_weights.split(",") if x.strip() != ""]
+        except ValueError as e:
+            raise ValueError(f"per_layer_weights must be comma-separated numbers: {e}") from e
+        if len(weights) != _NUM_TEXT_LAYERS:
+            raise ValueError(f"per_layer_weights must have exactly {_NUM_TEXT_LAYERS} values, got {len(weights)}.")
+        if not all(math.isfinite(weight) for weight in weights):
+            raise ValueError("per_layer_weights must contain only finite values.")
+        return weights
+
+    @torch.no_grad()
+    def invoke(self, context: InvocationContext) -> Krea2ConditioningOutput:
+        weights = self._parse_weights()
+
+        cond_data = context.conditioning.load(self.conditioning.conditioning_name)
+        assert len(cond_data.conditionings) == 1
+        conditioning = cond_data.conditionings[0]
+        assert isinstance(conditioning, Krea2ConditioningInfo)
+
+        embeds = conditioning.prompt_embeds  # (B, seq, 12, hidden)
+        gains = torch.tensor(weights, dtype=embeds.dtype, device=embeds.device).view(1, 1, _NUM_TEXT_LAYERS, 1)
+        embeds = embeds * gains * self.multiplier
+
+        new_data = ConditioningFieldData(
+            conditionings=[
+                Krea2ConditioningInfo(prompt_embeds=embeds, prompt_embeds_mask=conditioning.prompt_embeds_mask)
+            ]
+        )
+        conditioning_name = context.conditioning.save(new_data)
+        return Krea2ConditioningOutput.build(conditioning_name)