🌍 DeMark-World

A Universal Framework for AI Video Watermark Removal

DeMark-World is the evolution of SoraWatermarkCleaner. While its predecessor focused on Sora, DeMark-World is a "ready-to-roll" universal framework designed to detect and remove unwanted watermarks from any AI-generated video (Sora, Runway, Pika, Kling, etc.) with high fidelity and temporal consistency.

Prompt: a cat running in a forest.

Sora2	sora2_cat_run_sora_watermark_removed_combined_10M.mp4
Google Veo3.1	Veo3_Cat_Running_In_Forest_Video_sora_watermark_removed_combined_10M.mp4
Ruanway Gen4	gen4_cat_run_sora_watermark_removed_combined_10M.mp4

✨ Features

Moving beyond just Sora v2 in SoraWatermarkCleaner, DeMark-World is engineered to handle watermarks from the latest generation of video models, including Google Gemini/Veo, Runway Gen-3/Gen-4, Pika, Kling, and Luma Dream Machine.

Feature	Details
Universal detection	Works with Sora, Runway, Pika, Kling, Veo, and more
2 inpainting backends	LaMa (fast) · E2FGVI-HQ (temporally consistent)
Batch YOLO detection	Processes frames in GPU batches (--batch-size 4) for ~2× detection throughput
BF16 inference	Auto-enabled on supported GPUs (Ampere+) via torch.cuda.is_bf16_supported()
TorchCompile	torch.compile with artifact caching for persistent speedup across runs
4 interfaces	Web UI · Python API · CLI · REST Server
Docker support	GPU-enabled docker-compose.yaml included

🛠️ Installation

Prerequisites: You must have FFmpeg installed and added to your system PATH.

We use uv for project management. It is significantly faster and more reliable than pip/poetry.

1. Clone the repository

   Bash

   git clone https://github.com/linkedlist771/DeMark-World.git
   cd DeMark-World
   

2. Install environment

   # This creates the virtual environment and installs all dependencies
   uv sync
   

3. Activate environment

   # Linux/MacOS
   source .venv/bin/activate
   
   # Windows
   .venv\Scripts\activate
   

Note on Models: Detector weights and Inpainting models will be downloaded automatically to the cache directory upon the first run.

🚀 Quick Start

1. Interactive Web UI (Recommended)

The easiest way to use DeMark-World is via the modern Streamlit interface.

streamlit run app.py

Then open your browser to http://localhost:8501.

2. CLI

Process an entire folder of videos from the command line with a rich progress display:

# Basic usage (LaMa, fast)
demark-world -i /path/to/videos -o /path/to/output

# High quality, time-consistent (E2FGVI-HQ)
demark-world -i /path/to/videos -o /path/to/output --model e2fgvi_hq

# Max performance on CUDA (BF16 + TorchCompile, auto-enabled on supported GPUs)
demark-world -i /path/to/videos -o /path/to/output --model e2fgvi_hq --torch-compile

# Adjust detection batch size
demark-world -i /path/to/videos -o /path/to/output --batch-size 8

# Suppress internal progress bars
demark-world -i /path/to/videos -o /path/to/output --quiet

3. Python API

You can integrate DeMark-World into your own pipelines easily.

from pathlib import Path
from src.demark_world.core import DeMarkWorld
from src.demark_world.schemas import CleanerType

if __name__ == "__main__":
    input_video = Path("resources/Veo3_Cat_Running_In_Forest_Video.mp4")
    output_video: Path = Path("outputs/cleaned.mp4")

    # Option 1: LaMa (Fast)
    demarker = DeMarkWorld(cleaner_type=CleanerType.LAMA)

    # Option 2: E2FGVI_HQ (High Quality + Time Consistent)
    # BF16 and batch detection are auto-configured based on your hardware
    # demarker = DeMarkWorld(cleaner_type=CleanerType.E2FGVI_HQ)

    demarker.run(input_video, output_video)

4. REST Server

Start the API server for queue-based async processing:

python start_server.py --host 0.0.0.0 --port 5344

Endpoints:

Method	Path	Description
POST	/submit_remove_task	Submit a video; accepts ?cleaner_type=lama|e2fgvi_hq
GET	/get_results?remove_task_id=<id>	Poll task status and progress
GET	/get_queue_status	View queue depth and all task states
GET	/download/<task_id>	Download the processed video

# Submit a task (LAMA, default)
curl -X POST http://localhost:5344/submit_remove_task \
  -F "video=@input.mp4"

# Submit with E2FGVI_HQ
curl -X POST "http://localhost:5344/submit_remove_task?cleaner_type=e2fgvi_hq" \
  -F "video=@input.mp4"

# Check queue
curl http://localhost:5344/get_queue_status

5. Docker

docker compose up

The server will be available at http://localhost:5344. GPU passthrough is configured automatically.

🧠 How It Works

DeMark-World operates in a two-stage pipeline:

1. Detection — A YOLO model scans every frame (in GPU batches) to locate the watermark bounding box. Missed detections are imputed using interval-averaged bounding boxes from neighboring frames.
2. Inpainting — The detected region is inpainted using the selected backend:
   • LaMa: per-frame, fast, suitable for static watermarks.
   • E2FGVI-HQ: processes overlapping temporal segments for smooth, flicker-free results.

⚡ Performance

Optimization	How it's applied
Batch YOLO detection	Frames are accumulated into batches (DEFAULT_DETECT_BATCH_SIZE=4) before GPU inference
BF16 inference	Auto-detected via torch.cuda.is_bf16_supported() — enabled by default on Ampere (RTX 30xx/40xx, A100, H100) and newer
TorchCompile	torch.compile(mode="default") with artifact caching in ~/.cache/torch_compile/e2fgvi_hq/ — first run compiles once, subsequent runs load from cache

📜 License

Distributed under the Apache 2.0 License. See LICENSE for more information.

🖊️ Citation

If you find this project helpful in your research or work, please cite:

@misc{DeMark-World2025,
  author = {linkedlist771},
  title = {DeMark-World},
  year = {2025},
  url = {https://github.com/linkedlist771/DeMark-World}
}

🙏 Acknowledgments

• For the incredible implementation of SOTA inpainting models.
• For the YOLO object detection framework.

---

💝 If you find this project helpful, please consider starring the repo!