A watermark eraser for a video model that no longer exists

What it does

SoraWatermarkCleaner is a two-stage pipeline that detects and removes watermarks from videos generated by OpenAI’s Sora2. A custom-trained YOLOv11s model locates the watermark, and then either a LAMA model quickly inpaints it frame-by-frame or an E2FGVI_HQ model does the same with temporal consistency—at the cost of speed and a hard CUDA requirement. The project wraps these engines in a Streamlit GUI, a FastAPI server, a CLI, and even a Docker image that weighs in around 20 GB.

The interesting bit

The whole effort is now officially archived because OpenAI discontinued Sora entirely. The author has already pivoted to DeMark-World, a universal watermark remover for Veo, Runway, and others—making this repo less a utility and more a time capsule of the generative-video arms race.

Key highlights

• Two-stage deep learning: YOLOv11s detection plus LAMA or E2FGVI_HQ inpainting
• Performance tuning table shows bf16 inference, TorchCompile, and batch detection cutting E2FGVI processing time from roughly 142 s to 59 s
• Ships with multiple interfaces: CLI, Streamlit app, FastAPI web server, Replicate Cog API, and a one-click Windows portable build
• Published a labeled Sora watermark dataset on Hugging Face for custom detector training
• Explicitly archived; successor project DeMark-World targets other video models

Caveats

• The project is archived and no longer maintained since OpenAI shut down Sora
• E2FGVI_HQ requires CUDA and is very slow on CPU or MPS
• The Docker image is roughly 20 GB, and the fast LAMA mode can introduce flicker

Verdict

Grab it if you are maintaining a backlog of Sora2 clips or researching watermark-removal pipelines. Everyone else should skip straight to the author’s newer DeMark-World, which aims for universal coverage and is actually breathing.