Summary

This project addresses watermark removal from AI-generated videos, specifically targeting OpenAI's Sora model. It offers a deep learning-powered solution for users needing to clean video content, providing high-quality watermark removal for subsequent use or analysis.

How It Works

A two-stage deep learning pipeline detects and removes watermarks. A YOLOv11s model identifies watermark locations, followed by a cleaning stage using either the fast LAMA model (IOPaint reference) or the time-consistent E2FGVI_HQ model. This purely deep learning-driven approach aims for effective watermark removal across various generated videos.

Quick Start & Requirements

• Installation: FFmpeg is a prerequisite. Environment setup via uv is recommended. A one-click portable build for Windows and Docker Compose deployment are available.
• Prerequisites: CUDA-capable GPU is highly recommended for optimal performance, especially with E2FGVI_HQ and Docker. The Docker image is large (~20 GB) due to NVIDIA libraries and PyTorch.
• Links: Hosted Service: https://www.sorawatermarkremover.ai/; Portable Build downloads; Docker Compose instructions; Streamlit App (streamlit run app.py); FastAPI Web Server (python start_server.py); Hugging Face Datasets: https://huggingface.co/datasets/LLinked/sora-watermark-dataset.

Highlighted Details

• Performance optimizations include TorchCompile, bf16 inference (up to 2.43x speedup with potential minor quality loss), and batch processing for watermark detection.
• Two cleaning models: LAMA (faster, potential flicker) and E2FGVI_HQ (time-consistent, slower, CUDA-dependent).
• Batch processing is supported via CLI (cli.py) and an interactive Streamlit web UI.
• Docker Compose deployment simplifies setup for users with NVIDIA Container Toolkit and a CUDA-enabled GPU.
• A commercial hosted service is available, utilizing the project's backend.

Maintenance & Community

This project is archived due to OpenAI discontinuing the Sora model. The maintainer directs users to DeMark-World for a universal watermark removal solution. Labeled datasets are available on Hugging Face for custom model training.

Licensing & Compatibility

Licensed under the Apache License 2.0, generally permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The project is archived and no longer maintained. E2FGVI_HQ performs slowly on CPU/MPS. Bf16 inference, while faster, may introduce minor quality degradation. The Docker image is substantial (~20 GB) and requires CUDA.