Docker Compose Deployment Support

What's New

Docker Compose deployment is now available, making it significantly
easier to get SoraWatermarkCleaner up and running as a service without
manual environment setup. See more details at #96 .

Motivation

This addresses a common pain point where users struggled with CUDA
environment configuration, FFmpeg installation, and dependency
management. Docker Compose encapsulates all of this into a single
reproducible deployment.

Performance Boost Edition

Highlights

This release focuses on significant performance improvements for video watermark removal, achieving up to 2.43× speedup compared to the baseline E2FGVI cleaner. See #91 #90 #87 and #69

✨ New Features

1. Bf16 Inference Support

• Enable bfloat16 precision inference with use_bf16=True
• Up to 2.43× faster processing on supported GPUs (Ampere+, e.g., RTX 30/40 series, A100)
• Separate compile cache for bf16 mode to avoid recompilation overhead

2. YOLO Batch Detection

• Batch processing for watermark detection (detect_batch_size=4 by default)
• ~40% speedup on detection phase

3. Torch Compile Integration

• Automatic torch.compile optimization for E2FGVI model
• Compile artifacts caching for faster subsequent runs
• ~22% speedup on cleaning phase

📊 Performance Benchmarks

Detector	Batch	Cleaner	TorchCompile	Bf16	Time (s)	Speedup
YOLO	×	LAMA	×	×	44.33	-
YOLO	×	E2FGVI	×	×	142.42	1.00×
YOLO	×	E2FGVI	✓	×	117.19	1.22×
YOLO	4	E2FGVI	✓	×	82.63	1.72×
YOLO	4	E2FGVI	✓	✓	58.60	2.43×

Tested on a 10-second video (280 frames, 704×1280) with RTX GPU.

V0.0.3: With time consistency model for flicker and batch process

1. We provide two models to remove watermark. LAMA is fast but may have flicker on the cleaned area, which E2FGVI_HQ compromise this only requires cuda otherwise very slow on CPU or MPS. See more details on #65.

2. add batch process in streamlit and cli with #45 and #54.

V0.0.2

1. based on the temporal consistency feature of the watermark's appearing, we use this to impure the missed detect: #13

2. bit rate is fixed #7

3. dataset is released: https://huggingface.co/datasets/LLinked/sora-watermark-dataset

4. provide a windows one-click portable distribution: https://github.com/linkedlist771/SoraWatermarkCleaner?tab=readme-ov-file#3-one-click-portable-version

SoraWM yolo weights

The first version of the yolo weights that detect the sora2 watermark posiitons.