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.