video-subtitle-remover

Video-subtitle-remover (VSR) is a video hardcoded subtitle removal software based on AI technology. Main features include: lossless-resolution removal of video hardcoded subtitles with file generation; inpainting repair of subtitle-removed regions through AI algorithm models (non-adjacent pixel filling and mosaic removal); extraction of original video subtitles (can be used in conjunction with video-subtitle-extractor); support for custom subtitle positions to remove only specified regions; full-video automatic removal of all text; support for multi-selected images for batch watermark text removal. The software provides both GUI and command-line interfaces. The command-line interface supports specifying input/output file paths, subtitle region coordinates, and multiple inpainting modes (sttn-auto, sttn-det, lama, propainter, opencv). Pre-built packages are available for different environments: Windows CPU general version, Windows DirectML version (for non-Nvidia GPUs), and Windows Nvidia CUDA 11.8/12.6/12.8 versions, covering GPUs across a range of compute capabilities. Docker images are also provided, supporting Nvidia 10-50 series GPUs, AMD/Intel GPUs, and CPU execution. Source installation supports four running modes: CUDA (NVIDIA acceleration), CPU, DirectML (AMD/Intel acceleration), and macOS (Apple Silicon). Python 3.12+ is required, with dependencies including PaddlePaddle and PyTorch. Common issues describe the characteristics of three inpainting algorithms: STTN performs well on real-person videos and is fast, allowing subtitle detection to be skipped; LAMA performs well on images and anime-style videos; ProPainter works well on videos with intense motion but requires significant VRAM and is slower. Users can adjust processing speed and quality via configuration parameters, or use the project's training code to train custom models to replace the defaults.