video-subtitle-remover

Video-subtitle-remover (VSR) is an AI-based video hard subtitle removal software. Main features include: lossless resolution removal of video hard subtitles and generation of files; filling and repairing the area where subtitles were removed using AI algorithm models (non-adjacent pixel filling and mosaic removal); extracting original video subtitles (can be used with video-subtitle-extractor); supporting custom subtitle position to remove subtitles from specified areas only; supporting automatic removal of all text from the entire video; supporting batch removal of watermark text from multiple images. The software provides both GUI graphical interface and command-line usage. The command-line supports specifying input/output file paths, subtitle area coordinates, and various repair modes (sttn-auto, sttn-det, lama, propainter, opencv). Multiple pre-built packages are provided to adapt to different environments: Windows CPU general version, Windows DirectML version (non-Nvidia GPU), Windows Nvidia CUDA 11.8/12.6/12.8 versions, covering different GPU computing capabilities. It also provides Docker images supporting Nvidia 10-50 series GPUs, AMD/Intel GPUs and CPU. Source code installation supports four runtime modes: CUDA (NVIDIA acceleration), CPU, DirectML (AMD/Intel acceleration), and macOS (Apple Silicon). It requires Python 3.12+, with dependencies on PaddlePaddle and PyTorch libraries. The FAQ section introduces the characteristics of three repair algorithms: The STTN algorithm works well with human video and is fast, can skip subtitle detection; The LAMA algorithm works well with images and animated videos; The ProPainter algorithm works well with videos with intense motion but requires large GPU memory and is slower. Users can adjust processing speed and effect through configuration parameters, or use the project's provided training code to train and replace the default model.