This project provides an AI-powered tool for removing hard-coded subtitles and text-based watermarks from images and videos, offering lossless resolution output. It's designed for users who need to clean up media content without relying on external APIs, with a focus on local processing and advanced inpainting techniques.

How It Works

The tool utilizes advanced AI algorithms, specifically mentioning STTN, LAMA, and PROPAINTER, for inpainting the areas where subtitles are removed. This approach aims to intelligently fill in the removed content rather than using simple pixel duplication or blurring, leading to higher quality results. Users can either specify subtitle locations for targeted removal or allow the AI to automatically detect and remove all text across the entire video.

Quick Start & Requirements

• Installation: Download pre-compiled Windows GPU version or install from source.
• Prerequisites:
  • NVIDIA GPU (GTX 1060 or higher recommended).
  • CUDA 11.7 (or 11.8 for RTX 4090) and cuDNN 8.4.1.
  • Python 3.8+.
  • Miniconda for environment management.
  • CPU with AVX instruction set support.
• Setup: Source installation involves creating a conda environment, installing dependencies (including paddlepaddle-gpu and pytorch), and configuring CUDA/cuDNN.
• Running: Execute python gui.py for the GUI or python ./backend/main.py for CLI.
• Links: GUI Demo

Highlighted Details

• Supports multiple inpainting algorithms (STTN, LAMA, PROPAINTER) with configurable parameters for optimizing speed and quality.
• Offers batch processing for images and automatic detection for videos.
• Allows customization of subtitle detection areas.
• Provides options to fine-tune algorithm parameters for specific content types (e.g., real-person videos, animations).

Maintenance & Community

The project lists QQ groups for discussion and support. A table of donors indicates community engagement and financial backing.

Licensing & Compatibility

The README does not explicitly state a license. The project's reliance on NVIDIA GPUs and specific CUDA/cuDNN versions implies compatibility is limited to NVIDIA hardware.

Limitations & Caveats

The project explicitly states it is not compatible with AMD GPUs. Performance and quality are highly dependent on the chosen algorithm and user-configured parameters, with some settings potentially increasing VRAM usage or skipping crucial detection steps. The project also mentions the possibility of needing to train custom models for optimal results.