Download
Real-ESRGAN is a highly popular open-source project that provides practical algorithms for general image and video restoration using deep learning-based super-resolution techniques. It extends the original Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) approach by training on synthetic degradations to make results more robust on real-world images, effectively enhancing resolution, reducing noise/artifacts, and reconstructing fine detail in low-quality imagery. The repository includes inference and training scripts, a model zoo with different pretrained models (including general and anime-oriented variants), and support for batch and arbitrary scaling, making it adaptable for diverse enhancement tasks. It emphasizes usability with utilities that handle alpha channels, gray/16-bit images, and tiled inference for large inputs, and can be run via Python scripts or portable executables.
Features
- Deep-learning super-resolution and general image/video restoration
- Extensive pretrained model zoo for general images and anime/video content
- Flexible inference: Python scripts or portable executables for multiple platforms
- Handles alpha channels, gray/16-bit images, and tiled large-image processing
- Supports arbitrary scaling and optional face enhancement integration
- Designed for robust real-world performance on diverse image degradations
Project Samples
