Real-ESRGAN

Real-ESRGAN is a local AI upscaler for people enlarging images or restoring degraded images and video. It builds on ESRGAN, with dedicated models for general images, anime illustrations and animated video.

• General image restoration: RealESRGAN_x4plus enlarges images using a model trained on synthetic degradation data.
• Anime models: RealESRGAN_x4plus_anime_6B targets anime illustrations, while AnimeVideo-v3 handles animation footage.
• Output sizing: The Python implementation can produce sizes beyond the model's native enlargement scale by applying additional resizing.

The code is open source under the BSD 3-Clause license. Its Python implementation uses PyTorch. Portable builds provided through Real-ESRGAN-ncnn-vulkan run on Windows, Linux and macOS with Intel, AMD or NVIDIA GPUs. They include the required models and binaries and don't need a CUDA or PyTorch environment.

Local builds process media on your machine. Replicate and Colab demos provide online alternatives. The portable implementation doesn't support every Python capability, including arbitrary output sizing. Separate graphical applications that use Real-ESRGAN include Upscayl and Waifu2x-Extension-GUI.