Real-ESRGAN

Real-ESRGAN restores and upscales real-world images and video. Developed at Tencent ARC Lab, it extends ESRGAN by training purely on synthetic degradations designed to mimic the noise, blur, and compression artifacts found in real photos, which makes it practical outside lab conditions. The default model upscales 4x with arbitrary output scales, dedicated anime variants including AnimeVideo-v3 handle illustrations and animated footage, and optional GFPGAN integration enhances faces in the same pass. It supports alpha channels, grayscale, and 16-bit images, and tile-based processing keeps memory use manageable on smaller GPUs. Beyond the Python package with pretrained weights, portable NCNN Vulkan executables for Windows, Linux, and macOS run without any Python environment, and numerous third-party GUIs build on it. Released under the BSD-3-Clause license, it is widely used by content creators, archivists, and developers who need dependable upscaling.