Hysen LabsReal-ESRGAN
Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration.
Real-ESRGAN ships as executables with no CUDA required
An image restoration project from Tencent ARC Lab that extends ESRGAN with pure synthetic training data, and that hands you portable binaries with the models already baked in.
From ESRGAN to something usable
The aim is practical algorithms for general image and video restoration. The approach extends the powerful ESRGAN into a practical restoration application trained with pure synthetic data, the honest way of saying the training pairs are degraded images rather than laboriously collected ones. Development happened at Tencent ARC Lab together with the Shenzhen Institutes of Advanced Technology of the Chinese Academy of Sciences. Python, BSD-3-Clause.
Anime gets special treatment
Two model lines target animated material. AnimeVideo-v3 works on anime videos, and the RealESRGAN x4plus anime model handles anime images with a much smaller model size, with comparisons against waifu2x referenced along the way. A realesr-general-x4v3 model covers general scenes and adds a denoising strength option.
Try it before you install anything
For the curious, the README links an online Replicate demo, a Colab demo for Real-ESRGAN, and another Colab demo aimed at anime videos. A website demo currently supports a specific anime model. That is a low friction path to judging the output quality yourself before committing to any setup.
The portable builds
Portable executables exist for Windows, Linux, and macOS, covering Intel, AMD, and Nvidia GPUs, and they include all required binaries and models, no CUDA or PyTorch environment needed. Named models include realesrgan-x4plus, realesrnet-x4plus, realesrgan-x4plus-anime, and realesr-animevideov3. The ncnn implementation lives in a separate Real-ESRGAN-ncnn-vulkan project.
Neighbors and downstream users
A Python script exposes an outscale argument for arbitrary output sizes alongside the pretrained models. Recommended companions include GFPGAN for face restoration, BasicSR as a restoration toolbox, and facexlib for face functions. Downstream, NCNN Android apps and VapourSynth plugins build on Real-ESRGAN, and the repo carries around 36,500 stars.
Editorial conclusion
Restoration quality always deserves an eyeball test, and the online demos make that easy. The portable executables remove the last excuse for not trying it.
Community notes