Real-ESRGAN

Real-ESRGAN is best suited for upscaling degraded photos, cleaning JPEG compression artifacts from web assets, sharpening digital illustrations and anime, and restoring facial details in portraits. It uses high-order degradation modeling to reconstruct high-frequency textures and crisp line boundaries without introducing blur.

Select RealESRGAN_x4plus_anime_6B for stylized anime, manga panels, and 2D digital illustrations. This specialized 6-block architecture processes line art quickly while preventing halo and ringing artifacts around high-contrast ink boundaries.

Enable face enhancement when upscaling images containing real human faces that suffer from blur, compression, or distortion. This triggers an integrated GFPGAN facial restoration pass that reconstructs eyes, mouths, and skin micro-textures. Avoid enabling it on abstract art or stylized cartoons, as it can impose uncanny realistic facial features on non-human subjects.

Real-ESRGAN does not convert raster artwork into vector SVG files, struggles with heavily pixelated tiny text, and cannot correct extreme motion blur. Because it is a deterministic upscaler rather than a generative diffusion model, it enhances and sharpens existing structures rather than hallucinating entirely new compositions.

Set the tile parameter between 200 and 400 when processing images larger than 2048x2048 pixels. Tiling splits the canvas into manageable patches during processing, allowing you to upscale up to 8x without exhausting hardware memory.

Real-ESRGAN was developed by Tencent ARC Lab and the Shenzhen Institutes of Advanced Technology, initially introduced at ICCVW in 2021.