Upscaling

The upscaling stage uses AI super-resolution to increase frame resolution. Unlike traditional interpolation (bicubic, Lanczos), Real-ESRGAN generates new detail by predicting what the high-resolution original would have looked like.

Scale factor

The default 4x scale factor converts 1080p to 4K. A 2x option is available for less aggressive enhancement or when VRAM is limited. The scale factor is applied uniformly to both width and height.

Denoise strength

Real-ESRGAN includes a built-in denoise parameter. This is separate from the preprocessing denoise stage. The upscale denoise is applied during the super-resolution process itself, which can produce better results than denoising before upscaling because the model accounts for noise in its prediction. The trade-off is that aggressive denoising may soften fine texture.

Model variants

The standard Real-ESRGAN 4x model is trained on photographic content. The anime variant is optimized for cel animation, flat color fills, and line art — it preserves sharp edges and flat regions that the standard model might add unwanted texture to.

When to skip upscaling

If your source is already at your target resolution (e.g., 4K source for 4K output), skip the upscale stage. Running upscaling on content that does not need it wastes GPU time and can introduce subtle artifacts.