Animation vs realistic video upscaling

One network family, three weight sets

Animation uses the Anime4K weights distributed with WebSR. Realistic and 3D use the corresponding retrained WebSR weights on the same network architectures. The upstream weight documentation describes training those variants on Xiph material. This is not a claim that the weights are optimal for every live-action or generated video. Model names and the exact distributed files are recorded in the licence notices.

Lines and textures behave differently

Drawn animation often contains flat fills and deliberate outlines. A network can make those outlines more distinct, but can also thicken delicate line work. Live-action scenes contain grain, pores, fabric and out-of-focus areas. An aggressive result can make skin look unnatural or emphasize noise. For 3D scenes, inspect fine geometry and specular highlights for shimmering during movement.

Compare the same frame under different content types. Choose the version that preserves the intended look rather than simply the one that has the strongest edges. A comparison frame cannot reveal every temporal artifact; export a short test when a scene is important.

What Fast, Balanced and Quality mean

Fast selects cnn-2x-s, Balanced selects cnn-2x-m, and Quality selects cnn-2x-l. These names describe actual networks, not a simulated progress speed. Auto chooses one using input size, duration and exposed device capabilities. Larger models require more GPU memory. If a network cannot safely allocate its working buffers, select Fast or reduce the input resolution.