How Accurate are Video Quality Models for Diffusion-Based Video Super-Resolution?

A new study evaluated video quality models on diffusion-based video super-resolution and found they still cannot match human judgment.

Researchers tested full-reference and no-reference metrics on six upscaling methods using both compressed and uncompressed low-resolution videos.

CNN-based full-reference metrics performed best, but even the strongest models were not accurate enough to replace subjective testing.

The results suggest common metrics like LPIPS, DISTS, and VMAF can miss artifacts in modern diffusion-based upscalers.