FAST-VQA (ECCV 2022) is a deep learned no-reference video-quality model designed for speed on live and user-generated content, where there is no pristine original and often only seconds to produce a score. Its central technique is fragment sampling: instead of scoring every pixel of every frame, it evaluates a handful of small space-time patches sampled from the clip, which cuts cost enough for near-real-time use while preserving much of the accuracy. It looks only at the decoded pixels and is among the accuracy leaders for blind video quality, reaching roughly 0.83 SROCC with human scores on in-distribution large UGC datasets such as LSVQ, alongside DOVER. Its catch is the latency-versus-everything trade and the limit all learned blind models share: published accuracy is in-distribution only and collapses on content unlike its training set, so it must be re-validated on your own clips. It needs a GPU, carries a research licence, and every score should be reported as a band.
