VBench

Why we included this project

If you build or maintain text-to-video models, the hardest part is often telling whether a fresh version is genuinely better or just different. VBench addresses that by defining "video generation quality" as 16 separate dimensions, from subject consistency and motion smoothness to temporal flickering and spatial relationships, each with its own prompt suite and automated scoring pipeline. Because the benchmark was validated against human preference annotations across every dimension, its scores tend to line up with how real viewers perceive the output, which matters when you are comparing model candidates or chasing regressions. The repository ships the full prompt sets, generated sample videos, and unified evaluation code, and it extends to VBench-2.0 for broader capability coverage. For head-to-head comparisons, debugging a specific failure mode, or backing up a paper or product evaluation with credible numbers, this is a practical toolkit rather than a research toy.