Vchitect /VBench

About VBench

VBench rates video generation models across many named quality dimensions. It's for researchers and teams comparing models, not for making videos. You'll need to download prompt suites, sample videos and pretrained scoring models. Some parts are still in beta, so expect rough edges.

Best for: Model-selection teams who need to compare video models on named dimensions.

• Scores text-to-video models on quality, faithfulness and safety
• Adds image-to-video scoring with an adaptive image suite
• Handles long videos, like Sora-style clips
• Checks fairness, bias, culture and safety in outputs
• VBench-2.0 adds 18 fine-grained dimensions in five categories
• Ships prompt suites, sample videos and a human anomaly detector

1. 1
   Install the package with pip install vbench.
2. 2
   Clone the repo to get the prompt suites, sample videos and scoring scripts.
3. 3
   Open the Usage section of the README and run the evaluation on your own clips or the sample videos.
4. 4
   For image-to-video, long video or trustworthiness scoring, use the folders vbench2_beta_i2v, vbench2_beta_long and vbench2_beta_trustworthiness.
5. 5
   For VBench-2.0 dimensions, use the VBench-2.0 folder.

• VBench vs Awesome-Video-Diffusion
• VBench vs awesome-video-generation