t2v_metrics

Why we included this project

Judging whether a generated image, video, or 3D scene actually matches the prompt is often the slowest part of working with generative models, and this library turns that into a one-line Python call. It implements VQAScore, which has a vision-language model answer a yes/no question about how well the visual aligns with the text, and the resulting score tracks human judgment more closely than older CLIP-based measures. You can run it over a batch of outputs in a few lines, with backends ranging from open-weight models like Qwen and Gemma to hosted GPT and Gemini. The repo also bundles the GenAI-Bench and CameraBench benchmarks, so you get reference points for comparison instead of having to build your own evaluation setup. That makes it a practical choice for anyone iterating on model versions or researching generative visual systems.