PhyGround Project Page GitHub Paper PhyGround is a criteria grounded benchmark for evaluating physical reasoning in video generation. The benchmark contains 250 curated prompts, each augmented with an expected physical outcome, and a taxonomy of 13 physical laws across solid body mechanics, fluid dynamics, and optics. Sample Usage You can download the benchmark prompts and first frame images using the Hugging Face CLI: Contents Total: 250 prompts × 10 generation configs = 2,500 videos. The two LTX 2 dev models each ship both one stage and two stage generations on the same 250 prompts. Prompt JSON prompts/phyground.json — top level JSON array, 250 entries: File naming ( id stem ) id stem is a single shared filename stem that maps to both the video files and the first frame image: videos/ / .mp4 — generated video for each of the 8 models first images/ .jpg — first frame / conditioning image (one per prompt) So given a prompt entry with "id stem": "collision 156" : first images/ contains the conditioning frame used as input by the image to video (I2V) models in this release (e.g. wan2.2 i2v a14b ). Text to video (T2V) models do not consume these images. Human annotations annotations/…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy