PhyGround: leaderboard
PhyGround (2026) asks video generation models to continue a first frame under a prompt that states the expected physical outcome, and has hundreds of screened annotators rate the 2,000 resulting videos from 1 to 5. The overall score weights the three general dimensions and the per-law physics ratings equally; higher is better.
Metric: Overall human rating (1-5): half the mean of the three general dimensions (semantic alignment, physical temporal validity, object persistence) plus half the per-law physics rating pooled over solid-body, fluid and optical laws, human ratings on a 1-5 scale from a quality-controlled pool of 352 annotators (after filtering 459), on 2,000 videos generated from 250 text-plus-first-frame prompts with explicit physical outcomes, each model at its released inference defaults; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Veo 3.1 | 3.28 | |
| 2 | Wan2.2-27B-A14B | 3.28 | |
| 3 | OmniWeaving | 3.1 | |
| 4 | Cosmos-Predict2.5-14B | 2.91 | |
| 5 | LTX-2.3-22B | 2.69 | |
| 6 | Wan2.2-TI2V-5B | 2.63 | |
| 7 | Cosmos-Predict2.5-2B | 2.58 | |
| 8 | LTX-2-19B | 2.56 |
Interactive version: theaggregate.ai/benchmark?slug=phyground · How It Works · Data refreshed daily, snapshot 2026-11-21.