Physion Labs publishes Galileo-0, a critical model of physical consistency for generated video

Physion Labs releases Galileo-0, a model detecting physical flaws in videos from Veo 3, Kling, Sora 2, WAN 2.6 and Seedance with a 63.26% F1 score.

Physion Labs unveils Galileo-0, a system designed to detect and diagnose physical inconsistencies in AI-generated videos. While existing evaluation tools produce an overall score, Galileo-0 precisely identifies what failed, when, where in the image, and why the observed behavior violates the physical rules of the scene. The model relies on a two-stage pipeline: a spatiotemporal perception phase that proposes anomaly candidates, followed by a reasoning phase that determines if each candidate constitutes a true physical violation. It covers four categories of glitches: object existence discontinuity, identity or attribute drift, structural inconsistency, and textual inconsistency. The evaluation videos are sourced from leading current generation models, including Veo 3, Kling, Sora 2, WAN 2.6, and Seedance. On the internal benchmark, Galileo-0 achieves an F1 score of 63.26% across all categories, compared to 38.89% for GPT-5.4 and 36.48% for Gemini 3.1 Pro. The gap is particularly significant in textual inconsistencies, where it reaches 84.10%. Physion Labs states it developed the model, its datasets, and its evaluation pipeline in three months for a budget under $200,000. Google… Pablo Nastar 10 months ago AI is running wild: Midjourney V7, Llama 4, and Nova Sonic… Pablo Nastar 1 year ago