The authors use a frozen, pretrained spatio-temporal encoder to extract features that capture appearance and change over time.
PhyProbe scores physical inconsistencies in generated videos more reliably
The evaluator combines ranking, regression and calibration supervision to produce continuous, bounded scores that generalize across several video comparison settings.
Big Tech
Max Ku · Jiaojiao Fan · Zekun Hao · Francesco Ferroni · Heng Wang · Wenhu Chen · +2 more
NVIDIA · University of Waterloo · Vector Institute
Research Digest··3 min read
Ku and colleagues introduce PhyProbe, a learned evaluator for measuring the severity of physical violations in generated videos.
Why this paper
From NVIDIA and 2 others
In one line
PhyProbe outperforms prior evaluators on physical consistency benchmarks, especially in generated-generated and no-correspondence settings.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§