Video Diffusion Models Encode Physical Plausibility Signals
A recent investigation published on arXiv (ID 2603.14294) explores the capability of video diffusion models to capture signals indicative of physical plausibility. Researchers examined the intermediate denoising representations of pretrained Diffusion Transformers (DiTs) and found that videos deemed physically plausible can be partially distinguished from those that are not, even amidst high noise levels. Controls for within-source and perceptual quality indicate that this signal is not solely determined by generator identity or overall visual quality. The team created a compact, backbone-specific physics verifier, utilizing two complementary inference-time strategies within a fixed multi-trajectory framework: progressive trajectory selection and reward-gradient guidance. Experiments were performed on PhyGenBench and Physics-IQ benchmarks, revealing that diffusion models may inherently learn physical principles, enhancing the realism of generated videos. The full paper can be accessed at https://arxiv.org/abs/2603.14294.
Key facts
- Study probes intermediate denoising representations of pretrained Diffusion Transformers (DiTs).
- Physically plausible and implausible videos are partially separable in mid-layer feature space at high noise levels.
- Signal is not fully explained by generator identity or generic visual quality.
- Distilled signal into a lightweight, backbone-specific physics verifier trained on frozen features.
- Two inference-time mechanisms: progressive trajectory selection and reward-gradient guidance.
- Experiments conducted on PhyGenBench and Physics-IQ benchmarks.
- Paper available on arXiv with ID 2603.14294.
- Findings suggest diffusion models implicitly encode physical plausibility.
Entities
Institutions
- arXiv