Seeking Physics in Diffusion Noise

1. 1Analyze intermediate denoising representations of a pretrained Diffusion Transformer (DiT)
2. 2Probe mid-layer feature space across noise levels to identify separability of physically plausible and implausible videos
3. 3Investigate whether separability can be attributed to visual quality or generator identity
4. 4Explore recoverable physics-related cues in frozen diffusion models

AI engineers and researchers working on computer vision and diffusion models can benefit from this study, as it provides insights into the capabilities and limitations of these models

💡 Diffusion models can capture physically plausible signals, even in noise

💡 Diffusion models can encode physics-related cues, enabling separation of plausible & implausible videos