Feature Information Dynamics in Diffusion
The paper proposes a way to measure when specific features emerge during diffusion denoising.
The authors use the I-MMSE identity to tie feature information change to the gap between unconditional and feature-conditioned denoising losses. They add a chained decomposition to separate shared and incremental information across feature hierarchies. They test it on spectral autoregression in pixel diffusion, then compare class, mask, and Canny conditioning across pixel, SDVAE, VAVAE, and RAE representations. The result points to different feature-generation dynamics across representations, with ordered generation suggested as useful for training diffusion models. ArXiv · AI/CL/LG's note
The authors use the I-MMSE identity to tie feature information change to the gap between unconditional and feature-conditioned denoising losses. They add a chained decomposition to separate shared and incremental information across feature hierarchies. They test it on spectral autoregression in pixel diffusion, then compare class, mask, and Canny conditioning across pixel, SDVAE, VAVAE, and RAE representations. The result points to different feature-generation dynamics across representations, with ordered generation suggested as useful for training diffusion models. ArXiv · AI/CL/LG's note
score 5