ReDiverse: probing how distilled diffusion models collapse diversity in early denoising layers

anand_bhattad · x · 2026-11-19

Researcher anandbhattad unveiled ReDiverse, applying his probe→analyze→improve methodology to boost diversity in distilled generators.

The team probed activations of several distilled diffusion models across images and videos and found a consistent pattern: a collapse in the model's principal directions, concentrated in one layer and especially early in the denoising process. Compared to the teacher model, the distilled model loses diversity in a very specific way, which the work then targets for improvement.

Related event: ReDiverse Diagnoses and Fixes Diversity Collapse in Distilled Diffusion Models(9 posts)→

Original post →

More from Multimodal

Multimodal channel →