ReDiverse: probing distilled diffusion models reveals principal-direction collapse — PCA + light tuning restores diversity

anand_bhattad · x · 2026-11-19

Anand Bhattad shares ReDiverse, applying his Probe→Analyze→Improve loop to diversity loss in distilled diffusion generators. Probing activations across image and video models found a consistent pattern: principal directions collapse, concentrated in one layer early in denoising, explaining the diversity gap vs teachers. The fix: project these directions back with PCA, plus lightweight tuning focused on early denoising steps. The approach generalizes across many models; the project was led by sophomore Furkan Gozukeser with fal and Siebel School collaborators.

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

Original post →

More from Multimodal

Multimodal channel →