ReDiverse finds distilled image/video generators collapse variance into fewer directions — and fixes it

anand_bhattad · x · 2026-11-19

Distilling diffusion models into few-step students causes severe diversity collapse: different seeds yield nearly identical samples. By probing internal activations, the authors show students retain teacher-level total activation variance but concentrate it into far fewer directions, localized to an intermediate block at the first denoising step — a variance misallocation, not a loss of generative capacity. They propose ReDiverse: ReDiverse-FT fine-tunes a lightweight adapter regressing student activations onto the teacher's, while ReDiverse-TF is a training-free inference method steering activations along precomputed teacher principal directions. Across image and video distillation benchmarks, ReDiverse consistently restores diversity while preserving quality (backed by user studies). Paper, code and models are open-sourced; the project was led by a second-year undergrad Furkan Horoz with collaborators from fal and Siebel School.

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

Original post →

More from Multimodal

Multimodal channel →