ReDiverse fixes diversity collapse in distilled diffusion models via activation steering
ReDiverse · hf · 2026-11-17
A team from UIUC, fal, and Johns Hopkins identifies the mechanism behind diversity collapse in few-step distilled diffusion models and proposes ReDiverse:
- Diagnosis: at the initial denoising step, activation variance in an intermediate block concentrates into fewer directions, causing near-identical samples;
- Fix: reintroduce teacher variance directions via adapter fine-tuning or training-free activation steering;
- Validated on SDXL and SD3 for images and video, restoring diversity while preserving quality—their comparisons show the student model producing nearly identical outputs before, clearly varied outputs after.
The training-free option makes this immediately practical for anyone deploying distilled generators.
Related event: ReDiverse Restores Diversity Lost in Distilled Diffusion Models(2 posts)→
More from Multimodal
- MOSS Generates Stunning Backlit Butterfly, Showcasing Light-Dark Control — DevKkw · 2026-11-17
- VHS-Style AI Horror Short 'Human Alien' Leans Into Analog Horror Aesthetic — brythehuman · 2026-11-17
- Open-source AI video editor Scarlet ships major update: long-to-shorts, pro NLE handoff — MatthewBerman · 2026-11-17
- AI-generated series: Schrödinger explains modern dating in 'The Encounter' Ep 1 — FuzzByteStudios · 2026-11-17
- Seedance 2.5 4K Now Available on fal — petewoodbridge · 2026-11-17
- Higgsfield launches Short Drama Studio: turn a script into a full season — azed_ai · 2026-11-17