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.
More from Multimodal
- Genie 4 arrives in Project Genie with native audio, real-time editing and a creation agent — GoogleDeepMind · 2026-11-19
- Sora-generated 'waiting for GTA VI' meme video takes off on Reddit — __arex__ · 2026-11-19
- CapCut + Seedance workflow: AI Extend and timeline edits for Black Friday ads — HeyNayeem · 2026-11-19
- Building snowy product reveals from a reference image with AI video tools — HeyNayeem · 2026-11-19
- Seedance 2.5 4K video generation goes global in CapCut PC, 90% off first month — HeyNayeem · 2026-11-19
- AI-generated Shakespeare vs Dostoevsky rap battle stuns with realistic visuals — madebyprospero · 2026-11-19