ReSeed: tiny first-step biases restore diversity lost in diffusion distillation
kwangmoo_yi · x · 2026-11-05
Tang et al.'s paper ReSeed tackles diversity loss in diffusion model distillation.
- Distillation (e.g. DMD) collapses sample diversity: 4-step outputs look far more alike than the original 80-step SD3.5-M.
- The fix: train a tiny set of biases for the earliest denoising steps (e.g. step 1), reintroducing diversity at negligible cost.
- Qualitative comparisons show DMD + ReSeed (4 NFEs) recovers much of the original sample variety.
Related event: ReSeed: A First-Step Bias Restores Diversity Lost in Diffusion Distillation(2 posts)→
More from Multimodal
- Open-source text-to-image model AuroraIMG2-100M released, GenEval 2 now near SDXL — Tall_Abrocoma_3533 · 2026-11-05
- VFX artist tests Firefly Quick Cut: turning unrelated clips into one action showreel — LudovicCreator · 2026-11-05
- MiniMax-generated Warcraft x Skyrim fan film segment starring Grommash Hellscream — Dangerous_Tip_5023 · 2026-11-05
- Midjourney v9 coming by year-end with sound-enabled video, real-time, 3D and an LLM — koltregaskes · 2026-11-05
- AI fantasy short film 'The Abyssal Oracle — Awakening the Leviathan' showcases polished video generation — Andr3wITA · 2026-11-05
- Four parallel approaches emerging to solve persistence in AI video generation — cocktailpeanut · 2026-11-05