ReSeed: 0.002% extra params in first denoising step restore diversity lost to distillation

kwangmoo_yi · x · 2026-11-05

The paper 'ReSeed: A first-step bias restores lost diversity in diffusion model distillation' tackles diversity loss in distilled diffusion models by training a tiny bias adaptor applied only to the first denoising step, supervised by a teacher target computed with the same prompt and initial noise. The student stays frozen, adds just 0.002–0.004% trainable parameters, needs no extra sampling steps or teacher at inference, and works for both image and video diffusion models.

Original post →

ReSeed: 0.002% extra params in first denoising step restore diversity lost to distillation

More from Multimodal