Jinchao Zhu, Yuxuan Wang, Siyuan Pan, Pengfei Wan, Di Zhang...
TL;DR通过模型组装策略重构轻量级模型,以及通过特征继承策略加快 SDM 生成速度。
Abstract
The stable diffusion model (SDM) is a prevalent and effective model for
text-to-image (t2i) and image-to-image (I2I) generation. Despite various
attempts at sampler optimization, model distillation, and network
q
Hyper-SD is a novel framework that combines the advantages of ODE Trajectory Preservation and Reformulation, achieving state-of-the-art performance in multi-step inference with diffusion models.