search

vimar-gu / MinimaxDiffusion

[CVPR2024] Efficient Dataset Distillation via Minimax Diffusion
74 stars 7 forks source link

Request for useful information about skip step diffusion #2

Closed vvJesse closed 4 months ago

vvJesse commented 5 months ago

https://github.com/vimar-gu/MinimaxDiffusion/blob/969fc6f0a5e392d22fc2da39d8a130a4c55c6033/diffusion/respace.py#L67-L81

Hi I just find out here you mention skip step in diffusion model, which I am not familiar with. It seems that you do not use all time steps right? Would you please give me more explanation/reference papers about this method? Thank you.

vimar-gu commented 5 months ago

Thanks for your interest. The respacing is used to accelerate the inference stage with fewer sampling steps. Then the scheduling parameter alpha also requires to be resampled according to the fewer steps. You can refer to Improved Denoising Diffusion Probabilistic Models and Denoising Diffusion Implicit Models for more details on the acceleration.