notesum.ai

Published at November 14

Multi-scale Generative Modeling for Fast Sampling

cs.AI

Released Date: November 14, 2024

Authors: Xiongye Xiao¹, Shixuan Li¹, Luzhe Huang², Gengshuo Liu¹, Trung-Kien Nguyen¹, Yi Huang³, Di Chang¹, Mykel J. Kochenderfer⁴, Paul Bogdan¹

Aff.: ¹University of Southern California; ²University of California, Los Angeles; ³University of Chinese Academy of Sciences, Shenzhen; ⁴Stanford University

Arxiv: http://arxiv.org/abs/2411.09356v1

Methods	FID $\downarrow$				Parameters $\downarrow$	Sampling Time(s) $\downarrow$
Methods	CelebA-HQ (5K)	CelebA-HQ (30K)	AFHQ-Cat (5K)	Colon	Parameters $\downarrow$	Sampling Time(s) $\downarrow$
SGM	90.83	78.50	80.93	109.23	160M	15.093
WSGM	49.97	26.74	17.12	54.35	351M	11.097
WMGM(Ours)	30.58	25.38	16.29	45.76	89M	2.31