notesum.ai

Published at October 22

One-Step Diffusion Distillation through Score Implicit Matching

cs.SE

cs.AI

cs.CL

Released Date: October 22, 2024

Authors: Weijian Luo¹, Zemin Huang², Zhengyang Geng³, J. Zico Kolter³, Guo-jun Qi²

Aff.: ¹Peking University; ²Westlake University; ³Carnegie Mellon University

Arxiv: https://arxiv.org/abs/2410.16794v1

Refer to caption

Model	Steps	Type	Params	Aes Score	Image Reward	Pick Score	User Pref	Distill Cost
SD15-Base [57]	25	UNet	860 M	5.26	0.18	0.217
SD15-LCM [39]	4	UNet	860 M	5.66	-0.37	0.212		8 A100 $\times$ 4 Days
SD15-TCD [90]	4	UNet	860 M	5.45	-0.15	0.214		8 A800 $\times$ 5.8 Days
PeRFlow [80]	4	UNet	860 M	5.64	-0.35	0.208		M GPU $\times$ N Days
Hyper-SD15[56]	1	UNet	860 M	5.79	0.29	0.215		32 A100 $\times$ N Days
SDXL-Base [57]	25	UNet	2.6 B	5.54	0.87	0.229
SDXL-LCM [39]	4	UNet	2.6 B	5.42	0.48	0.224		8 A100 $\times$ 4 Days
SDXL-TCD [90]	4	UNet	2.6 B	5.42	0.67	0.226		8 A800 $\times$ 5.8 Days
SDXL-Lightning [34]	4	UNet	2.6 B	5.63	0.72	0.229		64 A100 $\times$ N Days
Hyper-SDXL[56]	4	UNet	2.6 B	5.74	0.93	0.232		32 A100 $\times$ N Days
SDXL-Turbo [63]	1	UNet	2.6 B	5.33	0.78	0.228		M GPU $\times$ N Days
SDXL-Lightning [34]	1	UNet	2.6 B	5.34	0.54	0.223		64 A100 $\times$ N Days
Hyper-SDXL[56]	1	UNet	2.6 B	5.85	1.19	0.231		32 A100 $\times$ N Days
PixArt- $\alpha$ [6]	30	DiT	610 M	5.97	0.82	0.226
SIM-DiT-600M	1	DiT	610 M	6.42	0.67	0.223		4 A100 $\times$ 2 days
PixArt- $\alpha^{*}$ [6]	30	DiT	610 M	5.93	0.53	0.223	54.88%
SIM-DiT-600M^∗	1	DiT	610 M	5.91	0.44	0.223	45.12%	4 A100 $\times$ 2 days