notesum.ai

Published at November 26

Contrastive CFG: Improving CFG in Diffusion Models by Contrasting Positive and Negative Concepts

cs.LG

cs.AI

cs.CV

Released Date: November 26, 2024

Authors: Jinho Chang¹, Hyungjin Chung¹, Jong Chul Ye¹

Aff.: ¹KAIST, Seoul, South Korea

Arxiv: http://arxiv.org/abs/2411.17077v1

[Uncaptioned image]

	Negative	pos. prompt[ $\uparrow$ ]		neg. prompt[ $\downarrow$ ]
	Guidance	HPS-v2	GPT4	HPS-v2	GPT4
(a)	None	0.260	1.00	0.213	0.10
	DNG	0.258	0.99	0.211	0.09
	nCFG	0.220	0.66	0.158	0.00
	CCFG	0.251	0.99	0.184	0.00
(b)	None	0.322	0.99	0.232	0.75
	DNG	0.312	1.00	0.219	0.31
	nCFG	0.297	0.85	0.192	0.01
	CCFG	0.314	1.00	0.204	0.02
(c)	None	0.258	0.99	0.192	0.45
	DNG	0.257	0.99	0.189	0.38
	nCFG	0.241	0.90	0.173	0.01
	CCFG	0.246	0.96	0.175	0.04
(d)	None	0.254	1.00	0.271	0.94
	DNG	0.252	1.00	0.266	0.87
	nCFG	0.222	0.89	0.179	0.00
	CCFG	0.235	0.99	0.205	0.03