notesum.ai

Published at December 10

Efficient Diversity-Preserving Diffusion Alignment via Gradient-Informed GFlowNets

cs.LG

cs.CV

Released Date: December 10, 2024

Authors: Zhen Liu¹, Tim Z. Xiao², Weiyang Liu³, Yoshua Bengio¹, Dinghuai Zhang¹

Aff.: ¹Mila, Université de Montréal; ²Max Planck Institute for Intelligent Systems - Tübingen; ³University of Cambridge

Arxiv: http://arxiv.org/pdf/2412.07775v1

Method	Aesthetic Score			HPSv2			ImageReward
Method	Reward ( $\uparrow$ )	Diversity DreamSim ( $\uparrow$ , $10^{-2}$ )	FID ( $\downarrow$ )	Reward ( $\uparrow$ , $10^{-1}$ )	Diversity DreamSim ( $\uparrow$ , $10^{-2}$ )	FID ( $\downarrow$ )	Reward ( $\uparrow$ , $10^{-1}$ )	Diversity DreamSim ( $\uparrow$ , $10^{-2}$ )	FID ( $\downarrow$ )
Base Model	5.83 $\scriptscriptstyle\pm$ 0.01	35.91 $\scriptscriptstyle\pm$ 0.00	216 $\scriptscriptstyle\pm$ 1	2.38 $\scriptscriptstyle\pm$ 0.13	37.75 $\scriptscriptstyle\pm$ 0.21	563 $\scriptscriptstyle\pm$ 5	-0.38 $\scriptscriptstyle\pm$ 0.12	41.09 $\scriptscriptstyle\pm$ 0.03	468 $\scriptscriptstyle\pm$ 1
DDPO	6.68 $\scriptscriptstyle\pm$ 0.14	32.96 $\scriptscriptstyle\pm$ 1.04	312 $\scriptscriptstyle\pm$ 9	2.52 $\scriptscriptstyle\pm$ 0.04	3.49 $\scriptscriptstyle\pm$ 0.03	681 $\scriptscriptstyle\pm$ 16	0.27 $\scriptscriptstyle\pm$ 0.38	38.51 $\scriptscriptstyle\pm$ 1.49	714 $\scriptscriptstyle\pm$ 25
ReFL	9.53 $\scriptscriptstyle\pm$ 0.46	8.20 $\scriptscriptstyle\pm$ 3.06	1765 $\scriptscriptstyle\pm$ 51	3.67 $\scriptscriptstyle\pm$ 0.06	19.84 $\scriptscriptstyle\pm$ 1.70	1191 $\scriptscriptstyle\pm$ 46	1.36 $\scriptscriptstyle\pm$ 0.30	36.50 $\scriptscriptstyle\pm$ 0.52	597 $\scriptscriptstyle\pm$ 10
DRaFT-1	10.16 $\scriptscriptstyle\pm$ 0.13	4.24 $\scriptscriptstyle\pm$ 0.45	1665 $\scriptscriptstyle\pm$ 182	3.70 $\scriptscriptstyle\pm$ 0.06	18.96 $\scriptscriptstyle\pm$ 1.35	1222 $\scriptscriptstyle\pm$ 84	1.59 $\scriptscriptstyle\pm$ 0.25	37.27 $\scriptscriptstyle\pm$ 0.49	531 $\scriptscriptstyle\pm$ 13
DRaFT-LV	10.21 $\scriptscriptstyle\pm$ 0.34	6.39 $\scriptscriptstyle\pm$ 1.66	1854 $\scriptscriptstyle\pm$ 296	3.75 $\scriptscriptstyle\pm$ 0.08	21.13 $\scriptscriptstyle\pm$ 1.19	1164 $\scriptscriptstyle\pm$ 43	1.44 $\scriptscriptstyle\pm$ 0.25	37.56 $\scriptscriptstyle\pm$ 0.09	529 $\scriptscriptstyle\pm$ 18
DAG-DB	7.73 $\scriptscriptstyle\pm$ 0.07	15.88 $\scriptscriptstyle\pm$ 0.70	595 $\scriptscriptstyle\pm$ 87	2.52 $\scriptscriptstyle\pm$ 0.06	32.50 $\scriptscriptstyle\pm$ 0.59	866 $\scriptscriptstyle\pm$ 41	4.70 $\scriptscriptstyle\pm$ 0.75	26.78 $\scriptscriptstyle\pm$ 0.64	809 $\scriptscriptstyle\pm$ 34
residual DB	7.20 $\scriptscriptstyle\pm$ 0.92	19.38 $\scriptscriptstyle\pm$ 4.07	1065 $\scriptscriptstyle\pm$ 587	2.55 $\scriptscriptstyle\pm$ 0.06	32.49 $\scriptscriptstyle\pm$ 0.47	840 $\scriptscriptstyle\pm$ 107	6.47 $\scriptscriptstyle\pm$ 0.54	25.52 $\scriptscriptstyle\pm$ 0.35	772 $\scriptscriptstyle\pm$ 19
$\nabla$ -GFlowNet ( $\scriptscriptstyle w_{B}=0$ )	7.86 $\scriptscriptstyle\pm$ 0.06	29.21 $\scriptscriptstyle\pm$ 0.18	318 $\scriptscriptstyle\pm$ 13	3.53 $\scriptscriptstyle\pm$ 0.03	24.40 $\scriptscriptstyle\pm$ 0.56	973 $\scriptscriptstyle\pm$ 5	5.33 $\scriptscriptstyle\pm$ 0.62	35.85 $\scriptscriptstyle\pm$ 0.66	638 $\scriptscriptstyle\pm$ 15
$\nabla$ -GFlowNet ( $\scriptscriptstyle w_{B}=1$ )	7.90 $\scriptscriptstyle\pm$ 0.09	29.67 $\scriptscriptstyle\pm$ 0.51	317 $\scriptscriptstyle\pm$ 15	3.53 $\scriptscriptstyle\pm$ 0.04	24.39 $\scriptscriptstyle\pm$ 0.87	1000 $\scriptscriptstyle\pm$ 39	2.23 $\scriptscriptstyle\pm$ 0.34	39.85 $\scriptscriptstyle\pm$ 0.67	501 $\scriptscriptstyle\pm$ 7