notesum.ai

Published at November 15

Enhancing Diffusion Posterior Sampling for Inverse Problems by Integrating Crafted Measurements

cs.CV

cs.AI

cs.LG

Released Date: November 15, 2024

Authors: Shijie Zhou¹, Huaisheng Zhu², Rohan Sharma¹, Ruiyi Zhang³, Kaiyi Ji¹, Changyou Chen¹

Aff.: ¹University at Buffalo; ²The Pennsylvania State University; ³Adobe Research

Arxiv: http://arxiv.org/abs/2411.09850v1

Refer to caption

Dataset	Method	Gaussian Deblur				Motion Deblur				super-resolution ( $4\times$ )
Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
FFHQ	DPS-CM (Ours)	27.45	0.7819	0.2059	56.04	30.31	0.8585	0.1609	42.58	27.81	0.7942	0.2008	51.72
	DDRM	25.20	0.7289	0.2976	105.06	-	-	-	-	26.89	0.7821	0.2667	95.23
	DPS	26.35	0.7533	0.2178	65.20	27.84	0.8023	0.2129	70.10	27.55	0.7886	0.2077	57.20
	OPC	27.06	0.7651	0.2196	61.76	26.20	0.7371	0.2535	66.63	26.90	0.7637	0.2400	76.33
	DiffPIR	24.13	0.6649	0.2919	76.96	26.35	0.7394	0.2487	64.25	25.74	0.7143	0.2711	67.19
	PSLD	28.45	0.7689	0.3019	99.90	26.19	0.6779	0.3667	137.88	27.91	0.7783	0.2783	87.82
	ReSample	26.10	0.6441	0.3361	100.21	28.26	0.7221	0.2937	89.80	23.23	0.4517	0.5061	151.20
ImageNet	DPS-CM (Ours)	22.73	0.6147	0.3397	128.92	24.36	0.6814	0.3163	97.54	23.78	0.6450	0.3242	97.38
	DDRM	21.48	0.5527	0.4948	242.38	-	-	-	-	22.72	0.6225	0.4252	196.32
	DPS	16.36	0.3449	0.5041	208.49	17.41	0.3970	0.4920	204.27	21.17	0.5404	0.3668	106.71
	OPC	18.93	0.4254	0.4579	132.38	18.43	0.3782	0.4989	174.44	19.44	0.4113	0.4836	169.71
	DiffPIR	20.62	0.4753	0.4445	154.12	23.38	0.6285	0.3711	120.58	22.97	0.5985	0.3840	119.66
	PSLD	23.45	0.6089	0.3419	131.90	23.18	0.5688	0.4206	167.57	23.79	0.6371	0.3346	120.68
	ReSample	22.79	0.5147	0.4355	172.17	23.95	0.5723	0.3929	125.33	21.04	0.3973	0.5001	203.56