notesum.ai

Published at December 4

RFSR: Improving ISR Diffusion Models via Reward Feedback Learning

cs.CV

Released Date: December 4, 2024

Authors: Xiaopeng Sun¹, Qinwei Lin², Yu Gao, Yujie Zhong, Chengjian Feng, Dengjie Li, Zheng Zhao, Jie Hu, Lin Ma

Aff.: ¹Meituan Inc.; ²Tsinghua University

Arxiv: http://arxiv.org/pdf/2412.03268v1

Refer to caption

Datasets	Metrics	DiffBIR	DiffBIR-RFSR	DiffBIR-tag	DiffBIR- tag-RFSR	PASD	PASD-RFSR	SeeSR	SeeSR-RFSR
DIV2K-val	$\text{MANIQA}\uparrow$	0.4869	0.5058	0.5193	0.5341	0.3412	0.4517	0.5091	0.5954
	$\text{MUSIQ}\uparrow$	67.88	69.2538	69.27	70.2611	50.26	59.88	67.40	69.97
	$\text{CLIPIQA}\uparrow$	0.7036	0.7231	0.7114	0.7377	0.4619	0.6056	0.6989	0.7944
	$\text{Aesthetic}\uparrow$	5.0447	5.0973	5.1826	5.2521	4.7717	5.0758	5.1475	5.2683
	$\text{LPIPS}\downarrow$	0.3659	0.3882	0.3556	0.3693	0.4410	0.4277	0.3329	0.3369
DRealSR	$\text{MANIQA}\uparrow$	0.4923	0.4978	0.5229	0.5388	0.3688	0.5130	0.5146	0.5922
	$\text{MUSIQ}\uparrow$	65.73	66.2777	67.44	68.7545	50.18	63.79	64.92	67.48
	$\text{CLIPIQA}\uparrow$	0.6842	0.6830	0.7038	0.7174	0.4872	0.6708	0.6813	0.7596
	$\text{Aesthetic}\uparrow$	4.6101	4.6184	4.7418	4.8171	4.4174	4.6433	4.6985	4.8158
	$\text{LPIPS}\downarrow$	0.3497	0.3933	0.3480	0.3729	0.2413	0.2654	0.2346	0.2761
RealSR	$\text{MANIQA}\uparrow$	0.4857	0.4805	0.5365	0.5431	0.3941	0.5316	0.5428	0.6057
	$\text{MUSIQ}\uparrow$	68.19	68.507	70.04	70.9008	59.83	68.99	69.77	71.22
	$\text{CLIPIQA}\uparrow$	0.6897	0.6862	0.7048	0.7155	0.4788	0.6491	0.6611	0.7438
	$\text{Aesthetic}\uparrow$	4.8153	4.8299	4.8964	4.9212	4.675	4.7879	4.8046	4.8985
	$\text{LPIPS}\downarrow$	0.2760	0.2838	0.2969	0.3153	0.2252	0.2468	0.2354	0.2642