notesum.ai

Published at December 4

TASR: Timestep-Aware Diffusion Model for Image Super-Resolution

cs.CV

Released Date: December 4, 2024

Authors: Qinwei Lin¹, Xiaopeng Sun², Yu Gao², Yujie Zhong², Dengjie Li², Zheng Zhao², Haoqian Wang¹

Aff.: ¹Tsinghua University; ²Meituan Inc.

Arxiv: http://arxiv.org/pdf/2412.03355v1

Refer to caption

Datasets	Metrics	GAN-based methods			Diffusion-based methods
Datasets	Metrics	RealESRGAN	BSRGAN	SwinIR	SeeSR	PASD	ResShift	DiffBIR	SUPIR	Ours
DIV2K-val	$\text{PSNR}\uparrow$	22.34	22.84	23.29	21.53	21.46	22.02	21.24	20.61	20.92
	$\text{SSIM}\uparrow$	0.5768	0.5969	0.6083	0.5356	0.5321	0.5485	0.5168	0.4717	0.5174
	$\text{LPIPS}\downarrow$	0.3370	0.4851	0.4951	0.3311	0.4408	0.3644	0.3691	0.4203	0.3762
	$\text{MANIQA}\uparrow$	0.3892	0.2500	0.2306	0.5094	0.3558	0.3540	0.4573	0.4861	0.6007
	$\text{MUSIQ}\uparrow$	57.50	37.71	31.32	67.62	52.63	55.56	67.43	54.97	68.14
	$\text{CLIPIQA}\uparrow$	0.5365	0.2869	0.3110	0.6987	0.4917	0.5479	0.7110	0.6374	0.7681
RealSR	$\text{PSNR}\uparrow$	25.69	27.27	27.42	25.18	26.56	26.42	25.22	23.74	23.79
	$\text{SSIM}\uparrow$	0.7618	0.7983	0.7999	0.7200	0.7614	0.7569	0.7028	0.6631	0.6650
	$\text{LPIPS}\downarrow$	0.2172	0.2312	0.2440	0.2354	0.2217	0.2385	0.2577	0.2871	0.2986
	$\text{MANIQA}\uparrow$	0.3743	0.3269	0.2816	0.5427	0.3875	0.3969	0.4583	0.5025	0.6113
	$\text{MUSIQ}\uparrow$	60.17	53.12	45.22	69.78	59.10	60.18	66.62	61.56	69.94
	$\text{CLIPIQA}\uparrow$	0.4444	0.2952	0.3166	0.6611	0.4829	0.5563	0.6700	0.6565	0.7076
DRealSR	$\text{PSNR}\uparrow$	28.64	29.91	30.36	28.17	29.05	28.78	26.87	25.00	27.25
	$\text{SSIM}\uparrow$	0.8052	0.8394	0.8496	0.7674	0.793	0.7878	0.7116	0.6416	0.7381
	$\text{LPIPS}\downarrow$	0.2121	0.2569	0.2571	0.2346	0.2371	0.2508	0.3045	0.3340	0.2869
	$\text{MANIQA}\uparrow$	0.3448	0.2771	0.2621	0.5146	0.3753	0.3517	0.4548	0.4980	0.5551
	$\text{MUSIQ}\uparrow$	54.17	41.76	36.48	64.93	52.60	52.49	63.34	59.66	65.23
	$\text{CLIPIQA}\uparrow$	0.4418	0.3145	0.3460	0.6810	0.5035	0.5429	0.6680	0.6791	0.7155