notesum.ai

Published at November 25

Efficient Video Face Enhancement with Enhanced Spatial-Temporal Consistency

cs.CV

Released Date: November 25, 2024

Authors: Yutong Wang¹, Jiajie Teng², Jiajiong Cao³, Yuming Li³, Chenguang Ma³, Hongteng Xu⁴, Dixin Luo¹

Aff.: ¹Beijing Institute of Technology; ²Zhejiang University; ³Unknown; ⁴Renmin University of China

Arxiv: http://arxiv.org/abs/2411.16468v1

Refer to caption

Task	Method	Quality and Fidelity			Pose Consistency			Temporal Consistency		Efficiency
Task	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AKD $\downarrow$	Face-Cons $\uparrow$	IDS $\uparrow$	FVD $\downarrow$	Flow-Score $\downarrow$	Runtime(s) $\downarrow$
BFIR	VQFR [14]	25.94	0.7852	0.2467	5.978	0.9947	0.6659	388.2	1.451	15.60
	GFPGAN [45]	27.15	0.8207	0.2279	4.134	0.9950	0.9206	246.9	1.316	14.44
	CodeFormer [58]	26.77	0.8102	0.2373	4.543	0.9947	0.8596	261.8	2.672	28.18
VSR	BasicVSR++ [4]	27.22	0.8218	0.2742	5.129	0.9965	0.9234	392.7	1.286	72.21
VSR	Real-BasicVSR [5]	27.45	0.7929	0.2968	4.780	0.9936	0.8785	305.7	1.404	12.20
BFVR	PGTFormer [49]	28.68	0.8426	0.1752	3.519	0.9942	0.9296	107.6	1.154	7.085
BFVR	KEEP [12]	27.04	0.8223	0.2370	3.979	0.9953	0.8783	264.9	1.302	19.01
\hdashline	Ours	27.47	0.8641	0.1829	3.858	0.9954	0.9312	105.1	1.150	2.995