notesum.ai

Published at December 5

IF-MDM: Implicit Face Motion Diffusion Model for High-Fidelity Realtime Talking Head Generation

cs.CV

Released Date: December 5, 2024

Authors: Sejong Yang¹, Seoung Wug Oh², Yang Zhou², Seon Joo Kim¹

Aff.: ¹Yonsei University; ²Adobe Research

Arxiv: http://arxiv.org/pdf/2412.04000v1

Refer to caption

	Image Quality	Identity Preservation	Temporal Consistency	Lip-Sync		Speed
	FID $\downarrow$	CSIM $\uparrow$	VideoScore-TC $\uparrow$	LSE-D $\downarrow$	LSE-C $\uparrow$	FPS
Real3DPortrait[30])	74.68	0.982	2.10	8.23	6.58	10.21
AniPortrait[27]	49.13	0.978	2.57	11.56	3.00	0.88
IF-MDM (ours)	42.84	0.984	2.99	11.04	3.88	30.90
Ground Truth	0.00	1.00	2.71	8.48	6.28	-