notesum.ai

Published at November 25

MVGenMaster: Scaling Multi-View Generation from Any Image via 3D Priors Enhanced Diffusion Model

cs.CV

Released Date: November 25, 2024

Authors: Chenjie Cao¹, Chaohui Yu², Shang Liu², Fan Wang³, Xiangyang Xue⁴, Yanwei Fu⁴

Aff.: ¹Fudan University, DAMO Academy, Alibaba Group, Hupan Lab; ²DAMO Academy, Alibaba Group, Hupan Lab; ³DAMO Academy, Alibaba Group; ⁴Fudan University

Arxiv: http://arxiv.org/abs/2411.16157v1

Condition views (all)	Tanks-and-Temples			DTU			MipNeRF-360
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
2-view (25)
ViewCrafter-sparse [82]	13.408	0.416	0.462	12.411	0.406	0.525	12.966	0.233	0.600
CAT3D* [20]	12.525	0.375	0.531	11.756	0.354	0.618	12.900	0.211	0.618
MVGenMaster	14.790	0.491	0.332	15.574	0.536	0.325	13.836	0.287	0.498
3-view (100)
MVSplat [13]	8.602	0.190	0.649	10.772	0.271	0.557	11.379	0.171	0.691
CAT3D* [20]	11.758	0.351	0.745	11.268	0.365	0.662	13.609	0.263	0.714
MVGenMaster	14.669	0.473	0.473	15.856	0.585	0.314	15.543	0.356	0.539