notesum.ai

Published at December 10

Multi-Shot Character Consistency for Text-to-Video Generation

cs.CV

Released Date: December 10, 2024

Authors: Yuval Atzmon¹, Rinon Gal, Yoad Tewel, Yoni Kasten, Gal Chechik

Aff.: ¹NVIDIA

Arxiv: http://arxiv.org/pdf/2412.07750v1

Refer to caption

	Multi-Shot Consistency	Text Similarity	Dynamic Degree	Subject Consistency
ConsiS Im2Vid	63.7 $\pm$ 1.4	27.3 $\pm$ 0.5	3.3 $\pm$ 1.5	99.1 $\pm$ 0.1
VSTAR	83.9 $\pm$ 1.6	19.8 $\pm$ 0.4	90.7 $\pm$ 2.4	92.6 $\pm$ 0.3
Tokenflow	65.3 $\pm$ 1.5	27.9 $\pm$ 0.4	26.0 $\pm$ 3.6	97.7 $\pm$ 0.2
VideoCrafter2	63.2 $\pm$ 1.7	28.7 $\pm$ 0.4	29.3 $\pm$ 3.7	97.3 $\pm$ 0.2
Ours + VideoCrafter2	68.8 $\pm$ 1.8	27.7 $\pm$ 0.4	20.0 $\pm$ 3.3	97.7 $\pm$ 0.2
Turbo-V2	63.3 $\pm$ 1.7	28.6 $\pm$ 0.4	63.3 $\pm$ 3.9	96.2 $\pm$ 0.2
Ours + Turbo-V2	67.3 $\pm$ 2.1	27.4 $\pm$ 0.4	62.0 $\pm$ 4.0	96.8 $\pm$ 0.2