MultiLive: On-Policy Distillation of Video and Multi-View Priors for Multi-View Video Generation

Yifan Niu^1,2,* · Yang-Tian Sun^1,3,*,† · Haoran Feng^1,4 · Lu Sheng² · Yan-Pei Cao^1,✉ · Zehuan Huang^1,✉,†

¹VAST   ²BUAA   ³HKU   ⁴THU
_{* Equal contribution · ✉ Corresponding authors · † Project leads}

Project Page

MultiLive generates synchronized multi-view videos from text and camera conditions by distilling video and multi-view priors into a single generator.

Highlights

• No multi-view video training data required. Uses separately learned video and multi-view priors.
• Dual-teacher on-policy distillation. Both teachers supervise the same student-generated sample for temporal and cross-view consistency.
• 4-step generation. Jointly generates synchronized multi-view videos in 4 denoising steps.

Code

Code coming soon.