MV-Forcing: Long Multi-View Video Generation via 4D-Grounded Spatio-Temporal Self-Forcing

MV-Forcing, from researchers at the Hebrew University of Jerusalem and Cornell University, tackles a gap that existing video diffusion systems such as SynCamMaster and Self-Forcing leave open: generating long videos that stay consistent across multiple camera viewpoints. The key idea is to combine temporal autoregression with view-sequential autoregression, while using CUT3R as a 4D geometric bridge that reconstructs a persistent scene state from previously generated frames and renders a viewpoint-specific geometric prior for the next view. The student model is distilled from a bidirectional SynCamMaster teacher with Distribution Matching Distillation, then trained with spatio-temporal self-forcing and a joint denoising regime so the first view can be generated from text and later views can be generated autoregressively without a fixed temporal window. On the SynCamVideo benchmark, which contains 3,400 synthetic scenes captured from 10 synchronized cameras, MV-Forcing reaches strong long-horizon performance at 3 views and 162 frames, improving camera accuracy to 3.72 rotation error and 8.41 translation error, while raising cross-view synchronization to 243.63 matching pixels and 89.88 CLIP-V in the synthetic setting. Scaling is notably stable: performance remains close to constant from 2 to 5 views, and across 81 to 648 frames the method preserves camera metrics with only mild temporal degradation. Runtime is also practical, with a 4-step causal student running in about 70 seconds for 2 views at 81 frames and using 23 GB peak VRAM, roughly a 5× speedup over the 50-step SynCamMaster baseline.