Paper Overview
Research Area: Computer Vision Authors: Gal Fiebelman, Hadar Averbuch-Elor, Sagie Benaim Published: 2026-07-06 arXiv: 2607.05376
Summary
Recent advances in video diffusion models enable generating long single-view videos via temporal autoregression, or short multi-view synthesis via bidirectional attention. However, generating long-range, multi-view consistent videos of dynamic scenes remains an open problem.
This paper proposes MV-Forcing, which combines temporal and view autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views.
Key Idea
The central insight is that autoregressive 3D reconstruction models naturally serve as an interface between autoregressively generated views. Given a completed source view, the model reconstructs its 3D structure and renders geometric priors for the next target view; the diffusion model then refines these priors into high-quality video.
Method Highlights
- Spatio-temporal autoregression: Sequential generation over both time and viewpoints, grounded in 4D geometry.
- Distillation: Distribution matching distillation combined with spatio-temporal self-forcing closes the train-inference exposure bias inherent in sequential temporal and view autoregression.
- Efficiency: A single-step student model performs the generation.
Results
Experiments demonstrate that MV-Forcing generates geometrically consistent multi-view videos of dynamic scenes with arbitrary length and number of views.
---
*Auto-collected on 2026-07-06.*