Summary
MV-Forcing is a paper by Gal Fiebelman, Hadar Averbuch-Elor, and Sagie Benaim (arXiv:2607.05376) addressing a long-standing gap in video generation: producing long, multi-view-consistent videos of dynamic scenes. Existing video diffusion models can generate long monocular videos via temporal autoregression or short multi-view videos via bidirectional attention, but not both. MV-Forcing combines temporal and view autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. The key insight is that autoregressive 3D reconstruction models naturally interface between generated views: given a completed source view, the framework reconstructs its 3D structure and renders a geometric prior for the next target view, which the diffusion model then refines into high-quality video. The model is distilled via distribution matching distillation and spatio-temporal self-forcing, closing the train-inference exposure bias of both temporal and view-wise sequential autoregression. Experiments show that MV-Forcing generates geometrically consistent multi-view videos of dynamic scenes of arbitrary length and number of views using a single-step student model.
Paper Overview
Field: Computer Vision
Authors: Gal Fiebelman, Hadar Averbuch-Elor, Sagie Benaim
arXiv: 2607.05376
Abstract
Recent advances in video diffusion models have enabled generating long single-view videos via temporal autoregression, or short multi-view videos via bidirectional attention. However, generating long, multi-view consistent videos of dynamic scenes remains an open problem.
This paper introduces MV-Forcing, which combines temporal and view autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. The key insight is that autoregressive 3D reconstruction models naturally interface between autoregressively generated views.
Method
- 4D-grounded bridging: Given a completed source view, the framework reconstructs its 3D structure and renders a geometric prior for the next target view.
- Refinement: The diffusion model refines this geometric prior into high-quality video.
- Distillation: The model is distilled via distribution matching distillation and spatio-temporal self-forcing, closing the train-inference exposure bias of both temporal and view-wise sequential autoregression.
Results
Experiments demonstrate that MV-Forcing generates geometrically consistent multi-view videos of dynamic scenes with arbitrary length and number of views, using a single-step student model.
---
*Source: arXiv preprint 2607.05376*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178346223