Video World Models · 2026-10-05
SymRegFlow: Symmetry-Regularized Flow Matching for Video World Models
Xi Ye, Yuzhu Wang, Xiaoyang Liu, Jiayi Wang, Yangyang Xu, Ruyu Wang, Wenlin Chen, Duo Su, Jun Zhu
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
SymRegFlow is a symmetry-regularized flow-matching fine-tuning framework for cross-view-consistent video generation over continuously varying camera poses without ground-truth novel-view RGB supervision. It warps source views into noisy anchors and uses masked dual-anchor supervision plus cross-anchor denoising-output consistency, with directed cross-view attention to preserve observed-view fidelity.
Why it matters
The authors address the limitation of multi-view video world models that are restricted to fixed camera rigs, noting that continuous-view generation typically needs costly paired pose–video data with dense pose coverage. SymRegFlow enables continuous novel-pose generation using noisy geometry-derived anchors rather than ground-truth novel-view RGB, and reports improved video quality and motion consistency on nuScenes midpoint views.
Method
- For each target pose, SymRegFlow warps source views into geometry-derived noisy anchor videos and forms pseudo multi-view training examples without ground-truth novel-view RGB.
- It applies masked dual-anchor fine-tuning with cross-anchor denoising-output consistency over regions supported by either anchor, sharing sampled diffusion noise and noise level across branches.
- Directed Cross-View Attention (DCVA) restricts observed-view queries to observed-view tokens while allowing novel-view queries to attend to all tokens, and supports source-conditioned and joint-generated inference.
Limitation
The paper states that achieving continuous-view generation without ground-truth novel-view RGB supervision still requires geometry-derived anchor construction using an off-the-shelf 3D perception system; the resulting anchors are noisy and incomplete due to geometry errors, occlusions, and resampling.
Abstract (from arXiv)
Flow-matching-based multi-view world models generate realistic videos, but are commonly restricted to fixed camera rigs. Extending them to continuously varying camera poses requires paired pose--video observations with dense pose coverage, which are costly to acquire. We introduce \emph{SymRegFlow}, a symmetry-regularized flow-matching framework for multi-view-consistent video generation across continuous viewpoints without ground-truth novel-view RGB supervision. For each target pose, SymRegFlow geometrically warps source views into noisy anchors and combines masked dual-anchor supervision with cross-anchor denoising-output consistency to mitigate anchor-specific errors. Under an affine Gaussian surrogate, we prove that suitable consistency regularization recovers the clean-reference optimum at fixed noise levels, strictly outperforming single- and merged-anchor baselines. Experiments on Cosmos-Drive-Dreams and nuScenes demonstrate high-quality, multi-view-consistent autonomous-driving video generation: on nuScenes, SymRegFlow achieves the lowest FVD and FVMD among the evaluated baselines, reducing FVD by over 31\% relative to the best baseline, and source-conditioned inference also attains the best FID and instance preservation.
Related papers
- Frechet Forcing: Training Long-Horizon Streaming World Models from Short Clips
- World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
- World2Motion: Turning Video World Models into 3D Human Motion Generators
- CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion
- WorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-02 · Indexed by Pub-AI: 2026-10-05 UTC
SymRegFlow: Symmetry-Regularized Flow Matching for Video World Models
Preprint posted: 2026-10-02 · Original record ↗
- arxiv
- 2610.02726
Explicit source versions (1)
- v1 · 2026-10-02
- Preprint available
Source date: 2026-10-02 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Flow-matching-based multi-view world models generate realistic videos, but are commonly restricted to fixed camera rigs. Extending them to continuously varying camera poses requires paired pose--video observations with dense pose coverage, which are costly to acquire. We introduce \emph{SymRegFlow}, a symmetry-regularized flow-matching framework for multi-view-consistent video generation across continuous viewpoints without ground-truth novel-view RGB supervision. For each target pose, SymRegFlow geometrically warps source views into noisy anchors and combines masked dual-anchor supervision with cross-anchor denoising-output consistency to mitigate anchor-specific errors. Under an affine Gaussian surrogate, we prove that suitable consistency regularization recovers the clean-reference optimum at fixed noise levels, strictly outperforming single- and merged-anchor baselines. Experiments on Cosmos-Drive-Dreams and nuScenes demonstrate high-quality, multi-view-consistent autonomous-driving video generation: on nuScenes, SymRegFlow achieves the lowest FVD and FVMD among the evaluated baselines, reducing FVD by over 31\% relative to the best baseline, and source-conditioned inference also attains the best FID and instance preservation.
