Flow-matching multi-view world models typically require paired videos from fixed camera rigs. SymRegFlow extends them to continuously varying camera poses without needing expensive novel-view ground truth. It warps source views to target poses using off-the-shelf 3D perception, then applies masked dual-anchor supervision and cross-anchor denoising consistency to overcome errors from geometric warping.
On nuScenes, SymRegFlow achieves the lowest FVD and FVMD among evaluated baselines, reducing FVD by over 31% relative to the best baseline. Source-conditioned inference also achieves best FID and instance preservation. A theoretical proof under an affine Gaussian surrogate shows that the regularization recovers the clean-reference optimum.