Title: Aligning Latents Equivariantly Through Flow Matching Within Group Symmetries

Abstract: While geometry-aware generative models and techniques for novel view synthesis have demonstrated impressive capabilities in maintaining visual fidelity and consistency, equivariant representation learning offers a complementary advantage. This framework facilitates the construction of latent spaces where known group transformations can operate directly, thereby encoding geometric structures within data and boosting both interpretability and generalization performance in novel view synthesis. Despite these benefits, we observe that current methods frequently encounter latent misalignment—a gap between the theoretical group action and the actual transformations necessary within the latent space. As a result, learned latent representations often struggle to reliably maintain the equivariant relationships dictated by the underlying group symmetry. To resolve this issue, we introduce Residual Latent Flow, a flow-based approach designed to rectify misaligned latents and enhance adherence to the fundamental equivariance relations. Extensive experimental results indicate that our proposed method substantially diminishes latent misalignment and elevates the quality of novel view synthesis, particularly under SO(n) rotation groups.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC