DGFamba: Learning Flow Factorized State Space for Visual Domain Generalization

TL;DR

This work tackles visual domain generalization under dramatic style variation by introducing DG-Famba, a flow factorized state space method that preserves the global receptive field of selective state space models while achieving style invariance. The approach combines state style randomization (SSR) to diversify styles, state flow encoding (SFE) to map state embeddings into a latent flow space, and state flow constraint (SFC) to align pre- and post-style embeddings using a Wasserstein-based objective and a physics-informed Hamilton-Jacobi constraint. Empirically, DG-Famba achieves state-of-the-art performance across PACS, VLCS, OfficeHome, and TerraIncognita, substantially improving generalization to unseen domains compared with CNN, ViT, and VMamba-based baselines, and ablations confirm the importance of each component. The method offers a principled, scalable path to robust cross-domain visual representations, with potential impact on real-world deployment where style shifts are common yet content remains stable.

Abstract

Domain generalization aims to learn a representation from the source domain, which can be generalized to arbitrary unseen target domains. A fundamental challenge for visual domain generalization is the domain gap caused by the dramatic style variation whereas the image content is stable. The realm of selective state space, exemplified by VMamba, demonstrates its global receptive field in representing the content. However, the way exploiting the domain-invariant property for selective state space is rarely explored. In this paper, we propose a novel Flow Factorized State Space model, dubbed as DG-Famba, for visual domain generalization. To maintain domain consistency, we innovatively map the style-augmented and the original state embeddings by flow factorization. In this latent flow space, each state embedding from a certain style is specified by a latent probability path. By aligning these probability paths in the latent space, the state embeddings are able to represent the same content distribution regardless of the style differences. Extensive experiments conducted on various visual domain generalization settings show its state-of-the-art performance.

Figures (5)

• Figure 1: General idea of the proposed DGFamba. The latent flow space aligns the probability path between the pre- and post- style augmented state embeddings to enhance the robustness to the style change.
• Figure 2: Performance on unseen domains (classification accuracy in %) v.s. trainable parameter number (in million, M). The proposed DGFamba outperforms the state-of-the-art methods.
• Figure 3: The proposed flow factorized state space model, dubbed as DGFamba, consists of three key components, namely, State Style Randomization, State Flow Encoding, and State Flow Constraint.
• Figure 4: t-SNE feature space visualization between the Mamba baseline and the proposed DGFamba. Experiments conducted on the PACS dataset. A more generalized representation allows features from different domains (in different color) to be uniformly distributed.
• Figure 5: Attention map on the unseen target domain from four experiments on PACS.