Research map / Label-efficient and robust learning / Domain adaptation and generalization
DGFamba: Learning Flow Factorized State Space for Visual Domain Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DGFamba, a VMamba-based domain generalization method that randomizes state-embedding styles and aligns pre- and post-augmentation flow-factorized latent probability paths via a Hamilton-Jacobi constraint.
Problem
Visual domain generalization suffers from style-induced domain gaps; selective state space models (VMamba) have global receptive fields but are not explicitly style-invariant.
Approach
Hallucinate styles on VMamba state embeddings via AdaIN-style randomization, map original and stylized embeddings into a latent flow space via a VAE with factorized prior/posterior flows, and align them with a Hamilton-Jacobi physics-informed loss so the state representation becomes style-invariant.
Key results
- On PACS, DGFamba reaches 92.6% average accuracy, versus 91.2% for DGMamba (31M params).
- On VLCS, DGFamba averages 82.2%, versus 80.8% for DGMamba.
- On TerraIncognita, DGFamba averages 56.1%, versus 54.6% for DGMamba; OfficeHome average 78.5% vs 77.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DGMamba with DGFamba | VMamba-based domain generalization on PACS · PACS | average accuracy: 92.6 vs. 91.2; 1.4% | Helped |
| replace DGMamba with DGFamba | VMamba-based domain generalization on VLCS · VLCS | average accuracy: 82.2 vs. 80.8; 1.4% | Helped |
| replace DGMamba with DGFamba | VMamba-based domain generalization on OfficeHome · OfficeHome | average accuracy: 78.5 vs. 77.0 | Helped |
| replace DGMamba with DGFamba | VMamba-based domain generalization on TerraIncognita · TerraIncognita | average accuracy: 56.1 vs. 54.6 | Helped |
| add State Style Randomization (SSR) (ablation) | VMamba on PACS · PACS | average accuracy: 89.4 -> 90.4; 1.0% | Helped |
| add State Flow Encoding (SFE) after SSR (ablation) | VMamba on PACS · PACS | average accuracy: 90.4 -> 91.5; 1.1% | Helped |
| add State Flow Constraint (SFC) after SSR and SFE (ablation) | VMamba on PACS · PACS | average accuracy: 91.5 -> 92.6 | Helped |
| add the learning scheme to F1 (ablation) | VMamba on PACS · PACS | Art accuracy: 88.2 -> 89.7; 1.5% | Helped |
| add the learning scheme to F2 (ablation) | VMamba on PACS with F1 enabled · PACS | Art accuracy: 89.7 -> 91.0 | Helped |
| add the learning scheme to F3 (ablation) | VMamba on PACS with F1 and F2 enabled · PACS | Art accuracy: 91.0 -> 91.9 | Helped |
| add the learning scheme to F4 (ablation) | VMamba on PACS with F1, F2, and F3 enabled · PACS | Art accuracy: 91.9 -> 92.6; 0.7% | Helped |
| vary factorization steps T (ablation) | DGFamba on PACS · PACS | average accuracy: 91.6 (T=2); 92.1 (T=4); 92.4 (T=6); 92.6 (T=8); 92.4 (T=10); 92.0 (T=12) | Mixed |
Limitations
- Gains over DGMamba are modest (about 1.4-1.5% average on PACS, VLCS, OfficeHome) and the method adds a VAE and HJ-constrained MLPs.
- Theoretical justification of style invariance relies on flow-factorization formalism and is not empirically verified in depth.
- Hardware, training compute and code release are not reported.
Experimental setup
Datasets: PACS VLCS OfficeHome TerraIncognita ImageNet
Benchmarks: PACS VLCS OfficeHome TerraIncognita
Metrics: top-1 accuracy (%)
Baselines: GroupDRO VREx RSC MTL Mixstyle SagNet ARM SWAD PCL SAGM iDAG GMDG
Builds on: VMamba AdaIN DGMamba flow factorization physics-informed neural network
Closest papers by meaning
- START: A Generalized State Space Model with Saliency-Driven Token-Aware Transformation (NeurIPS 2024 · similarity 0.90)
- Learning Fine-grained Domain Generalization via Hyperbolic State Space Hallucination (AAAI 2025 · similarity 0.88)
- Disentangled Prompt Representation for Domain Generalization (CVPR 2024 · similarity 0.87)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.85)
- Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification (CVPR 2024 · similarity 0.85)
- Domain Generalization in CLIP via Learning with Diverse Text Prompts (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.