Research map / Probabilistic and scientific modeling / Variational and generative sampling
Flow Matching for Multimodal Distributions
Gaoxiang Luo, Frank Cole, Sihang Zhang, Yuxiang Wan, Yulong Lu, Ju Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Gaussian source with GMM source and mode coupling | DiTDH-XL on DINOv2-B latents · ImageNet 256 unconditional, 80 epochs, no guidance | FID: 9.33 -> 3.18 | Improved |
| replace Gaussian source with GMM source and mode coupling | DiTDH-XL · ImageNet 256 unconditional, 80 epochs, with autoguidance | FID: 5.82 -> 2.74 | Improved |
| GMM source without mode coupling (drop mode coupling) (ablation) | MM-FM · ImageNet 256 unconditional, 80 epochs, no guidance | FID: 3.82 -> 3.18 (adding mode conditioning) | Improved |
| use GMM source alone with independent coupling (ablation) | Flow matching, toy R10 · Toy R10 spread data | squared 2-Wasserstein distance: 1.582 -> 1.631 | Hurt |
| add mode coupling on top of GMM source (ablation) | Flow matching, toy R10 · Toy R10 spread data | average trajectory length: 2.30 -> 2.00 | Improved |
| increase number of GMM modes (ablation) | MM-FM without mode conditioning · ImageNet 256 unconditional | FID | Improved |
| use MM-FM sampling with fewer ODE steps | DiTDH-XL · ImageNet 256 unconditional | ODE steps for same FID: 25 -> 5 ODE steps | Improved |
| train on 10% of ImageNet with MM-FM | DiTDH-S · ImageNet 256, 10% data, 400 epochs | FID-50K: 24.65 -> 8.04 | Improved |
| vary GMM covariance type and assignment type (ablation) | MM-FM · ImageNet 256 unconditional | FID | No clear effect |
Closest papers by meaning
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.88)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.88)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.88)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.88)
- Categorical Flow Matching on Statistical Manifolds (NeurIPS 2024 · similarity 0.88)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.87)
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.
