Research map / Probabilistic and scientific modeling / Variational and generative sampling
Carré du champ flow matching: better quality-generalisation tradeoff in generative models
Jacob Bamberger, Iolo Jones, Dennis Duncan, Michael M. Bronstein, Pierre Vandergheynst, Adam Gosztolai
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 isotropic FM noise with CDC geometric noise | Flow matching (MLP) · CITE-seq single-cell interpolation | Earth mover distance: I-FM 48.276 -> 46.657; OT-FM 45.393 -> 44.410 | Improved |
| Replace isotropic FM noise with CDC geometric noise | Flow matching (MLP) · Multiomics single-cell interpolation | Earth mover distance: I-FM 57.262 -> 54.419; OT-FM 54.814 -> 52.043 | Improved |
| Use CDC-FM in latent space on 1000 CelebA-HQ images | Latent flow matching · CelebA-HQ subset | FID: 13.10 -> 10.85 at epoch 5000 | Improved |
| Use CDC-FM in latent space on 1000 CelebA-HQ images | Latent flow matching · CelebA-HQ subset | NLL: 6.80 -> 7.18 at epoch 1000; 6.68 -> 6.48 at epoch 5000 | Mixed |
| Increase sigma_min of FM as naive regularisation (ablation) | Flow matching · Fruit fly motion capture | quality-generalisation frontier | No clear effect |
| Add CDC regularisation (gamma>0, strongest near 0.3) (ablation) | Flow matching (transformer) · Fruit fly motion capture | quality, generalisation, memorisation | Improved |
| Increase rank of G (d_cdc) (ablation) | CDC-FM · Fruit fly motion capture | generalisation / quality | Mixed |
| Add CDC-FM on CIFAR-10 small subsets (<10k) late in training | Flow matching (UNet) · CIFAR-10 subsets | memorisation, FID | Improved |
| Add CDC-FM with larger CIFAR-10 training sets | Flow matching (UNet) · CIFAR-10 large subsets | FID, generalisation | No clear effect |
| Add CDC-FM on high-dimensional tori at fixed data size | Flow matching · Torus T^d synthetic | sample quality | Hurt |
| Add CDC-FM on high-dimensional tori | Flow matching · Torus T^d synthetic | memorisation | Improved |
Closest papers by meaning
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 2025 · similarity 0.91)
- An Error Analysis of Flow Matching for Deep Generative Modeling (ICML 2025 · similarity 0.90)
- Categorical Flow Matching on Statistical Manifolds (NeurIPS 2024 · similarity 0.90)
- Improving Flow Matching by Aligning Flow Divergence (ICML 2025 · similarity 0.89)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.89)
- Metric Flow Matching for Smooth Interpolations on the Data Manifold (NeurIPS 2024 · similarity 0.89)
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.
