Research map / Probabilistic and scientific modeling / Molecules and proteins
Exploring and Exploiting Stability in Latent Flow Matching
Rania Briq, Michael Kamp, Ohad Fried, Sarel Cohen, Stefan Kesselheim
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 |
|---|---|---|---|
| prune training data with balanced clustering (Cb) at pr=0.5 | DiT-S/2 LFM · CelebA-HQ | FID: 24.24 -> 22.80 | Improved |
| select samples with highest loss (L) or lowest gradient (G-1) instead of the inverse (ablation) | LFM pruning · CelebA-HQ pr=0.5 | FID: 24.24 -> 33.92 (L), 29.75 (G-1) | Hurt |
| use coreset selection instead of balanced clustering (Cb) (ablation) | LFM pruning · CelebA-HQ pr=0.5 | FID: 22.80 -> 26.94 (balanced coreset), 26.25 (global coreset) | Hurt |
| select cluster samples furthest from center instead of nearest (ablation) | Cb pruning · CelebA-HQ pr=0.5 | FID: 22.80 -> 26.77 | Hurt |
| use balanced instead of proportional clustering (ablation) | clustering pruning · CelebA-HQ pr=0.5 | FID: 25.19 -> 22.80 | Improved |
| random pruning compared with unpruned (ablation) | LFM training · CelebA-HQ pr=0.5 | FID: 24.24 -> 25.25±0.38 | Hurt |
| scale model from DiT-S/2 to DiT-XL/2 (20x weights) (ablation) | LFM · CelebA-HQ | FID: 24.24 -> 22.87 | Improved |
| switch backbone from DiT to U-Net (ablation) | LFM · CelebA-HQ | ArcFace similarity s: s = 0.55 | Hurt |
| replace flow matching with score-based diffusion (ablation) | LFM · CelebA-HQ | output stability | Hurt |
| change the VAE (training seed) or apply invertible sign flip to latent space (ablation) | LFM · CelebA-HQ | ArcFace similarity s: s = 0.32 (VAE swap), 0.48 (flip first feature), 0.32 (flip all) | Hurt |
| add coarse-to-fine two-model inference (t0 = 0.7) | DiT-XL/2 LFM · CelebA-HQ, H100, batch 128, 256x256 | inference cost (ms/img): 93.95 -> 43.53 ms/img (≈2.15x faster) | Improved |
| train Coarse on a single perceived gender (stability violated) instead of balanced pruned data (ablation) | coarse-to-fine model · CelebA-HQ | FID: FID = 44.92 | Hurt |
Closest papers by meaning
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.92)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.90)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.90)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.90)
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (ICLR 2025 · similarity 0.89)
- Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · 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.
