Research map / Probabilistic and scientific modeling / Variational and generative sampling
Posterior Augmented Flow Matching
George Stoica, Sayak Paul, Matthew Wallingford, Abhay Nori, Vivek Ramanujan, Winson Han, Ali Farhadi, Ranjay Krishna, Judy Hoffman
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 |
|---|---|---|---|
| Train with PAFM (posterior-weighted multi-target supervision) instead of FM | MMDiT · CC12M | FID50K: 10.37 -> 9.45 | Improved |
| Use PAFM with nearest-neighbor targets instead of FM | SiT-B/2 and SiT-XL/2 · ImageNet-1K 256 | FID, IS, sFID, precision, recall: up to 3.4 FID50K | Improved |
| Vary number of neighbors K for PAFM (ablation) | SiT-B/2 · ImageNet-1K 256 | FID: best at K=16 | Mixed |
| Use random-crop augmentation targets (aug scale 1.25) in PAFM (ablation) | SiT-B/2 (K=5) · ImageNet-1K 256 | FID: -3.42 FID at scale 1.25 | Improved |
| Sample K times from VAE moment distribution as targets (ablation) | SiT-B/2 · ImageNet-1K 256 | FID | Mixed |
| Use PAFM without REPA loss (ablation) | SiT-B/2 · ImageNet-1K 256 (no REPA) | FID, sFID, precision, recall | Improved |
| Use PAFM instead of FM for gradient variance | REPA-SiT-B/2 (K=16) · ImageNet-1K 256 | mean batch-gradient variance: 0.80 -> 0.22 | Improved |
| Use PAFM instead of FM with nearest neighbors precomputed | MMDiT · CC12M training cost | throughput: -6.6% throughput, +0.4% peak memory | Hurt |
Closest papers by meaning
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.90)
- Variational Rectified Flow Matching (ICML 2025 · similarity 0.90)
- Coreset-Induced Conditional Velocity Flow Matching (NeurIPS 2026 · similarity 0.90)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.90)
- A Theory on Flow Matching with Neural Networks (NeurIPS 2026 · similarity 0.90)
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 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.
