Research map / Probabilistic and scientific modeling / Variational and generative sampling
Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models
Xinyue Ai, Yutong He, Albert Gu, Ruslan Salakhutdinov, J Zico Kolter, Nicholas Matthew Boffi, Max Simchowitz
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 |
|---|---|---|---|
| add F2D2 joint divergence distillation head | Shortcut-Distill · CIFAR-10 NLL (2 steps) | NLL (BPD): -57.72 -> 2.38 | Improved |
| add F2D2 | MeanFlow · CIFAR-10 FID (2 steps) | FID: 2.84 -> 2.59 | Improved |
| add F2D2 | MeanFlow · CIFAR-10 FID (1 step) | FID: 2.80 -> 3.02 | Hurt |
| add F2D2 | Shortcut-Distill · CIFAR-10 FID (8 steps) | FID: 5.01 -> 5.68 | Hurt |
| add F2D2 | Shortcut-Distill · ImageNet 64x64 FID (8 steps) | FID: 19.47 -> 21.91 | Hurt |
| add F2D2 | Shortcut-Distill · ImageNet 64x64 NLL (8 steps) | NLL (BPD): -9.03 -> 3.51 | Improved |
| add maximum likelihood self-guidance on 2-step MeanFlow-F2D2 | MeanFlow-F2D2 · CIFAR-10 | FID | Improved |
| distill with flow matching teacher velocity (Shortcut-Distill vs Shortcut) | Shortcut Model · CIFAR-10 FID (2 steps) | FID: 16.04 -> 7.13 | Improved |
| scale the divergence target by an appropriate factor (ablation) | F2D2 · CIFAR-10 | NLL and FID | Mixed |
| reweight divergence loss lambda=0.1 instead of 1 (ablation) | Shortcut-Distill-F2D2 · CIFAR-10 | FID and NLL | Mixed |
| add more Adam optimization steps in self-guidance (ablation) | MeanFlow-F2D2 · CIFAR-10 | FID | No clear effect |
Closest papers by meaning
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.90)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.90)
- How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 2025 · similarity 0.89)
- One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling (NeurIPS 2025 · similarity 0.89)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.89)
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (ICLR 2025 · similarity 0.88)
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.
