Research map / Image generation / Fast diffusion sampling
LeapAlign: Post-training Flow Matching Models at Any Generation Step by Building Two-Step Trajectories
Zhanhao Liang, Tao Yang, Jie Wu, Chengjian Feng, Liang Zheng
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 |
|---|---|---|---|
| use gradient discounting alpha=0.3 instead of removing the nested gradient (alpha=0) (ablation) | LeapAlign · HPDv2 test | HPSv2.1: 0.4064 -> 0.4092 | Improved |
| use gradient discounting alpha=0.3 instead of no discounting (alpha=1) (ablation) | LeapAlign · HPDv2 test | HPSv2.1 | Improved |
| use leap trajectory design even without nested gradient, versus DRTune | LeapAlign (alpha=0) · HPDv2 test | HPSv2.1: 0.3882 -> 0.4064 | Improved |
| add trajectory-similarity weighting at both x_j and x_0 (ablation) | LeapAlign · HPDv2 test | HPSv2.1 | Improved |
| use two-step instead of one-step leap trajectory (ablation) | LeapAlign · HPDv2 test | HPSv2.1 | Improved |
| use three-step instead of two-step leap trajectory (ablation) | LeapAlign · HPDv2 test | HPSv2.1 | No clear effect |
| train on timestep range [0,1] instead of [0,1/2] (ablation) | LeapAlign · GenEval | GenEval score | Improved |
| select k and j randomly instead of fixed distance 1/2 (ablation) | LeapAlign · HPDv2 test | HPSv2.1 | Improved |
| evaluate reward on generated x_0 instead of one-step prediction x_0|j (ablation) | LeapAlign · HPDv2 test | HPSv2.1 | Improved |
| raise loss threshold lambda from 0.55 to 0.95 (ablation) | LeapAlign · HPDv2 test | HPSv2.1: 0.4092 -> 0.4023 | Hurt |
| lower loss threshold lambda from 0.55 to 0.35 (ablation) | LeapAlign · HPDv2 test | HPSv2.1: 0.4092 -> 0.3860 | Hurt |
| replace DRTune with LeapAlign | Flux · GenEval | GenEval overall: 0.7101 -> 0.7420 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MixGRPO — MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE (ECCV 2026) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.94)
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.91)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.90)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.90)
- Optimal Flow Matching: Learning Straight Trajectories in Just One Step (NeurIPS 2024 · similarity 0.89)
- Structured Coupling for Flow Matching (NeurIPS 2026 · 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.
