Research map / Image generation / Fast diffusion sampling
Trajectory Consistency for One-Step Generation on Euler Mean Flows
Zhiqi Li, Yuchen Sun, Duowen Chen, Jinjin He, Bo Zhu
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 |
|---|---|---|---|
| remove JVP computation (JVP-free training) | EMF vs MeanFlow · CelebA-HQ and ImageNet training cost | peak memory and time per iteration: ImageNet memory 101.9GB -> 57.9GB; 400.9ms -> 198.8ms | Improved |
| add local-linearization surrogate to the JVP-free objective (ablation) | EMF · CelebA-HQ | FID: 13.0 -> 10.9 | Improved |
| replace JVP-based objective with the JVP-free surrogate (ablation) | EMF · CelebA-HQ | FID: 12.4 -> 10.9 | Improved |
| use fixed delta t between 6e-4 and 1e-2 (ablation) | EMF · CelebA-HQ | FID: 10.9 at 6e-4 to 10.7 at 1e-2 | No clear effect |
| use very small or very large delta t (1e-4, 1e-1) (ablation) | EMF · CelebA-HQ | FID: 10.9 -> 25.8 (1e-4); 10.9 -> 25.2 (1e-1) | Hurt |
| add x1-prediction instead of u-prediction (ablation) | EMF in pixel space · CelebA-HQ-256 pixel space (JiT-B/16), 1-step | FID-50k: 324.6 -> 36.3 | Improved |
| use x1-prediction with u-space loss (x1/u) rather than x1/x1 (ablation) | EMF in pixel space · CelebA-HQ-256 pixel space, 1-step | FID-50k: 36.3 -> 30.6 | Improved |
| use EMF instead of MeanFlow objective | DiT-B/2 on ImageNet-256 · ImageNet-256, 1-step | FID-50k: 11.1 -> 7.2 | Improved |
| use EMF instead of MeanFlow objective | DiT-B/2 on CelebA-HQ · CelebA-HQ-256, 1-step | FID-50k: 12.4 -> 10.8 | Improved |
| use EMF instead of MeanFlow at 128 steps on CelebA-HQ | DiT-B/2 · CelebA-HQ-256, 128-step | FID-50k: 7.8 -> 7.6 (Flow Matching 7.3 is better) | No clear effect |
| add auxiliary branch to predict u(t->t) | EMF · CelebA-HQ | peak memory / FID: memory 23.3GB -> 17.6GB; FID 10.9 -> 11.7 | Mixed |
| use EMF instead of MeanFlow | SDF generation · ShapeNet SDF from 64 points, 1-step | Chamfer distance: 0.060 -> 0.046 | Improved |
Closest papers by meaning
- AlphaFlow: Understanding and Improving MeanFlow Models (ICLR 2026 · similarity 0.90)
- DuoFlow: JVP-Free Finite-Difference Mean Flows for One-Step Image Generation (ECCV 2026 · similarity 0.89)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.89)
- SCoT: Unifying Consistency Models and Rectified Flows via Straight-Consistent Trajectories (NeurIPS 2025 · similarity 0.89)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.88)
- How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 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.
