Research map / Image generation / Fast diffusion sampling
Mean Flows for One-step Generative Modeling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MeanFlow, which models the average velocity over a time interval and trains a one-step generator via a MeanFlow identity that needs only instantaneous velocity and a JVP, reaching 3.43 FID with 1-NFE on ImageNet 256x256 from scratch.
Problem
One-step (1-NFE) generative modeling: instantaneous-velocity flow matching yields curved trajectories that are poorly suited to single-step integration, and existing one-step methods (consistency models, shortcut, IMM) rely on extra consistency constraints or curricula.
Approach
Model the average velocity u(z_t, r, t) = (1/(t-r)) ∫_r^t v dτ directly; differentiating the definition gives the MeanFlow identity u = v - (t-r) d/dt u, which yields a training target using only the conditional velocity and a JVP, so one-step sampling is z_0 = z_1 - u(z_1, 0, 1). CFG is folded into the ground-truth field so 1-NFE sampling is preserved.
Key results
- MeanFlow-XL/2 achieves 3.43 FID with 1-NFE on ImageNet 256x256 trained from scratch, vs. 7.77 for IMM and 10.60 for Shortcut-XL/2 at 1-NFE.
- MeanFlow-XL/2+ reaches 2.20 FID with 2-NFE, on par with DiT-XL/2 (2.27) and SiT-XL/2 (2.06 reported with 250x2 NFE).
- On unconditional CIFAR-10 with 1-NFE, MeanFlow reaches 2.92 FID, competitive with iCT (2.83) and IMM (3.20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the Flow Matching target with MeanFlow training that samples r≠t (25%) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 328.91 -> 61.06 | Helped |
| Use the correct JVP tangent (v, 0, 1) instead of (v, 0, 0) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 268.06 -> 61.06 | Helped |
| Use the correct JVP tangent (v, 0, 1) instead of (v, 1, 0) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 329.22 -> 61.06 | Helped |
| Use the correct JVP tangent (v, 0, 1) instead of (v, 1, 1) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 137.96 -> 61.06 | Helped |
| Condition on (t, t−r) rather than (t, r, t−r) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 63.98 -> 61.06 | Helped |
| Condition on (t, t−r) rather than t−r only (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 63.13 -> 61.06 | Helped |
| Use lognorm(--0.4, 1.0) rather than uniform(0, 1) time sampling (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 65.90 -> 61.06 | Helped |
| Use adaptive loss weighting p=1.0 instead of squared L2 loss p=0.0 (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 79.75 -> 61.06 | Helped |
| Use CFG scale 3.0 rather than no CFG (scale 1.0) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 61.06 -> 15.53 | Helped |
| Use CFG scale 5.0 rather than no CFG (scale 1.0) (ablation) | MeanFlow B/4 · ImageNet 256×256, 1-NFE | FID: 61.06 -> 20.75 | Helped |
| Increase model size from MeanFlow-B/2 to MeanFlow-XL/2 | MeanFlow · ImageNet 256×256, 1-NFE | FID: 6.17 -> 3.43 | Helped |
| Train MeanFlow-XL/2 longer with selected configurations | MeanFlow-XL/2 · ImageNet 256×256, 2-NFE | FID: 2.93 -> 2.20 | Helped |
Surprising findings
- /invoke
Limitations
- Class-conditional ImageNet results rely on CFG; the CIFAR-10 comparison is described as competitive rather than best.
- Ablations are run on a B/4 model for 80 epochs; scaling beyond XL/2+ is not explored in depth.
- The JVP adds about 16% wall-clock overhead, and the method's advantage over multi-step models is reported as on par rather than superior.
Experimental setup
Datasets: ImageNet CIFAR-10
Benchmarks: ImageNet 256x256 CIFAR-10
Metrics: FID-50K NFE
Baselines: iCT ECT sCT IMM Shortcut Models DiT-XL/2 SiT-XL/2 SiT-XL/2+REPA
Builds on: Flow Matching Consistency Models Shortcut Models Inductive Moment Matching DiT
Closest papers by meaning
- Understanding, Accelerating, and Improving MeanFlow Training (CVPR 2026 · similarity 0.94)
- Improved Mean Flows: On the Challenges of Fastforward Generative Models (CVPR 2026 · similarity 0.93)
- AlphaFlow: Understanding and Improving MeanFlow Models (ICLR 2026 · similarity 0.92)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.89)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.88)
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (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.