MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators

Yushi Huang; Xiangxin Zhou; Jun Zhang; Liefeng Bo; Tianyu Pang

Anysampling stepsnative MeanFlow inference

6 / 8best image metricsamong few-step RL-tuned generators

+1.76VBench totalvs. LongCat Video RL · 50 steps

10×fewer evaluationsvs. 40-step DiffusionNFT

Selected generations

One policy, flexible test-time compute.

Explore the full comparison →

MeanFlowNFT

Text-to-Video4 steps

MeanFlowNFT

Text-to-Video16 steps

TL;DR

First forward-process RL for MeanFlow. A DiffusionNFT-style objective on an induced instantaneous-velocity predictor keeps training likelihood-free and leaves the few-step sampler unchanged.
Provable improvement. In an idealized setting the induced optimum matches DiffusionNFT's improved policy, and the gain provably carries over to the deployed average-velocity generator.
State-of-the-art few-step generation. Consistent gains on image (SD3.5-M) and video (Wan2.1 1.3B); with only 4 steps it surpasses multi-step RL and scales gracefully at test time.

Method

Optimize reward in instantaneous-velocity space, while preserving MeanFlow's native average-velocity parameterization and efficient few-step sampler.

Algorithm 1MeanFlowNFT (one update step)

Require: pretrained MeanFlow uθ, frozen reference uold (EMA of uθ), reward r, guidance β

// Sampling

Sample prompt c; roll out x0 with uold; evaluate r(x0,c) ∈ [0,1].
Sample an interval s ≤ t, and ε ∼ N(0,I).
xt ← αt x0 + σt ε; vt ← α̇t x0 + σ̇t ε.

// Training

x(t±Δt) ← xt ± Δt vt
d ← [uold(x(t+Δt),s,t+Δt) − uold(x(t−Δt),s,t−Δt)]/(2Δt)
Vθ ← uθ(xt,s,t) + (t−s)d
Vold ← uold(xt,s,t) + (t−s)d
Vθ+ ← (1−β)Vold + βVθVθ− ← (1+β)Vold − βVθ
LMFNFT ← r‖Vθ+−vt‖² + (1−r)‖Vθ−−vt‖²
Update θ by ∇θ LMFNFT.

Quantitative Results

Among few-step models, bold = best and underline = second-best. Our method is highlighted in green.

Text-to-Image SD3.5-Medium

512×512 training and 1024×1024 evaluation across eight metrics, using DrawBench and the official GenEval2 setting.

Text-to-Video Wan2.1 1.3B

480p, 81-frame generation; VBench and four reward metrics are measured on 256 held-out prompts.

Optimization CFG-free · 4-step rollouts

LoRA rank 32 with AdamW at 3×10⁻⁶; native MeanFlow sampling is retained at inference.

Table 1. Image generation on SD3.5-M at 1024×1024.

Method	ImageReward↑	CLIPScore↑	Aesthetic↑	PickScore↑	HPSv2↑	HPSv3↑	GenEval2↑	OCR↑
Multi-step models (40 steps)
SD3.5-M (w/o CFG)	-0.4770	0.2391	5.1514	20.6587	0.2109	3.4601	0.1171	0.1439
+ DiffusionNFT	1.4066	0.2889	5.9647	23.6440	0.3236	13.5959	0.2613	0.9098
SD3.5-M (w/ CFG)	0.9253	0.2880	5.3811	22.4638	0.2822	11.2489	0.2038	0.5449
+ Flow-GRPO	1.0193	0.2912	5.2843	22.4783	0.2657	9.8059	0.2638	0.6282
Few-step models (4 steps)
DMD	0.9241	0.2841	5.5055	22.2807	0.2874	11.7156	0.2042	0.3996
CDM	1.0307	0.2819	5.5721	22.4160	0.2976	12.5190	0.2020	0.3225
AnyFlow	1.1125	0.2886	5.4203	22.4789	0.2969	12.0691	0.1896	0.4520
RTDMD	1.2315	0.2775	6.1290	23.2825	0.3265	13.9253	0.2042	0.2965
R_dm	0.7236	0.2720	5.7537	22.0748	0.2759	11.2115	0.1619	0.3759
DMD + DiffusionNFT	0.7158	0.2843	5.3784	21.9571	0.2708	9.6985	0.2249	0.4865
CDM + DiffusionNFT	0.1455	0.2745	4.8335	21.3849	0.2143	-3.2834	0.2103	0.3303
AnyFlow + DiffusionNFT	1.2394	0.2915	5.9489	23.0876	0.2919	12.1378	0.2335	0.5948
MeanFlowNFT (Ours)	1.4504	0.2967	5.9275	23.5019	0.3269	13.8826	0.2375	0.6534

With only 4 steps, MeanFlowNFT matches or exceeds the 40-step forward-process RL method DiffusionNFT on several reward metrics at ~10× fewer function evaluations.

Table 2. Video generation on Wan2.1 1.3B. VBench scores and four metrics on 256 held-out prompts (HPSv3-G/-P: HPSv3 general/percentile; MQ/TA: VideoAlign motion-quality/text-alignment).

Method	VBench Total↑	Quality↑	Semantic↑	HPSv3-G↑	HPSv3-P↑	MQ↑	TA↑
Multi-step models (50 steps)
Wan2.1 1.3B (w/ CFG)	82.94	84.69	75.97	3.9099	8.2868	0.8684	1.2255
+ LongCat-Video RL	82.57	84.44	75.10	4.7099	9.2730	0.5493	1.6321
Few-step models (4 steps)
rCM	82.43	84.58	73.82	3.9660	8.7198	0.2740	1.6290
DMD	82.64	84.65	74.57	3.8598	8.7955	0.1810	1.6845
SC-DMD	83.36	84.76	77.77	–	–	–	–
AnyFlow	83.71	85.36	77.11	6.0536	10.450	0.7504	1.7356
MeanFlowNFT (Ours)	84.33	85.99	77.68	6.5959	10.793	0.9535	1.7235

Test-time scaling

Because u_θ approximates the average velocity over any interval, MeanFlowNFT supports any-step inference (N ∈ {2,4,8,16,32}). Most metrics improve with more steps, and MeanFlowNFT is markedly more step-consistent than AnyFlow.

Additional test-time scaling on SD3.5-M — **Figure 3.** Additional test-time-scaling metrics on **SD3.5-M**.

Test-time scaling on Wan2.1 1.3B — **Figure 4.** Test-time scaling on **Wan2.1 1.3B**: VBench, HPSv3 and VideoAlign metrics.

Training stability

**Figure 5.** Training reward curves on SD3.5-M. Directly plugging DiffusionNFT into AnyFlow / DMD / CDM trains unstably and collapses early, whereas **MeanFlowNFT** improves smoothly and monotonically.

Qualitative Comparison

Interactive side-by-side gallery. Text-to-Image uses SD3.5-M (1024×1024); Text-to-Video uses Wan2.1 1.3B (832×480, 81 frames). Browse each track with the ‹ › buttons, by swiping, or with the ←/→ keys; click any sample to enlarge.

Speed

BibTeX

@misc{huang2026meanflownftbringingforwardprocessrl,
  title   = {MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators},
  author  = {Huang, Yushi and Zhou, Xiangxin and Zhang, Jun and Bo, Liefeng and Pang, Tianyu},
  year    = {2026},
  eprint={2607.15273},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2607.15273},
}