Flow-based image generative models exhibit stable training and produce high quality samples when using multi-step sampling procedures. One-step generative models can produce high quality image samples but can be difficult to optimize as they often exhibit unstable training dynamics. Meanflow models exhibit excellent few-step sampling performance and tantalizing one-step sampling performance. Notably, MeanFlow models that achieve this have required extremely large training budgets. We significantly decrease the amount of computation and data budget it takes to train Meanflow models by noting and exploiting a particular discretization of the Meanflow objective that yields a consistency property which we formulate into a ``Discrete Meanflow'' (DMF) Training Curriculum. Initialized with a pretrained Flow Model, DMF curriculum reaches one-step FID 3.36 on CIFAR-10 in only 2000 epochs. We anticipate that faster training curriculums of Meanflow models, specifically those fine-tuned from existing Flow Models, drives efficient training methods of future one-step examples.
Figures & tables
Figure 1: Training convergence on unconditional CIFAR-10. DMF curriculums achieve better 1-step FID compared to the MF baseline with equal training data budget, despite starting from a pretrained flow model at 2000 epochs.
Method
Initialization
Budget (Epochs)
FID ( ↓ )
From Scratch
iCT
Random
8k
2.83
MF (ours imple.)
Random
4k
3.85
MF Geng et al. (2025a)
Random
16k
2.90
With DM / FM Initialization
sCT Lu and Song (2025)
Pretrained DM
4k + 4k
2.85
Table 1: CIFAR-10 comprehensive 1-step FID comparison . Methods are categorized by initialization strategy. Budgets represent total training epochs, for models initialized with pretrained models, the budget are formatted as “Cost of Pretraining” + “Cost of Tuning”. DMF curriculum training attains competitive or better FID compared to prior methods.
Table 3
Method
Training Epochs
Rel. Budget
Steps
FID ↓
SiT-XL/2 (Baseline)
1400 (Pretrain)
100.0%
50
11.52
DMF †
1400 + 6
+0.42%
1
21.18
DMF †
1400 + 12
+0.85%
1
18.03
DMF †
1400 + 24
+1.71%
1
16.95
DMF †
1400 + 48
+3.42%
1
14.53
DMF †
1400 + 96
+3.42%
1
294.13
Table 4: ImageNet 256×256 curriculum training budget w.r.t FID. We report the FID of our 1-step DMF curriculum relative to the 1400-epoch pretrained SiT-XL/2 baseline. The FID is computed on samples generated w/o CFG sampling or tuning.
Strategy
Objective Ratio (FM:MF)
Training Type
CFG
Standard MF Geng et al. (2024)
1:1 to 3:1
Direct Training
Yes
RAE-MF Hu et al. (2025b)
3:1
Mid-training Teacher
None
α -Flow Zhang et al. (2025)
1:1
Direct Curriculum
Yes
DMF Curriculum
0:1 (Pure MF)
Direct Curriculum
None
Table 5: Comparison of Training Paradigms on ImageNet-256. Unlike majority MF that are trained on hybrid objectives or distillation teachers, our DMF curriculum enables “pure” MF fine-tuning.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
FM Pretrain
MF (Scratch)
MF (Fine-tune)
DMF † curriculum
Batch Size
256
1024
1024
1024
Training Epochs
2000
4000
2000
2000
Optimizer
Adam
Adam
Adam
Adam
Learning Rate
2×10−4
6×10−4
6×10−4
6×10−4
EMA Rate
0.999
0.999
0.999
0.999
Network Dropout
0.2
0.2
0.2
0.35
Appendix
Table 6: CIFAR-10, hyperparameter configurations across different experimental setups.
Hyperparameter
DMF † 6ep
DMF † 12ep
DMF † 24ep
DMF † 48ep
DMF † 96ep
Batch Size
1024
1024
1024
1024
1024
Training Epochs
6
12
24
48
96
Optimizer
AdamW
AdamW
AdamW
AdamW
AdamW
Learning Rate
1×10−4
1×10−4
1×10−4
1×10−4
1×10−4
EMA Rate
0.995
0.999
0.999
0.9995
0.9995
Network Dropout
0.0
0.0
0.0
0.0
0.0
Appendix
Table 7: ImageNet 256×256 hyperparameter configurations across different experimental setups.
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.