Diffusion and flow matching are now dominant frameworks for generative modeling in continuous domains, but they typically require multiple network evaluations at inference. This has motivated extensive research on one-step generative modeling primarily built on diffusion and flow, though existing approaches often rely on complex training machinery. In this work, we introduce Priming Models, a simple two-stage framework for training one-step generators. Motivated by the difficulty of distributional training from scratch, Priming Models are first trained on a reconstruction objective, e.g., pixel-space denoising at a fixed noise level, before undergoing feature-space distributional post-training. Our central scientific finding is that this simple reconstruction stage is already enough to entirely alleviate the difficulty of distributional training. Empirically, across class-conditional image generation, text-to-image generation, and world modeling, Priming Models significantly advance the Pareto frontier of training compute vs. generation quality. On ImageNet at 256×256 resolution, we achieve a state-of-the-art FID of 1.19 among one-step generators using approximately one-tenth the training compute of prior methods with comparable FID. We open-source our codebase at https://github.com/JaeyeonKim01/priming.
Figures & tables
Figure 1: Priming Models dramatically improve the training compute vs. FID trade-off on ImageNet ( left ). Simple reconstruction pretraining provides an effective initialization that makes subsequent distributional post-training substantially more efficient ( right ).
Figure 2: One-step samples from Priming Models across three visual domains.
Figure 3: Fixing style embedding limits diversity (ImageNet ‘otter’, ‘arctic fox’); more samples in Appendix F.4 .
Figure 4: Pretraining strengthens the Jacobian alignment, improving post-training efficiency.
Figure 5: Priming Models on FDr5 .
Method
FID ↓
FDr5↓
EFLOPs
pMF-L
2.55
10.85
821.75
pMF-H
2.15
8.14
1968.54
JiT-H
1.88
8.69
839.00
iMF-XL/2
1.74
9.75
3201.12
Drifting-B
1.74
15.30
1048.23
Priming-B (80 ep.)
1.47
12.71
85.75
Table 1: ImageNet . Priming Models improve the quality–compute trade-off. All methods are one-step except JiT-H (CFG scale 2.2 and 200 NFEs).
Method
GenEval ↑
DPG-Bench ↑
EFLOPs
Flow matching (100 steps) †
0.873
84.20
408.04
MeanFlow (4 steps) †
0.842
83.23
546.64
Priming (2 steps)
0.844
82.81
455.02
MeanFlow (1 step)
0.548
76.49
546.64
Priming ( s=0.8 , 200k)
0.740
81.30
293.53
Priming (logit-normal, 290k)
0.818
82.16
420.16
Table 2: Text-to-Image. Priming Models outperform one-step MeanFlow; the logit-normal variant approaches four-step MeanFlow with less training compute. Priming extended to two-step is competitive with four-step MeanFlow. The scores of † are from Wang et al. (2026a) .
Generation quality
Efficiency
Dataset
Method
NFE
SSIM ↑
PSNR ↑
LPIPS ↓
FID ↓
FVD ↓
Time/frame (ms) ↓
Train (EFLOPs) ↓
Flow Matching
32
0.85
25.428
0.076
7.8
27.24
322
12.37
RT-1
DriftWorld ( Lu et al., 2026a )
1
0.841
24.203
0.097
8.99
62.84
26
30.71
Priming
1
0.878
26.765
0.063
3.16
30.96
19
12.37
Flow Matching
32
0.841
23.37
0.079
11.02
39.87
445
9.55
Bridge-V2
DriftWorld ( Lu et al., 2026a )
1
0.835
22.691
0.091
4.86
46.05
30
37.05
Table 3: World modeling. Priming Model outperforms both the one-step DriftWorld baseline and a 32-step flow-matching baseline across all quality metrics while using less compute than DriftWorld.
Figure 6: Jacobian alignment in additional feature spaces. The 11 outputs complement the two shown in Figure 4 , covering all 13 analyzed ResNet-MAE and ConvNeXt outputs. Colors and markers match the main-text figure; horizontal axes are linear with feature-specific ranges; all panels share the same FID range.
Method
Precision ↑
Recall ↑
JiT-H †
0.78
0.62
Drifting-B
0.806
0.601
Priming-B (80 ep.)
0.786
0.617
Drifting-L
0.815
0.602
Priming-B (160 ep.)
0.783
0.624
WF-XL/2
0.762
0.674
Appendix
Table 8: Precision and recall on ImageNet. All methods use one-step generation except JiT-H. † JiT-H values are taken from Table 13 of Li and He (2026) .
Figure 7: Text-to-image qualitative comparison (1/6). One-step samples from Priming with logit-normal pretraining and one-step MeanFlow with CFG scale 12.0 .
Figure 8: Text-to-image qualitative comparison (2/6). One-step samples from Priming with logit-normal pretraining and one-step MeanFlow with CFG scale 12.0 .
Figure 9: Text-to-image qualitative comparison (3/6). One-step samples from Priming with logit-normal pretraining and one-step MeanFlow with CFG scale 12.0 .
Figure 10: Text-to-image qualitative comparison (4/6). One-step samples from Priming with logit-normal pretraining and one-step MeanFlow with CFG scale 12.0 .
Figure 11: Text-to-image qualitative comparison (5/6). One-step samples from Priming with logit-normal pretraining and one-step MeanFlow with CFG scale 12.0 .
Figure 12: Text-to-image qualitative comparison (6/6). One-step samples from Priming with logit-normal pretraining and one-step MeanFlow with CFG scale 12.0 .
Figure 13: Autoregressive world-model rollouts. Qualitative examples on RT-1 and Bridge-V2. For each rollout, the top row shows ground-truth frames and the bottom row shows predictions from our model at the same time points. The leftmost column is the final observed context frame; the remaining columns are uniformly spaced over the predicted rollout. Predictions are generated autoregressively while conditioning on the provided action sequence.
Figure 14: Fixing the style embedding limits sample diversity. For each row, we fix the ImageNet class and learned style embedding while varying only the Gaussian input across ten independent samples. For both Drifting Pixel-L and W-Flow latent XL/2, the samples retain highly similar appearance and composition across columns, indicating that the learned style embedding accounts for much of the variation that would otherwise be attributed to the Gaussian input.