Physics-informed generative models should produce samples that are both data-aligned and physically valid, yet existing methods are often iterative and too slow for practical use. Achieving this with a one-step flow-map model is difficult because both data-distribution and physics-constraint learning are enforced only indirectly through surrogate objectives, making their joint learning fragile for long-range flow maps. We propose Physics-Informed Long-Short Flow Maps with Direct Guidance (LSFM-DG), which makes direct target computation feasible by decomposing one-step generation into a long flow map for global evolution and a near-terminal short flow map with computable closed-form solutions and physics potentials. As the short flow map shrinks to zero, it provides both direct distributional and physical targets for learning the long flow map. This yields a unified one-step flow-map model that removes Jensen's gap, improves training stability, and achieves stronger data fidelity and physical consistency without distillation.
Figures & tables
Figure 1: PI-LSFM enables physics-informed one-step generation from scratch via long-short flow-map direct guidance, achieving multi-step-level performance on unconditional generation, conditional generation, and inverse design without increasing batch size or training steps, and without extra inference-time parameters or iterative correction.
Figure 2: Indirect Losses. Ldata does not directly supervise utθ with its true target ut(⋅) , but instead regresses to the conditional velocity ut(xt∣x1) . Likewise, Lphysics is not applied directly to utθ , but only to the terminal prediction x^1θ(xt) obtained after several integration steps. The indirectly supervised quantities are highlighted in muted blue.
Figure 3: Long-Short Flow Map. The flow map ψ0→1 is decomposed into a long flow map ψ0→1−Δt and a short flow map ψ1−Δt→1 . The short flow map is modeled through dynamics driven by the potentials Uphys and Udata , providing guidance for the long flow map. As Δt→0 , this becomes direct guidance for learning ψ0→1 , integrating both data and physical information.
Table 2: Three standard discretizations of the Physics potential-driven dynamics.
Figure 4: Comparison of design choices and overall performance in our framework. (a) Pareto front of physical residual MSE vs. Wasserstein distance for Darcy flow. (b)–(d) compare different feature extractors, numerical discretizations for the physics term and the data term respectively. Training time per step (s) is also reported in these comparisons. More details are provided in Appendix D . (e) Overall comparison with competing methods on RE, WD, JS, NFE, and inference time.
Dataset
Metric
Ours
PBFM
OT-FM
DiffusionPDE
D-Flow †
PCFM †
Kolmogorov Flow
RE ⋅101
0.404
1.362
2.314
1.930
-
-
WD ⋅101
0.140
1.222
2.124
3.698
-
-
JS ⋅102
1.621
7.440
12.49
19.39
-
-
MMSE ⋅104
25.83
4.455
4.188
5.669
-
-
SMSE ⋅104
1.765
2.484
2.574
21.82
-
-
NFE
1
20
20
20
20
20
Table 3: Generative performance metrics for Kolmogorov flow and dynamic stall problems. RE: physical Residual MSE, WD: Wasserstein Distance, JS: Jensen-Shannon divergence, MMSE: MSE of the mean fields, SMSE: MSE of the standard deviation fields, NFE: Number of function evaluations, IT: inference wall-clock time. Best and second-best results are highlighted.
Method
RE ↓
WD ⋅102↓
JS ⋅101↓
IT [s] ↓
PI-Consistency
0.240
3.889
5.023
0.025
PI-MeanFlow
4603.573
4.243
5.457
0.024
PI-Shortcut
0.435
2.336
2.518
0.024
PI-LSFM
0.055
0.108
0.171
0.012
Table 4: Additional one-step Darcy comparisons. All methods use NFE =1 ; IT is inference time.
Method
Seconds/step
PBFM
0.566
PI-Consistency
0.223
PI-MeanFlow
0.264
PI-Shortcut
0.328
PI-LSFM
0.308
Table 5: Training time per generator-training step on Darcy.
Method
NFE
IT [s]
Diffusion
100
1.720
PG-Diffusion
100
10.488
CoCoGen
100
7.911
TopoDiff-G
100
11.182
DOM
100
3.314
PIDM
100
1.701
Table 6: Inference timing on the same hardware.
Model
Size
in-distribution
out-of-distribution
RMAE↓
MDN % CE ↓
% VFE ↓
RMAE↓
MDN % CE ↓
% VFE ↓
Diffusion
136M
1.86e−3
-0.20
2.93
1.97e−3
0.30
2.80
PG-Diffusion ( Shu et al. (2023) )
136M
1.82e−3
0.09
3.59
1.92e−3
0.81
3.23
CoCoGen ( Jacobsen et al. (2025) )
136M
1.51e−3
0.14
4.00
1.56e−3
0.58
3.64
TopoDiff-G ( Mazé & Ahmed (2023) )
239M
–
0.83
1.49
–
1.82
1.80
DOM Giannone et al. (2023)
121M
–
0.74
1.52
–
3.47
1.59
Table 7: Topology optimization results under both in-distribution and out-of-distribution boundary-condition settings. PI-LSFM attains competitive performance against 100-step diffusion-based baselines while using only a single function evaluation.
Table 10: Three standard discretizations of the data potential-driven dynamics in feature space.
Figure 6: Comparison of constant step sizes and scheduling strategies for different physics-term discretizations. Each row corresponds to one numerical scheme, while columns show metric trends, RE–WD trade-offs, and RE–JS trade-offs. Due to the large dynamic range of the metrics for the Backward Euler methods, logarithmic scales are used in the scatter plots for better visualization.
Figure 7: Effect of the coefficient λˉ under different numerical discretizations.
Figure 8: Effect of the coefficient λˉ for different feature encoders. Each plot shows RE, WD, and JS as λˉ varies.
Figure 9: Toy example visualization comparing the training dataset, the fully projected target distribution, our method, and physics-informed sCM with different values of λˉ . Our method preserves the three-mode structure while satisfying the circle constraint, and achieves up to approximately 30× lower physical residual error than sCM, as reported in Table 11 .
Batch size
Ours
sCM ( λˉ=0 )
sCM ( λˉ=1 )
sCM ( λˉ=3 )
sCM ( λˉ=10 )
sCM ( λˉ=30 )
sCM ( λˉ=100 )
B=3
8.015
3996.970
220.986
176.003
190.425
773.622
240.384
B=4
3.040
574.162
198.419
155.516
146.720
142.761
98.074
B=8
0.722
366.265
158.029
142.076
68.378
38.710
33.033
Appendix
Table 11: Constraint satisfaction on the toy example under different batch sizes. We report RE: physical Residual MSE ×104 .
Figure 10: Generated Darcy flow samples. Each row shows one sample, with columns corresponding to pressure p , permeability K , and residual.
Figure 11: Failure cases without feature-space optimization on Darcy flow. Each row shows one generated sample, with columns corresponding to pressure p , permeability K , and residual. Without a feature encoder, optimization fails to produce physically meaningful results.
Hyperparameter
Value
Architecture hyperparameters
In-, output channels
2, 2
ResNet blocks per down-/upsampling pass
2
ResNet block normalization
Group Normalization, 8 groups
ResNet block activation function
SiLU
Attention block normalization
Layer Normalization
Appendix
Table 12: Architecture and training hyperparameters of Darcy Flow.
Figure 12: Generated Kolmogorov flow samples. Each row shows one sample, with columns corresponding to u -velocity, v -velocity, and divergence residual. From top to bottom, the Reynolds number increases.
Hyperparameter
Value
Architecture hyperparameters
In-, output channels
2, 2
Global conditioning dimension
1
ResNet blocks per down-/upsampling pass
2
ResNet block normalization
Group Normalization, 8 groups
ResNet block activation function
SiLU
Appendix
Table 13: Architecture and training hyperparameters of Kolmogorov flow.
Hyperparameter
Value
Architecture hyperparameters
In-, output channels
6, 6
Global conditioning dimension
4
ResNet blocks per down-/upsampling pass
2
ResNet block normalization
Group Normalization, 6 groups
ResNet block activation function
SiLU
Appendix
Table 14: Architecture and training hyperparameters of dynamic stall.
Figure 13: Generated dynamic-stall samples. Each row corresponds to one generated sample, and the columns show pressure P , temperature T , density ρ , wall shear stress τw , and tangential velocity gradients ∂xu and ∂yu , respectively.
Hyperparameter
Value
Architecture hyperparameters
In-, output channels
10, 3
ResNet blocks per down-/upsampling pass
2
ResNet block normalization
Group Normalization, 8 groups
ResNet block activation function
SiLU
Attention block normalization
Layer Normalization
Appendix
Table 15: Architecture and training hyperparameters of topology optimization.
Figure 14: Generated topology-optimization samples. Columns 1 and 3 show the generated density fields ρ , while columns 2 and 4 show the corresponding equilibrium residuals.
Figure 15: Kernel profiles for different values of t .
Method
CF
RE ↓
WD ⋅102↓
JS ⋅101↓
PI-Consistency
No
0.240
3.889
5.023
PI-Consistency
Yes
0.152
3.017
3.933
PI-MeanFlow
No
4603.573
4.243
5.457
PI-MeanFlow
Yes
4.973
4.253
5.212
PI-Shortcut
No
0.435
2.336
2.518
PI-Shortcut
Yes
0.276
2.184
2.363
Appendix
Table 16: One-step Darcy baselines with and without conflict-free (CF) updates.
Seed
RE ↓
WD ⋅102↓
JS ⋅101↓
43
0.132
0.269
0.241
123
0.116
0.215
0.210
2026
0.127
0.308
0.259
Mean ± std
0.125±0.008
0.264±0.047
0.237±0.025
Appendix
Table 17: Three-seed Darcy stability check at 100K steps.
Batch size
RE ↓
WD ⋅102↓
JS ⋅101↓
64 (default)
0.117
0.213
0.216
32
0.193
0.503
0.441
16
13.715
7.572
2.771
8
5.971
7.395
2.570
Appendix
Table 18: Darcy batch-size sensitivity after 100K training steps per setting.
Benchmark
λˉ
ϵ
Encoder
Data
Physics
Darcy
3.0
0.1
MAE
TR
FE
Kolmogorov
1.0
0.1
MAE
TR
FE
Dynamic stall
1.0
0.1
None
TR
FE
Topology optimization
0.01
0.1
None
TR
FE
Appendix
Table 19: Per-benchmark default settings used in the reported experiments.