Learned super-resolution of physical fields must satisfy structural constraints---such as divergence-free flow (∇⋅u=0), curl-free fields (∇×u=0), and pointwise sum-to-one normalization (1⊤p=1)---that standard training objectives do not enforce. Soft penalties trade fidelity against satisfaction via a scalar hyperparameter λ but cannot drive the residual to zero; hard-constrained regression eliminates the residual by construction yet converges to the conditional mean, losing small-scale spectral structure. We observe that in Conditional Flow Matching with linear paths, a linear constraint Ax=c is preserved along the entire transport trajectory whenever the velocity field is compatible (Avθ=0) and the initial noise is consistent (Ax0=c); along the trajectory, the latter two conditions are also necessary. We term this the path-equivariance principle and instantiate it in PACE (Path-Aligned Constraint Equivariance), a framework that enforces these conditions through a null-space parameterization of the velocity field and a constraint-respecting noise sampler, with no modification to the loss function or the ODE integrator. Across three constraint families, PACE reduces the mean constraint residual by one to five orders of magnitude relative to unconstrained baselines, matching the residual level of the data-generation pipeline itself. The proposed framework requires only a closed-form parameterization onto the null space of A; within this scope, it offers exact constraint enforcement while recovering the spectral content and predictive diversity that conditional-mean regression systematically loses, all under a single unified recipe.
Figures & tables
Constraint
C
Operator A
Target c
Null-space primitive
Div-free
2
∇⋅
0
stream function ψ
Curl-free
2
∇×
0
scalar potential ϕ
Sum-to-one
3
1⊤
1
logits ℓ∈RC
Table 1: Three constraint families. C : number of data channels; A : linear constraint operator; c : target value. The null-space primitive is predicted by the backbone Bθ and mapped through the parameterization Γ to obtain the velocity field vθ=Γ(Bθ)∈kerA .
Testbed
Metric
λ=0
λ=10−2
λ=10−1
λ=100
Hard
PACE
Div-free
PSNR
65.66
63.05
59.66
59.25
60.64
65.65
∣∇⋅u∣
2.27×10−2
1.46×10−3
3.86×10−4
2.03×10−4
7.34×10−6
6.39×10−6
Curl-free
PSNR
45.37
44.03
45.22
46.24
42.86
44.00
∣∇×u∣
1.51×10−1
7.22×10−2
4.17×10−2
9.09×10−3
7.77×10−6
8.25×10−6
Sum-to-one
PSNR
57.90
57.90
57.86
58.22
57.91
55.83
∣1⊤p−1∣
6.97×10−7
5.82×10−7
3.05×10−7
9.91×10−8
2.39×10−8
3.51×10−8
Table 2: Constraint satisfaction and reconstruction fidelity. PSNR (dB) and mean constraint residual ∣r(x)∣ across three testbeds ( s=4 ). λ=0 is the unconstrained regressor; λ∈{10−2,10−1,100} are soft-penalty variants. Data-generation residual levels: 6.00×10−6 (div-free), 7.87×10−6 (curl-free), 1.68×10−8 (sum-to-one).
Method
Div-free
Curl-free
Sum-to-one
Hard low- k
4.67×10−4
1.44×100
6.39×100
Hard high- k
4.52×10−7
5.26×10−6
7.77×10−5
PACE low- k
3.01×10−4
1.47×10−1
4.93×10−4
PACE high- k
7.27×10−8
5.13×10−6
4.18×10−6
Table 3: Spectral fidelity gap. ∣ΔE(k)∣ averaged in each band. Lower value indicates closer proximity to the true spectrum. Here, low- k refers to k<kc and high- k refers to k≥kc , with the band split at the FNO mode cut-off kc=32 .
Testbed
Method
PSNR
CRPS
Spectral W2
Curl-free
Hard
42.86
5.02×10−2
4.27
PACE
43.96
2.71×10−2
1.04
Div-free
Hard
60.65
1.04×10−3
3.22×10−3
PACE
63.18
4.35×10−4
3.18×10−3
Sum-to-one
Hard
57.83
9.61×10−4
7.22×10−4
PACE
57.79
7.11×10−4
1.47×10−3
Table 4: Distribution-level ensemble metrics. PACE ( N=10 ensemble) vs. Hard regression across three testbeds ( s=4 , FNO backbone). Lower CRPS indicates better probabilistic forecast than deterministic MAE.
Figure 1: Ensemble-size sensitivity on the curl-free testbed. CRPS and ensemble-mean PSNR as functions of ensemble size N ( s=4 , FNO backbone).
Scale factor
Div-free
Curl-free
Sum-to-one
s=4
6.39×10−6
8.25×10−6
3.51×10−8
s=8
6.39×10−6
8.21×10−6
3.52×10−8
Table 5: Scale ablation. FNO + PACE at s=4 vs s=8 on all three testbeds. Constraint residual remains effectively unchanged across scale factors.
Testbed
Model
PSNR
∣r(x)∣
Div-free
U-Net
60.47
6.39×10−6
DiT
63.89
6.39×10−6
FNO
65.65
6.39×10−6
Curl-free
U-Net
40.10
9.14×10−6
DiT
38.01
7.65×10−6
FNO
44.00
8.25×10−6
Table 6: Backbone ablation under PACE. Constraint residual stays at the data-generation residual level for every architecture on every testbed ( s=4 ).
Testbed
Method
PSNR
∣r(x)∣
Curl-free
Unconstrained FM
44.73
1.38×10−1
Velocity-Only
25.30
2.92×101
Noise-Only
24.49
1.12×101
Endpoint-Only
43.92
6.32×10−6
Per-Step
41.44
6.44×10−6
PBFM
44.43
2.14×10−1
Table 7: Component ablation. Reconstruction fidelity and constraint satisfaction across eight configurations on the curl-free and div-free testbeds.
Testbed
t
Endpoint-Only
Per-Step
PACE
Curl-free
0.0
2.91×101
2.91×101
5.48×10−6
0.2
2.26×101
3.42×10−6
5.45×10−6
0.4
1.62×101
3.45×10−6
5.79×10−6
0.6
1.30×101
4.10×10−6
6.06×10−6
0.8
6.56×100
5.16×10−6
6.78×10−6
1.0
2.63×10−1
6.43×10−6
7.75×10−6
Table 8: Mean constraint residual ∣r(xt)∣ along the ODE trajectory at integration times t∈{0.0,0.2,0.4,0.6,0.8,1.0} , where A=∇× for the curl-free testbed and A=∇⋅ for the divergence-free testbed. For Endpoint-Only, the value at t=1 is measured immediately before the final endpoint projection; its post-projection residual is reported in Table 7 .
Conditional Flow Matching trains generative models by regressing a network onto the velocity of a prescribed noise-to-data interpolation path. The interpolation schedule that shapes this path is known to affect convergence and sample quality, yet it is invariably fixed in advance, independent of both the data and the model. We show that the regression difficulty of Conditional Flow Matching varies systematically along the path, and we propose Difficulty-Calibrated Flow Matching, which derives the schedule from the model itself: a short pilot run with the linear path records the per-time loss, and the schedule is set to the quantile function of this difficulty profile, so the trajectory lingers where the velocity is hardest to learn. The method has a single hyperparameter, leaves the training objective and its gradient equivalence intact, composes with classifier-free guidance, and adds about two percent training overhead. In controlled experiments on CIFAR-10, MNIST, and Fashion-MNIST with an identical compact U-Net, the calibrated path attains the best FID on CIFAR-10 at full sampling budget and clearly outperforms all fixed schedules in the large-batch, few-update regime, precisely the setting where compute is scarcest.
Airin Akter Tania, Md Raihan Khan
Dept. of Electrical and Electronic Engineering Khulna University of Engineering & Technology Khulna, Bangladesh · Dept. of Electrical and Electronic Engineering North Western University Khulna, Bangladesh
Flow matching models excel at generative modeling, and many downstream applications require their samples to satisfy prescribed constraints, such as observed measurements and physical laws. However, existing constrained samplers often face a trade-off: enforcing constraints can substantially displace samples from the pretrained data distribution. To address this trade-off, we introduce MintFlow, a training-free constrained sampling framework that formulates constraint enforcement as a minimal intervention on the pretrained flow trajectory. MintFlow seeks the minimal perturbation of an intermediate flow state such that its subsequent evolution under the pretrained flow field satisfies the target constraint. By minimally perturbing the flow state while keeping the pretrained flow field unchanged, MintFlow enforces the constraint while minimizing unnecessary deviation from the pretrained distribution. An adjoint formulation yields a closed-form expression for this perturbation, eliminating expensive iterative optimization. Furthermore, MintFlow adaptively selects the intervention time to balance the required perturbation magnitude with its amplification by the remaining flow. Across a range of tasks in generative vision and physical system modeling, MintFlow achieves competitive constraint satisfaction while preserving the pretrained generative distribution substantially better than state-of-the-art constrained methods.
Yesom Park, Kelvin Kan, Qifan Chen +3
Department of Mathematics, University of California, Los Angeles · Brookhaven National Laboratory
We study path-flow alignment as a unified training objective for flow matching. Instead of fixing the interpolation path and learning only the velocity field, we jointly train an endpoint-preserving path network and a flow network using the same alignment loss: the flow learns to match the path velocity, and the path learns to align its velocity to the current flow. Although every fixed learned path defines a valid flow-matching objective, the alignment loss alone is not a reliable criterion for path learning. We identify path overfitting, a failure mode in which the alignment loss decreases while sample quality worsens. We find that this failure is associated with low-entropy bottlenecks in the induced probability path, where the learned path routes samples through overly concentrated intermediate marginals. Motivated by this diagnosis, we introduce a stochastic path regularizer that hides part of the source information from the path network while preserving exact endpoints. The resulting regularization gives an explicit entropy floor for the stochastic training-path marginals and empirically suppresses the bottleneck in the learned sampler, making joint path-flow training effective. On ImageNet-256x256 with SiT backbones, our method consistently improves FID across model scales, extends to model-guidance training, and leaves the inference-time architecture and sampler unchanged. Code is available at https://github.com/lizeyu090312/traj_opt_paper