betaFeedback

WorksComputer Science

Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems

Abstract

Reconstructing high-fidelity flow fields from low-fidelity observations is a central problem in scientific machine learning, yet recent diffusion and flow-matching models typically rely on iterative sampling, making them costly for latency-sensitive workflows such as ensemble forecasting, real-time visualization, and simulation-in-the-loop inference. We study whether a high-fidelity flow-matching generative model can be compressed into a compact one-step model for fast scientific flow reconstruction. Our approach distills an optimal-transport flow-matching teacher into a one-step consistency model. Low-fidelity observations are incorporated at inference by initializing the generative trajectory from a noised observation along the transport path, allowing an unconditional high-fidelity flow model to perform conditional reconstruction without retraining the teacher. We evaluate this distillation strategy on three fluid benchmarks, Smoke Buoyancy, Turbulent Channel Flow, and Kolmogorov Flow, using coarse-to-fine reconstruction as a controlled testbed at field sizes up to 256×256256 \times 256. Across these settings, the distilled student retains similar performance of the teacher's model on spectrum metrics, while using roughly half as many parameters and achieving a 12×12\times inference speedup over the flow-matching teacher. Under the same training budget, the distilled student also outperforms a one-step consistency model trained directly from scratch by 23.1%23.1\% in SSIM, showing that teacher distillation improves training efficiency rather than merely accelerating sampling. These results suggest a promising route for turning future high-capacity scientific generative models into compact reconstruction models that are faster to train, cheaper to run, and easier to deploy.

Introduction

The high-fidelity simulation of physical systems such as turbulent flows, atmospheric dynamics, and combustion processes is central to modern scientific computing. These systems are governed by nonlinear partial differential equations whose solutions exhibit rich multi-scale structure, with small-scale features playing a critical role in energy transfer, dissipation, and long-term system evolution [1]. Direct Numerical Simulation (DNS) [2] resolves these features faithfully but at computational costs that are prohibitive for many practical workflows such as ensemble forecasting, parameter sweeps, real-time monitoring, and uncertainty quantification routinely require thousands or millions of forward evaluations, far beyond what DNS can support. Approximation methods such as Reynolds-Averaged Navier-Stokes (RANS) [3] and Large Eddy Simulation (LES) [4] reduce cost by filtering or modelling fine-scale dynamics, but at the expense of the very small-scale fidelity that often matters most.

A practical alternative is physical fidelity enhancement, also referred to as super-resolution or statistical downscaling: given a coarse, cheaply-computed low-resolution field, recover a high-resolution counterpart consistent with the underlying physics. The mapping is fundamentally ill-posed as multiple physically plausible high-resolution states are consistent with any low-resolution observation, which makes generative modelling a natural fit [5]. Diffusion-based approaches [6, 7, 8] and, more recently, Flow Matching (FM) [9, 10] have produced strong results on this class of problem [11, 12], yielding samples that respect the energy spectra of the target dynamics and capture multi-modal posteriors over fine-scale detail [13, 14, 15, 16, 17].

Despite their fidelity, these models are intrinsically multi-step at inference: producing a single high-resolution field requires several Neural Function Evaluations (NFEs) along an iterative denoising or integration trajectory. The cost compounds when many samples are needed, which is precisely the regime in which generative super-resolution would be most useful: large-ensemble forecasting, simulation-in-the-loop pipelines, real-time visualisation. This inference cost has become the binding constraint on deploying generative super-resolution in scientific workflows, and it cannot be circumvented by simply scaling hardware because the latency of a multi-step trajectory is irreducible at the algorithmic level.

Consistency models [18, 19] and their simplified variant (sCM) [20] address this bottleneck by learning a self-consistent mapping that collapses an entire generative trajectory into a single forward pass. Recent work by Chen et al. [21] demonstrates that a pretrained FM teacher can be distilled into a one-step sCM student via TrigFlow parameterisation and Jacobian-vector-product based tangent estimation, achieving teacher-level quality at a fraction of the inference cost. This framework has been validated extensively in the natural-image domain. Whether it transfers to scientific fields with their power-law spectra, conservation structure, and multi-scale coupling has not been demonstrated.

In this work we present the first such demonstration. We distil an FM teacher into a TrigFlow consistency student for physical fidelity enhancement of two-dimensional fluid systems, with both teacher and student trained fully unconditionally on the high-resolution distribution. The low-resolution observation is incorporated only at inference, by initializing the trajectory from a noised version of the coarse field at an intermediate point along the OT path using the FM path insertion mechanism [22, 23, 24] that carries through cleanly to the one-step consistency student.

Contributions.

  • We present the first demonstration of one-step consistency distillation from a flow-matching teacher for physical fidelity enhancement of fluid dynamical systems, adapting the sCM/TrigFlow framework [20, 21] to scientific super-resolution.

  • Across three fluid benchmarks, our 15M-parameter student matches a substantial fraction of the 30M-parameter teacher’s improvement over low-resolution baselines on both pointwise and spectral metrics, while reducing inference to a single network evaluation and achieving a 12×12 \times wall-clock speedup over the teacher.

  • We provide a systematic benchmark across resolutions (128×128128 \times128 to 256×256256 \times256) and flow regimes (Smoke Buoyancy, Turbulent Channel Flow, 2D Kolmogorov Flow), establishing reference results for one-step generative super-resolution in fluid dynamics.

Related Works

Diffusion models for physical fidelity reconstruction. Diffusion-based generative models have long been adapted for physically faithful reconstruction [25, 15, 13]. They have also demonstrated superior performance over deterministic reconstruction algorithms [15], largely because their distribution-matching nature enables better preservation of physical metrics [15, 13, 14]. Shu et al. [15] extensively studied the application of denoising diffusion probabilistic models (DDPMs) [6] and denoising diffusion implicit models (DDIMs) [26], showing that diffusion models consistently outperform deterministic baselines on super-resolution and sparse reconstruction tasks. Flow matching, as an alternative formulation of diffusion models [9, 10], has also been widely adopted for generative vision reconstruction and editing tasks [27, 28, 29, 30], including applications in Earth systems such as weather forecasting [30, 28] and fluid dynamics [25]. However, these models are intrinsically multi-step and incur substantial inference costs as the number of generated frames increases, limiting their applicability in real-world scenarios [14]. Although one-step generative models, such as consistency models [18] and shortcut diffusion [31], have been introduced to address this limitation, their application to physical systems remains largely unexplored.

Overview of the proposed one-step consistency distillation pipeline

Figure 1. Overview. Left: Our pipeline distills a multi-step OT-FM teacher (5-step RK5, 30 NFEs) into a one-step TrigFlow consistency student via sCM distillation. Both models are trained unconditionally; the low-resolution observation is incorporated at inference by initializing the trajectory at an intermediate time τ\tau along the OT path.

Diffusion distillation. To mitigate the inference cost of standard DDPM and flow matching models, distillation methods have been proposed to reduce the number of inference steps required by multi-step generative models. Diffusion distillation methods such as Distribution Matching Distillation (DMD) [32, 33] and Consistency Trajectory Models (CTM) [34] involve the joint training of student and adversarial networks, which can impose substantial training overhead. A simple yet effective alternative is consistency distillation [18, 20, 35], which distils FM-style models by enforcing consistency along generation trajectories. In this study, we focus on this formulation to achieve efficient and effective distillation.

Methodology

Figure 1 summarizes our system. We train an unconditional OT-FM teacher on the high-resolution distribution (§ 3.2), distill it into a one-step TrigFlow consistency student via sCD (§ 3.3), and incorporate the low-resolution observation purely at inference by initializing the trajectory along the OT path, reducing the full pipeline to a single forward pass through the student (§ 3.4).

Problem Setup

We address physical fidelity enhancement of two-dimensional fluid fields. Let xHR∈RH×W×Cx_{\mathrm{HR}} \in\mathbb{R}^{H \times W \times C} denote a high-resolution snapshot of a fluid field with CC channels (e.g. velocity components, density), and let xLR∈Rh×w×Cx_{\mathrm{LR}} \in\mathbb{R}^{h \times w \times C} denote a corresponding low-resolution observation obtained by stride-ss subsampling of xHRx_{\mathrm{HR}}, with H=shH = sh and W=swW = sw. The goal is to recover a high-resolution sample x^HR\hat{x}_{\mathrm{HR}} consistent with the underlying physics, conditioned on xLRx_{\mathrm{LR}}.

The mapping is fundamentally one-to-many: many physically plausible high-resolution fields are consistent with any given low-resolution observation, since the subsampling discards information at scales smaller than ss. We therefore model p(xHR∣xLR)p(\mathbf{x}_{\mathrm{HR}} \mid\mathbf{x}_{\mathrm{LR}}) as a conditional distribution and draw samples from it via a generative trajectory.

Throughout, we represent the low-resolution input on the high-resolution grid via nearest-neighbour upsampling, denoted x^LR∈RH×W×C\hat{\mathbf{x}}_{\mathrm{LR}} \in\mathbb{R}^{H \times W \times C}, so that all flow trajectories operate in the same ambient space.

Flow Matching

Flow Matching [9, 10] learns a continuous-time velocity field vϕ(z,t)\mathbf{v}_{\phi}(\mathbf{z}, t) that transports a Gaussian base distribution into the data distribution pdatap_{\mathrm{data}} along a time-indexed density path {pt}t∈[0,1]\{p_t\}_{t \in[0,1]}. The path is specified through a conditional coupling zt=atx+btϵ\mathbf{z}_t = a_t \mathbf{x} + b_t \boldsymbol{\epsilon} with x∼pdata\mathbf{x} \sim p_{\mathrm{data}}, ϵ∼N(0,I)\boldsymbol{\epsilon} \sim\mathcal{N}(\mathbf{0}, \mathbf{I}), and vϕ\mathbf{v}_{\phi} is trained to regress the corresponding conditional velocity dzt/dt\mathrm{d}\mathbf{z}_t/\mathrm{d}t via the conditional flow-matching loss [9]. We adopt the linear (optimal-transport) schedule at=1−ta_t = 1 - t, bt=tb_t = t, which yields straight-line trajectories

zt=(1−t)x+tϵ,t∈[0,1],(1)\mathbf{z}_t = (1-t)\mathbf{x} + t\boldsymbol{\epsilon}, \qquad t \in[0,1], \tag*{(1)}

and a constant conditional velocity ϵ−x\boldsymbol{\epsilon} - \mathbf{x} [9]. Our teacher is trained under this OT parameterization. The full FM objective, its conditional reformulation (CFM), and the OT-path derivation are reproduced in Appendix § 6.§ 1.

Consistency Distillation

Consistency models. A consistency model [18, 19] replaces iterative integration with a single network fθ(zt,t)f_{\theta}(\mathbf{z}_t, t) trained to map any point along a generative trajectory directly to that trajectory’s endpoint, enforcing the self-consistency property

fθ(zt,t)=fθ(zt′,t′)for all (t,t′) along the same trajectory.(2)f_{\theta}(\mathbf{z}_t, t) = f_{\theta}(\mathbf{z}_{t'}, t') \quad\text{for all } (t,t') \text{ along the same trajectory}. \tag*{(2)}

Once trained, generation reduces to a single forward pass, regardless of where on the trajectory the input lies.

Continuous-time consistency distillation. We adopt the simplified continuous-time consistency distillation (sCD) framework of Lu and Song [20], which differentiates the self-consistency identity directly rather than approximating it with finite differences over discrete trajectory pairs. Concretely, the student’s velocity is regressed onto a target constructed from an EMA copy of the student and the tangent of the consistency function along the trajectory. The tangent term is computed exactly using a Jacobian–vector product (JVP), requiring only a single additional forward-mode pass, with the trajectory direction provided by the teacher. We provide the full sCD loss and JVP-based tangent estimator in Appendix § 6.§ 3, and the per-step distillation algorithm in Appendix § 6.§ 4. The sCD framework is derived under a trigonometric coupling, referred to as TrigFlow. Following the lossless OT–TrigFlow equivalence of Chen et al. [21], we query our pretrained OT-FM teacher at TrigFlow timesteps, enabling it to provide consistent supervision to the student without retraining. The required transformations are given in Appendix § 6.§ 2.

Conditioning and Inference

Both the FM teacher and the sCM student are trained unconditionally on the high-resolution distribution p(xHR)p(\mathbf{x}_{\mathrm{HR}}). The low-resolution observation is incorporated only at inference by exploiting the structure of the OT path: we treat the upsampled LR field x^LR\hat{\mathbf{x}}_{\mathrm{LR}} as an approximation to a high-resolution sample that lies on the OT trajectory at some intermediate time τ∈(0,1)\tau\in(0,1), and initialize the inference trajectory at that point,

zτ=(1−τ)x^LR+τϵ,ϵ∼N(0,I).(3)\mathbf{z}_{\tau} = (1-\tau)\hat{\mathbf{x}}_{\mathrm{LR}} + \tau\boldsymbol{\epsilon}, \qquad\boldsymbol{\epsilon} \sim\mathcal{N}(\mathbf{0}, \mathbf{I}). \tag*{(3)}

This is precisely the affine interpolation of Eq. (1) that the velocity field was trained on, so the initialization lies on-manifold for vϕ\mathbf{v}_{\phi} at time τ\tau; the construction is the OT-path flow-matching analog of SDEdit [22]. The hyperparameter τ\tau trades off LR fidelity against high-resolution realism, small τ\tau stays close to the LR input but adds little fine-scale structure, large τ\tau erases the LR information, and within a wide plateau the choice is largely cosmetic (Appendix § 9.§ 1).

At test time the FM teacher integrates the unconditional velocity field from t=τt = \tau to t=0t = 0 with a fifth-order Runge–Kutta solver (5 NFEs). The sCM student needs no integration: it maps any on-trajectory point directly to the endpoint in a single forward pass, so x^HR=fθ(zτ,τTrig)\hat{x}_{\mathrm{HR}} = f_\theta(z_\tau, \tau_{\mathrm{Trig}}) with τ\tau expressed in TrigFlow time. The student is operated at a shallower noise injection (τ≈0.3\tau\approx0.3) than the multi-step teacher (τ≈0.6\tau\approx0.6), reflecting its reduced capacity to refine deeply noised inputs in a single step. The full inference pipeline thus reduces to one nearest-neighbour upsample, one noise sample, and one network evaluation.

Experiments

We evaluate the proposed method on three two-dimensional fluid benchmarks spanning different resolutions and flow regimes. Our central question is whether a one-step consistency student distilled from an OT-FM teacher can recover comparable physical fidelity at substantially lower inference cost. All experiments follow the unconditional training setup in § 3; task conditioning is introduced only at inference through trajectory initialization at a dataset-specific noising time τ\tau.

Benchmarks We consider three datasets: Smoke Buoyancy (32→\rightarrow128; buoyancy-driven transport; 173,824 snapshots), Turbulent Channel Flow (TCF) (64→\rightarrow192; anisotropic near-wall turbulence; 38,080 snapshots), and Kolmogorov Flow (64→\rightarrow256; two-dimensional periodic incompressible turbulence; 12,800 snapshots). Appendix E provides full dataset statistics, simulation provenance, and preprocessing details. Together, these benchmarks test whether the method generalizes across distinct fluid regimes and super-resolution scales rather than being tuned to a single setting.

Baselines We compare against four baselines. The LR baseline is the nearest-neighbour upsampled low-resolution input and serves as a no-model reference. The FM teacher is the multi-step OT-FM model used for distillation, evaluated by default with a 5-step RK5 ODE solver. The diffusion baseline is a separately trained DDPM-style model, providing a standard multi-step generative reference. Finally, sCM trained from scratch uses the same student architecture and consistency objective as our distilled model but receives no teacher supervision, isolating the effect of distillation.

Metrics We evaluate fidelity using both pointwise and spectral metrics. Pointwise reconstruction is measured by relative L2L_2 error (RL2) and Structural Similarity Index Measure (SSIM) [36] against the ground-truth high-resolution field. Spectral fidelity is measured by Power Spectral Density Discrepancy (PSDD), the radially averaged deviation between predicted and ground-truth power spectra. Because pointwise metrics can favor overly smooth reconstructions, PSDD is essential for assessing whether a model preserves the multi-scale energy structure of fluid fields. Inference cost is reported as wall-clock time per frame on a single NVIDIA GPU, averaged over 100 forward passes after 10 warm-up passes. All metrics are computed on the held-out 10% validation split.

Across all benchmarks, the FM teacher’s Euler, Heun, and RK5 solvers produce near-identical reconstruction metrics, indicating that solver order is not the limiting factor. We therefore focus the per-dataset discussion on the 5-step RK5 teacher.

Smoke Buoyancy

Table 1 reports reconstruction metrics on the Smoke Buoyancy dataset (32→\rightarrow128). The 30-step diffusion baselines, DDIM and DPM++, use a substantially smaller backbone (3.7 M parameters, approximately 9×\times smaller than the FM teacher) yet achieve slightly stronger pointwise reconstruction scores than the FM teacher itself. This provides useful context for the wall-clock comparison in Table 4. Despite using only a single function evaluation, the distilled sCM model (15.9 M parameters, NFE=1=1) outperforms the RK5 FM teacher on all reported metrics, improving RL2 by 5.9%, SSIM by 5.9%, and PSDD by 35.9%. In contrast, the from-scratch sCM ablation, which uses the same student architecture and training budget but receives no teacher supervision, performs substantially worse on both pixel metrics and spectral metrics. This gap isolates teacher distillation, rather than the consistency objective alone, as the key factor enabling accurate one-step generation on this task.

ModelNFERL2↓RL_2 \downarrowSSIM ↑\uparrowPSDD ↓\downarrow
Diffusion (multi-step)
DDIM300.1450.6904.35×10−24.35 \times10^{-2}
DPM++300.1560.6154.38×10−24.38 \times10^{-2}
Flow Matching teacher (multi-step)
Euler50.1620.6576.61×10−26.61 \times10^{-2}
Heun100.1680.6146.60×10−26.60 \times10^{-2}
RK5300.1680.6286.60×10−26.60 \times10^{-2}
Consistency Models (1-step)
sCM (trained)10.190.541.75×10−11.75 \times10^{-1}
sCM (distilled)10.1580.6654.23×10−24.23 \times10^{-2}
Δ\Delta vs. RK5 (%)−5.9-5.9+5.9+5.9−35.9-35.9

Table 1. Smoke Buoyancy (32 →\to 128) super-resolution metrics. The sCM-distilled model achieves comparable reconstruction quality to the flow matching teacher with roughly half the parameters and a single function evaluation. Diffusion rows compare DDIM and DPM-Solver++ (DPM++) at 30 reverse steps; flow-matching rows compare Euler, Heun, and the 5-step RK5 used by the teacher in the rest of the paper.

SolverDatasetParametersNFEInference Time (s)
Diffusion Models
DDIMSmoke Buoyancy (1282128^{2})3.7 M300.0311 ± 0.0001
Turbulent Channel Flow (1922192^{2})3.7 M300.0719 ± 0.0001
Kolmogorov Flow (2562256^{2})3.5 M300.1319 ± 0.0002
DPM++Smoke Buoyancy (1282128^{2})3.7 M300.0314 ± 0.0001
Turbulent Channel Flow (1922192^{2})3.7 M300.0724 ± 0.0001
Kolmogorov Flow (2562256^{2})3.5 M300.1327 ± 0.0002
Teacher Models — Flow Matching
EulerSmoke Buoyancy (1282128^{2})32.4 M50.0133 ± 0.0024
Turbulent Channel Flow (1922192^{2})29.5 M50.0272 ± 0.0049
Kolmogorov Flow (2562256^{2})29.5 M50.0462 ± 0.0060
HeunSmoke Buoyancy (1282128^{2})32.4 M100.0236 ± 0.0032
Turbulent Channel Flow (1922192^{2})29.5 M100.0483 ± 0.0053
Kolmogorov Flow (2562256^{2})29.5 M100.0836 ± 0.0066
RK5Smoke Buoyancy (1282128^{2})32.4 M300.0667 ± 0.0028
Turbulent Channel Flow (1922192^{2})29.5 M300.1382 ± 0.0061
Kolmogorov Flow (2562256^{2})29.5 M300.2415 ± 0.0006
Consistency Model — 1-step (Ours)
sCMSmoke Buoyancy (1282128^{2})15.9 M10.0056 ± 0.0021
Turbulent Channel Flow (1922192^{2})14.9 M10.0116 ± 0.0045
Kolmogorov Flow (2562256^{2})14.9 M10.0197 ± 0.0090

Table 4. Inference cost per frame across solvers and datasets. The distilled sCM reduces inference to a single function evaluation, yielding a substantial wall-clock speed-up over multi-step teachers at roughly half the parameter count, independent of solver, dataset, or output resolution. NFE counts model evaluations per frame: for the FM teacher, NFE = (outer integration steps) × (stages per step), giving 5×1, 5×2, and 5×6 for Euler, Heun, and RK5 respectively. All times measured at the same per-dataset batch size (Smoke Buoyancy 32, turbulent channel flow 16, Kolmogorov Flow 8).

Figure 2 confirms the metric numbers visually: across eight i.i.d. samples spanning a range of plume morphologies, the one-step student is nearly indistinguishable from the ground truth at this resolution, and its absolute-error map closely tracks the teacher’s both in spatial localisation (concentrated along

Smoke Buoyancy super-resolution samples

Figure 2. Smoke Buoyancy super-resolution samples (32 →\to 128). Six independently super-resolved frames (Samples A–F) drawn from across the dataset; columns are i.i.d. reconstructions of distinct frames. Row 1: nearest-neighbour-upsampled low-resolution input. Row 2: high-resolution ground truth. Row 3: one-step sCM student reconstructions, which preserve the buoyant plume’s vortical roll-ups, edge sheets, and fine-scale fingering. Rows 4–5: per-pixel absolute residual against ground truth for the sCM student and the five-step OT-FM teacher, plotted on a shared error colorbar. The student’s residual is comparable to and, on several samples, lower than the teacher’s despite using a single network evaluation versus thirty.

the plume’s steep density gradients) and in magnitude. The headline message is that the distillation does not collapse fine-scale structure: vortical features at the diffusive front are recovered in a single forward pass.

Turbulent Channel Flow

Table 2 reports results on Turbulent Channel Flow. Among FM teacher solvers, Euler gives the best pointwise scores (RL2=0.0056RL_2 = 0.0056, SSIM = 0.998), marginally exceeding RK5; at N=16N=16 frames, this difference is small and likely reflects the smooth bulk velocity profile that dominates the 1922192^2 field. DPM-Solver++ substantially improves over DDIM (RL2RL_2 0.019 vs. 0.036; PSDD 7.4×10−47.4 \times10^{-4} vs. 2.7×10−32.7 \times10^{-3}), suggesting that the higher-order solver better recovers near-wall gradients. The distilled sCM remains close to the FM teacher in SSIM (1.6% gap) but incurs larger errors in RL2RL_2 (∼2.4×\sim2.4 \times) and PSDD (about one order of magnitude), making TCF the only benchmark where one-step distillation carries a notable spectral cost. This gap is likely amplified by the dataset’s narrow dynamic range and by the teacher’s exceptionally low PSDD (3.58×10−43.58 \times10^{-4}), which sets a high spectral baseline; the student’s PSDD (3.51×10−33.51 \times10^{-3}) is comparable to its values on the other datasets. Qualitative trajectories in Appendix § 8 (Figure 4) show that structural content is largely preserved, while residuals reveal the teacher’s lower pointwise error.

ModelNFERL2↓RL_2 \downarrowSSIM ↑\uparrowPSDD ↓\downarrow
Diffusion (multi-step)
DDIM300.03550.96112.72×10−32.72 \times10^{-3}
DPM++300.01920.98327.36×10−47.36 \times10^{-4}
Flow Matching teacher (multi-step)
Euler50.00560.99842.94×10−42.94 \times10^{-4}
Heun100.01020.97091.98×10−41.98 \times10^{-4}
RK5300.00710.99463.58×10−43.58 \times10^{-4}
Consistency Models (1-step)
sCM (trained)10.04790.77814.06×10−24.06 \times10^{-2}
sCM (distilled)10.01730.97883.51×10−33.51 \times10^{-3}
Δ\Delta vs. RK5 (%)+145.7−1.6+880

Table 2. Turbulent Channel Flow (64 →\to 192) super-resolution metrics. The sCM-distilled model preserves the reconstruction fidelity of the flow matching teacher while operating at NFE = 1 with approximately half the parameter count. Diffusion rows compare DDIM and DPM-Solver++ at 30 reverse steps; flow-matching rows compare Euler, Heun, and the 5-step RK5 used by the teacher in the rest of the paper.

Turbulent Channel Flow super-resolved samples

Figure 4. Turbulent Channel Flow super-resolved samples (64→19264 \to192). Six independently super-resolved frames (Samples A–F) drawn from across the dataset; columns are i.i.d. reconstructions. Layout matches Figure 2. The streamwise-velocity field is dominated by the bulk channel profile, so visual differences between the nearest-neighbour-upsampled input, ground truth, and sCM reconstruction are subtle. The error rows make the residual gap explicit: the multi-step OT-FM teacher achieves roughly a third of the sCM student’s pointwise error here, consistent with the metric ratios in Table 2. The student residual is concentrated in the near-wall streaks where the small-scale fluctuations live; the bulk profile is recovered to within machine-precision-scale error in both models.

Kolmogorov Flow

Table 3 reports results on Kolmogorov Flow, the highest-resolution and most challenging benchmark in our suite (2562256^2). The distilled sCM outperforms all multi-step baselines: relative to the RK5 FM teacher, it reduces RL2RL_2 by 6.8%, increases SSIM by 5.1%, and lowers PSDD by 59.3%. The from-scratch sCM ablation achieves similar pointwise accuracy (RL2=0.282RL_2 = 0.282, SSIM = 0.723) but much worse spectral fidelity (PSDD 5.27×10−25.27 \times10^{-2} vs. 1.34×10−21.34 \times10^{-2}), again indicating that teacher distillation is critical for recovering small-scale structure. That a one-step student exceeds a 30-NFE teacher on this benchmark suggests that single-shot distillation can avoid integration error accumulation in highly turbulent fields, while inference-time OT-path noise may act as an implicit regularizer.

ModelNFERL2↓RL_2 \downarrowSSIM↑SSIM \uparrowPSDD↓PSDD \downarrow
Diffusion (multi-step)
DDIM300.2910.6884.75×10−24.75 \times10^{-2}
DPM++300.2930.6814.93×10−24.93 \times10^{-2}
Flow Matching teacher (multi-step)
Euler50.3040.6893.26×10−23.26 \times10^{-2}
Heun100.3090.6783.29×10−23.29 \times10^{-2}
RK5300.3080.6823.29×10−23.29 \times10^{-2}
Consistency Models (1-step)
sCM (trained)10.2820.7235.27×10−25.27 \times10^{-2}
sCM (distilled)10.2870.7171.34×10−21.34 \times10^{-2}
Δ\Delta vs. RK5 (%)−6.8-6.8+5.1+5.1−59.3-59.3

Table 3. Kolmogorov Flow (64→25664 \to256) super-resolution metrics. The sCM-distilled model matches the flow matching teacher across all metrics while requiring a single network evaluation and roughly half the parameters. On this benchmark the distilled student in fact improves over the FM teacher on every metric, with a −59.3%-59.3\% reduction in spectral error (PSDD). Diffusion rows compare DDIM and DPM-Solver++ at 30 reverse steps; flow-matching rows compare Euler, Heun, and the 5-step RK5 used by the teacher in the rest of the paper.

Figure 3 provides two diagnostics for the Kolmogorov student. Panel (a) shows that validation RL2RL_2 decreases monotonically with student capacity and saturates above roughly 12 M parameters, where the student matches the teacher on most validation frames. Panel (b) reports the radially averaged power spectral density. The nearest-neighbour LR baseline shows the expected dip near the input Nyquist frequency (k≈50k \approx50) and spurious high-frequency energy, whereas both the FM teacher and the 14.89 M one-step sCM closely track the ground truth across the inertial range. The student retains a small high-wavenumber residual, consistent with partial loss of fine-scale detail when collapsing multi-step integration into a single neural evaluation. Spectra for the remaining benchmarks are provided in Appendix § 8.

Figure 3. Capacity scaling and spectral fidelity on Kolmogorov Flow (64→25664 \to256). (a) Validation RL2RL_2 for four sCM students of varying capacity, all distilled from a common FM teacher (mean ±\pm SEM, 32 frames). (b) Radially-averaged power spectral density (mean, same 32 frames) of the ground truth, nearest-neighbour LR baseline, FM teacher (5 NFEs), and our 14.89 M sCM student (1 NFE).

Table 4 consolidates per-frame inference time across solvers and datasets. Three trends emerge. First, the FM teacher cost scales approximately linearly with the number of model evaluations per integration step: Heun is ∼1.8×\sim1.8\times slower than Euler, while RK5 is ∼5×\sim5\times slower, indicating that forward passes dominate runtime at the chosen batch sizes. Second, the distilled sCM delivers a consistent ∼12×\sim12\times wall-clock speed-up over the RK5 FM teacher across all resolutions (1282128^2, 1922192^2, and 2562256^2), with speed-up factors of 11.9×11.9\times, 11.9×11.9\times, and 12.3×12.3\times, respectively. This near-constant ratio reflects the reduction from 30 to 1 NFE together with the student’s smaller backbone. Third, the diffusion baselines remain 5.55.5–6.7×6.7\times slower than sCM despite using a ∼4×\sim4\times smaller backbone, because their 30-step sampling budget is unchanged; DPM-Solver++ improves the trajectory discretization but not its length. Overall, consistency distillation removes the dominant inference-time cost while preserving reconstruction quality, with the largest practical gains expected at high spatial resolutions where multi-step samplers are most expensive. Qualitative super-resolution results are shown in Appendix § 8 (Figure 5).

Kolmogorov Flow super-resolved samples

Figure 5. Kolmogorov Flow super-resolved samples (64→25664 \to256). Six independently super-resolved frames (Samples A–F) drawn from across the dataset; columns are i.i.d. reconstructions. Layout matches Figure 2. The vorticity field exhibits dense multi-scale structure with prominent same-sign vortex patches and steep shear sheets between them. The one-step sCM student recovers the location, sign, and amplitude of the vortices, and the residual rows show that its absolute error is comparable to and, on every sample shown here, modestly lower than the multi-step OT-FM teacher’s, consistent with the metric ratios in Table 3.

Conclusion

We introduced one-step consistency distillation from a flow-matching teacher for fidelity enhancement in two-dimensional fluid systems. Conditioning is supplied only at inference through OT-path trajectory initialization, allowing both teacher and student to remain unconditional models of the high-resolution distribution.

Across the Smoke Buoyancy, Turbulent Channel Flow, and Kolmogorov Flow benchmarks, the distilled student closely matches the multi-step OT-FM teacher on pointwise and spectral metrics while using roughly half the parameters and only one neural function evaluation. This yields an approximately 12×12 \times wall-clock speed-up at comparable reconstruction quality. Under a matched training budget, the distilled student also improves SSIM over a from-scratch sCM by 23.1% on Smoke Buoyancy, suggesting that the teacher trajectory provides an effective training curriculum rather than merely accelerating sampling.

Several limitations remain. The fidelity–realism trade-off is controlled by a single dataset-specific noising time τ\tau; adaptive or content-dependent schedules may further reduce the residual spectral gap. Our diffusion baselines use smaller backbones than the FM teacher and student, leaving parameter-matched comparisons to future work. Finally, our experiments are limited to two-dimensional, statistically stationary fields, and the conditioning mechanism assumes that the upsampled low-resolution input is well approximated by an on-trajectory OT state. Extending this framework to three-dimensional turbulence, non-stationary boundary conditions, sparse or unstructured observations, and other scientific domains such as weather and combustion is a natural direction for future work.

More broadly, our results suggest a transferable recipe: as larger flow-matching models for scientific data become available, consistency distillation can convert them into compact, deployable reconstruction models without retraining the teacher or adding task-specific conditioning machinery.

Background and Algorithms

This appendix consolidates the technical material underlying the flow-matching teacher and the TrigFlow consistency student: the full FM objective and its conditional reformulation (Appendix § 6.§ 1), the TrigFlow parameterization and its lossless equivalence with the OT-FM path (Appendix § 6.§ 2), the simplified continuous-time consistency distillation loss with its JVP-based tangent estimator (Appendix § 6.§ 3), and the per-step training and distillation algorithms (Appendix § 6.§ 4).

Flow Matching

Flow Matching constructs a continuous-time flow transforming a simple base distribution pbasep_{\mathrm{base}} (e.g. a Gaussian) into a complex target data distribution pdatap_{\mathrm{data}} [9]. The flow is characterized by a time-indexed family of densities {pt(zt)}t∈[0,1]\{p_t(z_t)\}_{t \in[0,1]} together with a velocity field vt(z)\mathbf{v}_t(\mathbf{z}) whose dynamics generate the path.

Conditional paths. A widely used construction specifies the path through a conditional coupling. Given a data sample x∼pdata\mathbf{x} \sim p_{\mathrm{data}} and noise ϵ∼pbase\boldsymbol{\epsilon} \sim p_{\mathrm{base}}, one defines a time-dependent interpolation

zt=atx+btϵ,(4)\mathbf{z}_t = a_t \mathbf{x} + b_t \boldsymbol{\epsilon}, \tag*{(4)}

where at,bt:[0,1]→Ra_t, b_t : [0,1] \to\mathbb{R} are differentiable scalar schedules satisfying a0=b1=1a_0 = b_1 = 1 and a1=b0=0a_1 = b_0 = 0. Differentiating Eq. (4) yields the corresponding conditional velocity

v(zt∣x,ϵ)=dztdt=at′x+bt′ϵ.(5)\mathbf{v}(\mathbf{z}_t \mid\mathbf{x}, \boldsymbol{\epsilon}) = \frac{\mathrm{d}\mathbf{z}_t}{\mathrm{d}t} = a'_t \mathbf{x} + b'_t \boldsymbol{\epsilon}. \tag*{(5)}

Flow matching objective. Let vϕ(z,t)\mathbf{v}_{\phi}(\mathbf{z}, t) denote the parametrized teacher velocity model. Ideally, one would match vϕ\mathbf{v}_{\phi} to the marginal velocity field that generates the density path {pt}\{p_t\},

LFM=Et,zt∼pt∥vϕ(zt,t)−Ept(x,ϵ∣zt)[v(zt∣x,ϵ)]∥22.(6)\mathcal{L}_{\mathrm{FM}} = \mathbb{E}_{t, \mathbf{z}_t \sim p_t} \left\| \mathbf{v}_{\phi}(\mathbf{z}_t,t) - \mathbb{E}_{p_t(\mathbf{x},\boldsymbol{\epsilon}\mid\mathbf{z}_t)}\left[\mathbf{v}(\mathbf{z}_t \mid\mathbf{x},\boldsymbol{\epsilon})\right] \right\|_2^2. \tag*{(6)}

The inner conditional expectation is generally intractable, as it requires posterior inference under the coupling. Conditional FM (CFM) [9] avoids this difficulty by directly matching the conditional velocity and leveraging the fact that its gradient coincides with that of the ideal FM objective,

LCFM=Et,x∼pdata,ϵ∼pbase∥vϕ(zt,t)−v(zt∣x,ϵ)∥22.(7)\mathcal{L}_{\mathrm{CFM}} = \mathbb{E}_{t, \mathbf{x}\sim p_{\mathrm{data}}, \boldsymbol{\epsilon}\sim p_{\mathrm{base}}} \left\| \mathbf{v}_{\phi}(\mathbf{z}_t,t) - \mathbf{v}(\mathbf{z}_t \mid\mathbf{x},\boldsymbol{\epsilon}) \right\|_2^2. \tag*{(7)}

Optimal-transport path. The linear schedule at=1−ta_t = 1-t, bt=tb_t = t used in the main paper (Eq. (1)) yields the interpolation zt=(1−t)x+tϵ\mathbf{z}_t = (1-t)\mathbf{x} + t\boldsymbol{\epsilon} and a constant conditional velocity v(zt∣x,ϵ)=ϵ−x\mathbf{v}(\mathbf{z}_t \mid\mathbf{x},\boldsymbol{\epsilon}) = \boldsymbol{\epsilon} - \mathbf{x}. This corresponds to straight-line trajectories in latent space and is referred to as the optimal-transport (OT) path [9]. Our teacher is trained under this parameterization.

TrigFlow Parameterization and OT Equivalence

TrigFlow. TrigFlow [20] replaces the linear OT schedule with a sinusoidal coupling

zt=cos⁡(t)x+sin⁡(t)ϵ,t∈[0,π2],(8)\mathbf{z}_t = \cos(t)\mathbf{x} + \sin(t)\boldsymbol{\epsilon}, \qquad t \in\left[0,\frac{\pi}{2}\right], \tag*{(8)}

with conditional velocity vt=cos⁡(t)ϵ−sin⁡(t)x\mathbf{v}_t = \cos(t)\boldsymbol{\epsilon} - \sin(t)\mathbf{x}. The continuous-time consistency framework of [20] is derived under this parameterization.

OT–TrigFlow equivalence. Chen et al. [21] establish a lossless equivalence between OT-FM and TrigFlow. Let tTrig∈[0,π/2]t_{\mathrm{Trig}} \in[0,\pi/2] and tOT∈[0,1]t_{\mathrm{OT}} \in[0,1]. The transforms are:

tOT=sin⁡(tTrig)sin⁡(tTrig)+cos⁡(tTrig),(9)t_{\mathrm{OT}} = \frac{\sin(t_{\mathrm{Trig}})}{\sin(t_{\mathrm{Trig}}) + \cos(t_{\mathrm{Trig}})}, \tag*{(9)}
zt,OT=zt,TrigtOT2+(1−tOT)2,(10)\mathbf{z}_{t,\mathrm{OT}} = \frac{\mathbf{z}_{t,\mathrm{Trig}}}{\sqrt{t_{\mathrm{OT}}^2 + (1-t_{\mathrm{OT}})^2}}, \tag*{(10)}
Fθ(zt,Trig,tTrig)=(1−2tOT)zt,OT+(1−2tOT+2tOT2)vOTtOT2+(1−tOT)2,(11)\mathbf{F}_{\theta}(\mathbf{z}_{t,\mathrm{Trig}}, t_{\mathrm{Trig}}) = \frac{(1-2t_{\mathrm{OT}})\mathbf{z}_{t,\mathrm{OT}} + (1-2t_{\mathrm{OT}}+2t_{\mathrm{OT}}^2)\mathbf{v}_{\mathrm{OT}}}{\sqrt{t_{\mathrm{OT}}^2 + (1-t_{\mathrm{OT}})^2}}, \tag*{(11)}

where vOT\mathbf{v}_{\mathrm{OT}} is the OT-FM velocity at (zt,OT,tOT)(z_{t,\mathrm{OT}}, t_{\mathrm{OT}}). Equation 9 maps TrigFlow time onto the OT interval, Eq. 10 rescales the latent for the differing path geometries, and Eq. 11 converts the OT velocity into its TrigFlow counterpart. Together these transforms allow our pretrained OT-FM teacher to provide TrigFlow-consistent supervision to the consistency student during distillation.

Consistency Models and sCD Loss

TrigFlow parameterization of the consistency function. Working in TrigFlow coordinates, the consistency function is parameterized in terms of the underlying velocity network as

fθ(zt,Trig,tTrig)=cos⁡(tTrig) zt,Trig−sin⁡(tTrig) Fθ(zt,Trig,tTrig),(12)f_\theta(z_{t,\mathrm{Trig}},t_{\mathrm{Trig}})=\cos(t_{\mathrm{Trig}})\,z_{t,\mathrm{Trig}}-\sin(t_{\mathrm{Trig}})\,F_\theta(z_{t,\mathrm{Trig}},t_{\mathrm{Trig}}), \tag*{(12)}

which automatically enforces the boundary condition fθ(z0,0)=xf_\theta(z_0,0)=\mathbf{x}.

Simplified continuous-time consistency distillation (sCD). Discrete-time consistency models train the self-consistency identity by sampling pairs (t,t+Δt)(t,t+\Delta t), but the resulting finite-difference approximation introduces discretization error [20]. The simplified continuous-time framework differentiates the self-consistency loss directly. The sCD loss is

LsCD(θ)=Ezt,t∥Fθ(zt,t)−Fθ−(zt,t)−cos⁡(t)ddtfθ−(zt,t)∥22,(13)\mathcal{L}_{\mathrm{sCD}}(\theta)=\mathbb{E}_{z_t,t}\left\|F_\theta(z_t,t)-F_{\theta^-}(z_t,t)-\cos(t)\frac{\mathrm{d}}{\mathrm{d}t}f_{\theta^-}(z_t,t)\right\|_2^2, \tag*{(13)}

where θ−\theta^- is an EMA of the student parameters with stop-gradient. The EMA target stabilizes the self-supervised objective by decoupling the regression target from the parameters being optimized.

JVP-based tangent estimation. The tangent ddtfθ−\frac{\mathrm{d}}{\mathrm{d}t}f_{\theta^-} in Eq. (13) decomposes as

ddtfθ−(zt,t)=∂fθ−∂ztdztdt+∂fθ−∂t,(14)\frac{\mathrm{d}}{\mathrm{d}t}f_{\theta^-}(z_t,t)=\frac{\partial f_{\theta^-}}{\partial z_t}\frac{\mathrm{d}z_t}{\mathrm{d}t}+\frac{\partial f_{\theta^-}}{\partial t}, \tag*{(14)}

and is computed exactly via a single Jacobian-vector product (JVP) against the tangent vector (dzt/dt,1)(\mathrm{d}z_t/\mathrm{d}t,1), where dzt/dt\mathrm{d}z_t/\mathrm{d}t is supplied by the teacher [20]. This avoids storing or differencing the network across multiple timesteps.

Algorithms

The two algorithms below give the per-step training procedures for the two stages of our pipeline. Algorithm 1 is the standard conditional flow-matching update (Eq. (7)) instantiated under the linear OT path of Eq. (1), and is used to train the unconditional OT-FM teacher on the high-resolution distribution. Algorithm 2 then distils that frozen teacher into a one-step TrigFlow consistency student: it samples a TrigFlow timestep, evaluates the teacher in OT coordinates via the lossless transforms of Appendix A.2, and uses the resulting trajectory tangent inside a single Jacobian–vector product to construct the sCD target of Eq. (13). The EMA copy Fθ−F_{\theta^-} is updated after each optimizer step to provide a stop-gradient regression target on the next iteration.

Algorithm 1 Single training step for the OT-FM teacher

Require: High-resolution sample x\mathbf{x}, velocity field vϕv_\phi, time sampler p(t)p(t) on (0,1)(0,1)

  1. Sample time t∼p(t)t\sim p(t), noise ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I)

  2. zt←(1−t)x+tϵ\mathbf{z}_t\leftarrow(1-t)\mathbf{x}+t\epsilon

  3. vt←ϵ−x\mathbf{v}_t\leftarrow\epsilon-\mathbf{x}

  4. v^←vϕ(zt,t)\hat{\mathbf{v}}\leftarrow v_\phi(\mathbf{z}_t,t)

  5. ℓ←∥v^−vt∥22\ell\leftarrow\|\hat{\mathbf{v}}-\mathbf{v}_t\|_2^2

  6. Update ϕ\phi with one optimizer step using ∇ϕℓ\nabla_\phi\ell

Algorithm 2 Single distillation step (OT-FM teacher →\to TrigFlow sCM student)

Require: Student FθF_\theta, EMA student Fθ−F_{\theta-}, OT-FM teacher vϕv_\phi, clean sample xx, TrigFlow–OT transforms TtT_t, TzT_z, TvT_v (Eqs. (9)–(11)), TrigFlow time sampler pTrigp_{\mathrm{Trig}}

  1. tTrig∼pTrigt_{\mathrm{Trig}} \sim p_{\mathrm{Trig}}, ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0,I)

  2. zt,Trig←xcos⁡(tTrig)+ϵsin⁡(tTrig)z_{t,\mathrm{Trig}} \leftarrow x\cos(t_{\mathrm{Trig}}) + \epsilon\sin(t_{\mathrm{Trig}})

  3. tOT←Tt(tTrig)t_{\mathrm{OT}} \leftarrow T_t(t_{\mathrm{Trig}}), zt,OT←Tz(zt,Trig,tOT)z_{t,\mathrm{OT}} \leftarrow T_z(z_{t,\mathrm{Trig}},t_{\mathrm{OT}})

  4. vOT←vϕ(zt,OT,tOT)v_{\mathrm{OT}} \leftarrow v_\phi(z_{t,\mathrm{OT}},t_{\mathrm{OT}})

  5. F^teacher←Tv(zt,OT,tOT,vOT)\hat{F}_{\mathrm{teacher}} \leftarrow T_v(z_{t,\mathrm{OT}},t_{\mathrm{OT}},v_{\mathrm{OT}}), Fstud←Fθ(zt,Trig,tTrig)F_{\mathrm{stud}} \leftarrow F_\theta(z_{t,\mathrm{Trig}},t_{\mathrm{Trig}}), Fema←Fθ−(zt,Trig,tTrig)F_{\mathrm{ema}} \leftarrow F_{\theta-}(z_{t,\mathrm{Trig}},t_{\mathrm{Trig}})

  6. fθ−(zt,Trig,tTrig)←cos⁡(tTrig)zt,Trig−sin⁡(tTrig)Femaf_{\theta-}(z_{t,\mathrm{Trig}},t_{\mathrm{Trig}}) \leftarrow\cos(t_{\mathrm{Trig}})z_{t,\mathrm{Trig}} - \sin(t_{\mathrm{Trig}})F_{\mathrm{ema}}

  7. ddtfθ−←JVP⁡(fθ−,(zt,Trig,tTrig),(F^teacher,1))\frac{\mathrm{d}}{\mathrm{d}t}f_{\theta-} \leftarrow\operatorname{JVP}\left(f_{\theta-},(z_{t,\mathrm{Trig}},t_{\mathrm{Trig}}),(\hat{F}_{\mathrm{teacher}},1)\right)

  8. target←stopgrad⁡(Fema+cos⁡(tTrig)ddtfθ−)\mathrm{target} \leftarrow\operatorname{stopgrad}\left(F_{\mathrm{ema}} + \cos(t_{\mathrm{Trig}})\frac{\mathrm{d}}{\mathrm{d}t}f_{\theta-}\right)

  9. LsCD←∥Fstud−target∥22\mathcal{L}_{\mathrm{sCD}} \leftarrow\left\|F_{\mathrm{stud}}-\mathrm{target}\right\|_2^2

  10. Update θ\theta with one optimizer step using ∇θLsCD\nabla_\theta\mathcal{L}_{\mathrm{sCD}}

Baseline Solver Details

This appendix specifies the multi-step solvers used by the diffusion and flow-matching baselines reported in § 4 (Table 1–Table 3 and Table 4). All three solvers share the inference-time conditioning of § 3.4, with the starting state xτ0x_{\tau_0} formed by noise injection (Eq. (3)) at the per-dataset τ\tau.

Denoising Diffusion Implicit Models

Denoising Diffusion Implicit Models (DDIM) [26] accelerates DDPM [6] sampling by replacing the Markovian reverse chain with a deterministic non-Markovian update that preserves the marginals of the forward diffusion process, so that fewer steps suffice while a consistent trajectory is preserved. Given a learned noise predictor ϵθ(xt,t)\epsilon_\theta(x_t,t) and a noise schedule {αˉt}t=0T\{\bar{\alpha}_t\}_{t=0}^{T}, the deterministic step from time ss to time u<su < s uses the implied clean prediction

x^0(xs,s)=xs−1−αˉsϵθ(xs,s)αˉs,(15)\hat{x}_0(x_s,s) = \frac{x_s-\sqrt{1-\bar{\alpha}_s}\epsilon_\theta(x_s,s)}{\sqrt{\bar{\alpha}_s}}, \tag*{(15)}

to advance the sample as

xu=αˉux^0(xs,s)+1−αˉuϵθ(xs,s).(16)x_u = \sqrt{\bar{\alpha}_u}\hat{x}_0(x_s,s) + \sqrt{1-\bar{\alpha}_u}\epsilon_\theta(x_s,s). \tag*{(16)}

We use 30 sampling steps on a uniformly-spaced grid over [0,T][0,T]. The full procedure is summarised in Algorithm 3.

Algorithm 3 DDIM sampling

Require: Noise predictor ϵθ\epsilon_\theta, schedule {αˉt}\{\bar{\alpha}_t\}, step grid T=τ0>τ1>⋯>τK=0T=\tau_0>\tau_1>\cdots>\tau_K=0 (we use K=30K=30), starting state xτ0x_{\tau_0}

  1. for k=0,…,K−1k=0,\ldots,K-1 do

  2. s←τk, u←τk+1s\leftarrow\tau_k,\ u\leftarrow\tau_{k+1}

  3. ϵs←ϵθ(xs,s)\epsilon_s\leftarrow\epsilon_\theta(x_s,s)

  4. x^0←(xs−1−αˉsϵs)/αˉs\hat{x}_0\leftarrow(x_s-\sqrt{1-\bar{\alpha}_s}\epsilon_s)/\sqrt{\bar{\alpha}_s}

  5. xu←αˉux^0+1−αˉuϵsx_u\leftarrow\sqrt{\bar{\alpha}_u}\hat{x}_0+\sqrt{1-\bar{\alpha}_u}\epsilon_s

  6. end for

Return: High-resolution sample x0x_0.

DPM-Solver++

DPM-Solver++ [37] reformulates the reverse diffusion ODE in log-SNR coordinates and integrates it with a multistep method, exploiting the semi-linear structure of the diffusion ODE to attain higher effective order than DDIM at the same number of function evaluations. With λt=log⁡(αˉt/1−αˉt)\lambda_t=\log(\sqrt{\bar{\alpha}_t}/\sqrt{1-\bar{\alpha}_t}) as the log-SNR variable, σt=1−αˉt\sigma_t=\sqrt{1-\bar{\alpha}_t}, and x^0(xt,t)\hat{x}_0(x_t,t) as defined for DDIM in Appendix § 7.§ 1, the second-order multistep variant DPM-Solver++(2M), which we use here, advances the sample from λs\lambda_s to λu\lambda_u via

xu=σuσsxs−αˉu(e−hk−1)Dk,hk=λu−λs,(17)x_u=\frac{\sigma_u}{\sigma_s}x_s-\sqrt{\bar{\alpha}_u}\left(e^{-h_k}-1\right)D_k,\qquad h_k=\lambda_u-\lambda_s, \tag*{(17)}

where the multistep operator DkD_k blends the current and previous clean predictions according to

Dk=(1+12rk)x^0(xs,s)−12rkx^0(xsk−1,sk−1),rk=hk−1hk,(18)D_k=\left(1+\frac{1}{2r_k}\right)\hat{x}_0(x_s,s)-\frac{1}{2r_k}\hat{x}_0(x_{s_{k-1}},s_{k-1}),\qquad r_k=\frac{h_{k-1}}{h_k}, \tag*{(18)}

with D0=x^0(xτ0,τ0)D_0=\hat{x}_0(x_{\tau_0},\tau_0) for the first step. We use 30 steps on the same uniformly-spaced grid as DDIM. The full procedure is summarised in Algorithm 4.

Algorithm 4 DPM-Solver++(2M) sampling

Require: Noise predictor ϵθ\epsilon_\theta, schedule {αˉt}\{\bar{\alpha}_t\}, step grid T=τ0>τ1>⋯>τK=0T = \tau_0 > \tau_1 > \cdots> \tau_K = 0 (we use K=30K = 30), starting state xτ0x_{\tau_0}

1: Pre-compute λk=log⁡(αˉτk/1−αˉτk)\lambda_k = \log\left(\sqrt{\bar{\alpha}_{\tau_k}}/\sqrt{1-\bar{\alpha}_{\tau_k}}\right) and σk=1−αˉτk\sigma_k = \sqrt{1-\bar{\alpha}_{\tau_k}} for k=0,…,Kk = 0,\ldots,K 2: x^0(0)←(xτ0−σ0ϵθ(xτ0,τ0))/αˉτ0\hat{x}_0^{(0)} \leftarrow(x_{\tau_0} - \sigma_0\epsilon_\theta(x_{\tau_0},\tau_0))/\sqrt{\bar{\alpha}_{\tau_0}} 3: for k=0,…,K−1k = 0,\ldots,K-1 do 4: hk←λk+1−λkh_k \leftarrow\lambda_{k+1} - \lambda_k 5: if k=0k = 0 then 6: Dk←x^0(k)D_k \leftarrow\hat{x}_0^{(k)} 7: else 8: rk←hk−1/hkr_k \leftarrow h_{k-1}/h_k 9: Dk←(1+1/(2rk))x^0(k)−(1/(2rk))x^0(k−1)D_k \leftarrow(1+1/(2r_k))\hat{x}_0^{(k)}-(1/(2r_k))\hat{x}_0^{(k-1)} 10: end if 11: xτk+1←(σk+1/σk)xτk−αˉτk+1(e−hk−1)Dkx_{\tau_{k+1}} \leftarrow(\sigma_{k+1}/\sigma_k)x_{\tau_k}-\sqrt{\bar{\alpha}_{\tau_{k+1}}}(e^{-h_k}-1)D_k 12: x^0(k+1)←(xτk+1−σk+1ϵθ(xτk+1,τk+1))/αˉτk+1\hat{x}_0^{(k+1)} \leftarrow(x_{\tau_{k+1}}-\sigma_{k+1}\epsilon_\theta(x_{\tau_{k+1}},\tau_{k+1}))/\sqrt{\bar{\alpha}_{\tau_{k+1}}} 13: end for

Return: High-resolution sample xτKx_{\tau_K}.

Heun Method

The Heun (improved-Euler) method is a two-stage explicit Runge–Kutta integrator that is second-order accurate in the step size [38]. Applied to the unconditional flow-matching ODE dz/dt=vϕ(z,t)\mathrm{d}z/\mathrm{d}t = v_\phi(z,t), the step from time tt to time t′=t+Δtt^\prime= t + \Delta t uses one predictor evaluation followed by one corrector evaluation,

k1=vϕ(zt,t),z~=zt+Δt k1,k2=vϕ(z~,t′),zt′=zt+Δt2(k1+k2),(19)k_1 = v_\phi(z_t,t), \quad\tilde{z} = z_t + \Delta t\,k_1, \quad k_2 = v_\phi(\tilde{z},t^\prime), \quad z_{t^\prime} = z_t + \frac{\Delta t}{2}(k_1+k_2), \tag*{(19)}

costing 2 NFEs per outer step. We integrate backward from t=τt = \tau to t=0t = 0 over 5 outer steps (10 NFEs total) on a uniform grid. The full procedure is summarised in Algorithm 5.

Algorithm 5 Heun sampling for the flow-matching ODE

Require: Velocity field vϕv_\phi, step grid τ=t0>t1>⋯>tK=0\tau= t_0 > t_1 > \cdots> t_K = 0 (we use K=5K = 5), starting state zt0z_{t_0}

1: for k=0,…,K−1k = 0,\ldots,K-1 do 2: Δt←tk+1−tk\Delta t \leftarrow t_{k+1} - t_k 3: k1←vϕ(ztk,tk)k_1 \leftarrow v_\phi(z_{t_k},t_k) 4: z~←ztk+Δt k1\tilde{z} \leftarrow z_{t_k} + \Delta t\,k_1 5: k2←vϕ(z~,tk+1)k_2 \leftarrow v_\phi(\tilde{z},t_{k+1}) 6: ztk+1←ztk+Δt2(k1+k2)z_{t_{k+1}} \leftarrow z_{t_k} + \frac{\Delta t}{2}(k_1+k_2) 7: end for

Return: High-resolution sample ztKz_{t_K}.

Per-Dataset Qualitative Samples and Power Spectra

This appendix collects the per-dataset super-resolved comparisons and power spectra that are not shown in the main paper. The qualitative figures follow the same layout as Figure 2 — row 1 is the nearest-neighbour-upsampled low-resolution input, row 2 is the high-resolution ground truth, row 3 is the one-step sCM student reconstruction, and rows 4–5 are the per-pixel absolute residuals of the sCM student and the multi-step OT-FM teacher against the ground truth. The three field rows share a single field colorbar and the two error rows share a single sequential error colorbar, so brighter pixels denote larger absolute deviation from the ground truth.

Reproducibility. Samples A–F of Figure 2 (Smoke Buoyancy) are at indices 4173, 30000, 50000, 78901, 135000, 165000 of 128_smoke_hr.npy. Samples A–F of Figure 4 (TCF) are at indices 1287, 9461, 14906, 24561, 29871, 35402 of 192_tcf_hr.npy. Samples A–F of Figure 5 (Kolmogorov Flow) are at indices 487, 2589, 4823, 7184, 9523, 11947 of 256_kmflow.npy. Indices were chosen to span the full dataset with non-consecutive spacing.

Power spectra on the remaining benchmarks. We complement the Kolmogorov spectrum in Figure 3(b) with the corresponding spectra on the Smoke Buoyancy and Turbulent Channel Flow benchmarks (Figure 6). The protocol matches that figure — per-channel mean-centered 2D DFT, fftshift, normalised power, and a radial average over wavenumber bins (Appendix § 11.§ 3); spectra are averaged over the same validation subset used for the metric tables. On Smoke Buoyancy (panel a), the nearest-neighbour baseline shows a sharp drop near the input Nyquist with spurious energy beyond, while both the multi-step FM teacher and the one-NFE sCM student recover the ground-truth tail across the resolved range, consistent with the PSDD ranking in Table 1. On TCF (panel b), the bulk channel profile dominates the low-wavenumber content, and the residual gap between the sCM student and the FM teacher at intermediate wavenumbers tracks the per-table PSDD discrepancy in Table 2.

(a) (a) Smoke Buoyancy (32→128)

(b) (b) Turbulent Channel Flow (64→192)

Figure 6. Power spectra on the remaining benchmarks. Radially-averaged spectra of the ground truth, nearest-neighbour LR baseline, multi-step FM teacher, and the one-NFE sCM student (15.9 M for Smoke Buoyancy, 14.9 M for TCF).

Inference-Time Sensitivity Studies

This appendix collects sensitivity studies on the two stochastic ingredients of the inference pipeline of § 3.4: the noise-injection time τ\tau that governs the fidelity–realism trade-off, and the Gaussian noise vector ϵ\epsilon drawn each time a sample is generated. The first study justifies the per-dataset τ\tau used in the main paper; the second quantifies how much of the metric values reported in § 4 can be attributed to the random draw of ϵ\epsilon.

Noise-strength sweep (τ\tau)

We sweep τ∈(0,1]\tau\in(0, 1] on a 50-point uniform grid and, at each grid point, run the distilled sCM student on Nsamples=30N_{\mathrm{samples}} = 30 held-out low-resolution frames, recording the RMSE against the matched ground-truth high-resolution targets. The sweep is implemented as an Optuna [39] GridSampler study with a sqlite backend to enable resumption. Each trial draws a fresh ϵ\epsilon, so the per-trial RMSE inherits a small amount of sampling noise that is characterised separately in Appendix D.2.

Distilled sCM L₂ loss versus noise strength τ for Smoke (32 → 128)

Figure 7. Smoke Buoyancy noise-strength sweep. RMSE of the distilled sCM student against the matched ground truth as a function of the inference noise-injection time τ\tau, on the Smoke Buoyancy (32→\to128) task (Nsamples=30N_{\mathrm{samples}} = 30, 50-point uniform τ\tau grid). The empirical optimum is τ⋆=0.56\tau^\star= 0.56 with RMSE 0.1977; the default τ=0.3\tau= 0.3 used in the main paper yields RMSE 0.2031 (+2.7% relative to the optimum), well inside the plateau (RMSE within +5% of optimum for τ∈[0.18,0.76]\tau\in[0.18, 0.76]). RMSE rises sharply only past τ≈0.85\tau\approx0.85, reaching 0.619 at τ=1\tau= 1 where the LR observation is essentially erased.

The curve has the characteristic shape predicted by the trade-off in § 3.4: a wide low-RMSE plateau spanning roughly half the τ\tau range, and a sharp rise as τ→1\tau\to1 where the noised initialization zτz_\tau approaches pure Gaussian noise and the conditioning information from xLRx_{\mathrm{LR}} is destroyed. Within the plateau the choice of τ\tau is largely cosmetic; the practical sensitivity emerges only near the upper boundary. The operating point τ=0.3\tau= 0.3 used throughout the main paper is suboptimal by under three percent in RMSE and lies comfortably inside the plateau, while also matching the regime in which the consistency student was trained.

Impact of generation randomness

The sCM student is a stochastic generator: each call draws a fresh Gaussian noise vector ϵ\epsilon to form the noised initialization zτz_\tau in Eq. (3). To verify that the headline metrics reported in § 4 are not artefacts of a particular noise draw, we repeat the validation-split evaluation K=100K = 100 times with independent random seeds, holding τ=0.3\tau= 0.3 and the model weights fixed at their main-paper values. To make each per-seed estimate itself stable, every run scores 64 evenly-spaced frames from the held-out validation split (a 4×\times wider subset than the main-table evaluation), so the spread we report below is dominated by the random draw of ϵ\epsilon rather than by frame sub-sampling. For each run we record RL2RL_2, SSIM, and PSDD using the implementations from Section 4 and Appendix F.

MetricMean ±\pm SD5th / 95th %min / maxσ/μ\sigma/\mu
RL2↓RL_2 \downarrow0.1470 ±\pm 0.00070.1459 / 0.14830.1454 / 0.14890.50%
SSIM ↑\uparrow0.6720 ±\pm 0.00410.6647 / 0.67790.6590 / 0.67920.61%
PSDD ↓\downarrow(3.70 ±\pm 0.46) ×10−2\times10^{-2}2.96×10−22.96 \times10^{-2} / 4.43×10−24.43 \times10^{-2}2.79×10−22.79 \times10^{-2} / 4.97×10−24.97 \times10^{-2}12.3%

Table 5. Sensitivity of the distilled sCM student’s evaluation metrics on Smoke Buoyancy (32→\to128) to the inference-time noise draw, over K=100K = 100 random seeds at fixed τ=0.3\tau= 0.3 and a fixed 64-frame validation subset. Reported values are mean ±\pm standard deviation, with 5th/95th percentiles and full min/max ranges. The relative spread σ/μ\sigma/\mu characterises how much a single-seed evaluation can shift in either direction.

The pointwise metrics (RL2RL_2, SSIM) cluster within roughly half a percent of their mean across 100 seeds: the digits cited for the distilled sCM student in Table 1 are therefore stable to the random draw of ϵ\epsilon at the precision reported, and the comparison against the FM teacher is not sensitive to seed choice. PSDD has a substantially wider relative spread (σ/μ≈12%\sigma/\mu\approx12\%), as expected: it is a relative L2L_2 distance between one-dimensional radial spectra (Appendix F), and small fluctuations in the high-wavenumber tail produce a disproportionately large response in this metric. Even at the 95th-percentile worst case (4.43×10−24.43 \times10^{-2}), the distilled student remains close to the multi-step diffusion baseline (4.35×10−24.35 \times10^{-2} at NFE = 30 in Table 1) and only modestly above its own mean.

Datasets

We evaluate on three two-dimensional fluid-dynamics benchmarks spanning different resolutions, flow regimes, and dataset sizes. All datasets are split 90%/10% into training and validation; data statistics (per-channel mean and standard deviation) are pre-computed on the training split and used to standardise inputs to both teacher and student.

2D Kolmogorov Flow

Short description. 12,800 snapshots of 2D periodic incompressible Kolmogorov Flow generated from the two-dimensional incompressible Navier-Stokes equations with a sinusoidal forcing term at resolution 256×256256 \times256. The coarse input is downsampled to 64×6464 \times64 by stride-4 subsampling and nearest-neighbour upsampled back to 256×256256 \times256 before inference, giving a 4× super-resolution task.

Governing Equations. The 2D Kolmogorov Flow is simulated as a canonical benchmark for statistically stationary turbulence governed by the incompressible Navier-Stokes equations. The flow is described in the vorticity formulation of the 2D incompressible Navier-Stokes system,

∂ω∂t+u⋅∇ω=ν∇2ω+∇×F,∇⋅u=0,ω=∇×u,(20)\frac{\partial\omega}{\partial t}+\mathbf{u}\cdot\nabla\omega=\nu\nabla^{2}\omega+\nabla\times\mathbf{F},\qquad\nabla\cdot\mathbf{u}=0,\qquad\omega=\nabla\times\mathbf{u}, \tag*{(20)}

where u(x,y,t)\mathbf{u}(x,y,t) is the velocity field, ω(x,y,t)\omega(x,y,t) is the scalar vorticity, ν=1/Re\nu=1/\mathrm{Re} is the kinematic viscosity, and F\mathbf{F} is a steady, large-scale sinusoidal body force

F(x,y)=0.1[sin⁡(2π(x+y))+cos⁡(2π(x+y))],(21)\mathbf{F}(x,y)=0.1\left[\sin(2\pi(x+y))+\cos(2\pi(x+y))\right], \tag*{(21)}

that drives the flow into a statistically stationary turbulent regime. The system is solved on the periodic square [0,1]×[0,1][0,1]\times[0,1] discretised on a 256×256256\times256 grid, with Re=1000\mathrm{Re}=1000. Time integration uses a pseudo-spectral scheme with Crank–Nicolson stepping at Δt=10−3 s\Delta t=10^{-3}\,\mathrm{s}, with the nonlinear advective term computed in Fourier space and dealiased by the standard 2/3-truncation rule. Each trajectory is initialised from a Gaussian random vorticity field with energy spectrum E(k)∝[1+(τk)2]−αE(k)\propto[1+(\tau k)^2]^{-\alpha} (α=2.5\alpha=2.5, τ=7\tau=7), and snapshots are recorded at uniform temporal intervals over a horizon of T=5000 sT=5000\,\mathrm{s} across multiple independent realisations to provide diverse, statistically stationary turbulent samples.

Turbulent Channel Flow

Short description. 38,080 snapshots at resolution 192×192192\times192 at cross-section of 3D Turbulent Channel Flow. The low-resolution input is obtained by stride-3 sub-sampling to 64×6464\times64 and nearest-neighbour upsampled back to 192×192192\times192, preserving the anisotropic structure of near-wall turbulence.

Governing Equations. We use the dataset obtained from Xue et al. [40], in which the three-dimensional Turbulent Channel Flow is simulated by a large-eddy lattice Boltzmann method (LES-LBM). The LBM evolves a discrete-velocity distribution function fi(x,t)f_i(\mathbf{x},t) on a Cartesian lattice with velocity set {ci}\{\mathbf{c}_i\} via the Bhatnagar–Gross–Krook (BGK) collision-streaming equation

fi(x+ciΔt,t+Δt)−fi(x,t)=−1τ[fi(x,t)−fieq(x,t)],(22)f_i(\mathbf{x}+\mathbf{c}_i\Delta t,t+\Delta t)-f_i(\mathbf{x},t)=-\frac{1}{\tau}\left[f_i(\mathbf{x},t)-f_i^{\mathrm{eq}}(\mathbf{x},t)\right], \tag*{(22)}

with the local equilibrium fieqf_i^{\mathrm{eq}} taken in its standard low-Mach polynomial form and the macroscopic density and momentum recovered as ρ=∑ifi\rho=\sum_i f_i and ρu=∑icifi\rho\mathbf{u}=\sum_i\mathbf{c}_i f_i. The relaxation time τ\tau is linked to an effective kinematic viscosity through νeff=cs2(τ−12)Δt\nu_{\mathrm{eff}}=c_s^2(\tau-\frac{1}{2})\Delta t, in which a Smagorinsky sub-grid eddy viscosity is embedded,

νeff=ν0+νt,νt=CsmagΔ2S‾,(23)\nu_{\mathrm{eff}}=\nu_0+\nu_t,\qquad\nu_t=C_{\mathrm{smag}}\Delta^2\overline{S}, \tag*{(23)}

where ν0\nu_0 is the molecular viscosity, S‾\overline{S} is the magnitude of the filtered strain-rate tensor, and Δ\Delta is the lattice spacing. In the low-Mach, weakly-compressible limit, this scheme is consistent with the filtered incompressible Navier-Stokes equations,

∂u∂t+u⋅∇u=−1ρ∇p+∇⋅[(ν0+νt)(∇u+∇uT)]+f,∇⋅u=0,(24)\frac{\partial\mathbf{u}}{\partial t}+\mathbf{u}\cdot\nabla\mathbf{u}=-\frac{1}{\rho}\nabla p+\nabla\cdot\left[(\nu_0+\nu_t)(\nabla\mathbf{u}+\nabla\mathbf{u}^{\mathsf{T}})\right]+\mathbf{f},\qquad\nabla\cdot\mathbf{u}=0, \tag*{(24)}

with u=(ux,uy,uz)\mathbf{u} = (u_x, u_y, u_z) along the streamwise, wall-normal and spanwise directions and f\mathbf{f} a uniform body force driving the channel. The simulation is carried out on a domain Lx×Ly×Lz=768×192×192L_x \times L_y \times L_z = 768 \times192 \times192 at friction Reynolds number Re⁡τ=180\operatorname{Re}_{\tau} = 180 (bulk Re⁡=3250\operatorname{Re} = 3250), with periodic boundary conditions in the streamwise and spanwise directions and no-slip walls in the wall-normal direction. The flow is initialised from a quiescent state and perturbed by a cubic obstruction of size 20×20×10020 \times20 \times100 grid points placed at x=192x = 192 to trigger transition; after a long transient and removal of the obstruction the simulation is advanced for a further 50 domain-through times to reach a statistically stationary turbulent state. The 2D fields used in this work are obtained by sampling the mid-plane cross-section at x=384x = 384 over ∼100\sim100 turnover times, yielding the 192×192192 \times192 snapshots that serve as ground-truth high-fidelity data.

2D Smoke Buoyancy

Short description. 173,824 snapshots at resolution 128×128128 \times128, capturing buoyancy-driven scalar transport. The coarse input is downsampled to 32×3232 \times32 by stride-4 subsampling and nearest-neighbour upsampled back to 128×128128 \times128, giving a 4×4 \times super-resolution factor.

Governing Equations. We use the dataset from Zhou et al. [41], which is based on data from Gupta and Brandstetter [42] and was originally generated using the Φflow\Phi_{\mathrm{flow}} (PhiFlow) solver [43]. The simulations contain 56 timesteps, spanning t=18t = 18 to t=102t = 102, and are solved on a two-dimensional spatial domain, x∈[0,32]x \in[0, 32] and y∈[0,32]y \in[0, 32], at a resolution of 128×128128 \times128. The data describes smoke driven by a buoyant force in a constrained box. The dynamics are governed by the incompressible Navier-Stokes equations, coupled with an advection equation for the smoke density dd:

∂v∂t+v⋅∇v=ν∇2v−1ρ∇p+f,∇⋅v=0.(25)\frac{\partial\mathbf{v}}{\partial t} + \mathbf{v} \cdot\nabla\mathbf{v} = \nu\nabla^2 \mathbf{v} - \frac{1}{\rho} \nabla p + \mathbf{f}, \qquad\nabla\cdot\mathbf{v} = 0. \tag*{(25)}
∂d∂t+v⋅∇d=0.(26)\frac{\partial d}{\partial t} + \mathbf{v} \cdot\nabla d = 0. \tag*{(26)}

Evaluation Metrics

We evaluate predicted 2D physical fields against ground truth using three complementary metrics: a pointwise relative error, a local structural similarity score, and a spectral discrepancy in Fourier space. Together they capture pixel-level fidelity, local geometric agreement, and the distribution of energy across spatial scales.

Relative L2L_2 Error

Let z^kj\hat{z}_k^j and zkjz_k^j denote the predicted and ground-truth states at step kk of the jj-th test trajectory. The relative L2L_2 error is

rel.⁡L2(k)=1N∑j=1N∥z^kj−zkj∥2∥zkj∥2,(27)\operatorname{rel.} L_2(k) = \frac{1}{N}\sum_{j=1}^{N}\frac{\left\lVert\hat{z}_k^j-z_k^j \right\rVert_2}{\left\lVert z_k^j \right\rVert_2}, \tag*{(27)}

averaged over NN test trajectories. We report this quantity per step to characterize how error accumulates over the prediction horizon.

Structural Similarity Index Measure (SSIM)

To quantify local structural agreement between a reference field x∈RH×Wx \in\mathbb{R}^{H \times W} and a prediction y∈RH×Wy \in\mathbb{R}^{H \times W}, we use SSIM [36], which compares local luminance, contrast, and structure within a sliding Gaussian window.

Local statistics. Let H(⋅)\mathcal{H}(\cdot) denote convolution with a normalized 2D Gaussian kernel of window size w×ww \times w (with ww odd) and standard deviation σ\sigma, implemented as a separable 1D convolution with “same” padding. The local means, variances, and cross-covariance are

μx=H(x),μy=H(y),σx2=H(x2)−μx2,σy2=H(y2)−μy2,σxy=H(xy)−μxμy.(28)\mu_x = \mathcal{H}(x), \qquad\mu_y = \mathcal{H}(y), \qquad\sigma_x^2 = \mathcal{H}(x^2)-\mu_x^2, \qquad\sigma_y^2 = \mathcal{H}(y^2)-\mu_y^2, \qquad\sigma_{xy} = \mathcal{H}(xy)-\mu_x\mu_y. \tag*{(28)}

SSIM map and aggregation. The pointwise SSIM is

SSIM⁡(x,y)=(2μxμy+C1)(2σxy+C2)(μx2+μy2+C1)(σx2+σy2+C2),(29)\operatorname{SSIM}(x,y) = \frac{(2\mu_x\mu_y+C_1)(2\sigma_{xy}+C_2)}{(\mu_x^2+\mu_y^2+C_1)(\sigma_x^2+\sigma_y^2+C_2)}, \tag*{(29)}

with stability constants C1=(K1L)2C_1=(K_1L)^2 and C2=(K2L)2C_2=(K_2L)^2, using the standard choices K1=0.01K_1=0.01 and K2=0.03K_2=0.03. We report the spatial mean

SSIM⁡‾(x,y)=1HW∑i=1H∑j=1WSSIM⁡ij(x,y),(30)\overline{\operatorname{SSIM}}(x,y) = \frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}\operatorname{SSIM}_{ij}(x,y), \tag*{(30)}

computed independently per sample for batched inputs of shape (B,H,W,1)(B,H,W,1) and averaged over the batch.

Per-sample data range. Because our fields are unnormalized physical quantities whose magnitudes vary across samples, we set the data range adaptively as L=max⁡(x)−min⁡(x)L=\max(x)-\min(x) rather than using a fixed global value. In the degenerate case L=0L=0 (constant reference), we define SSIM⁡=1\operatorname{SSIM}=1 when xx and yy agree to within numerical tolerance, and 00 otherwise.

Power Spectral Density Discrepancy (PSDD)

To assess whether predictions reproduce the correct distribution of energy across spatial scales—a property of central importance for turbulent and other multi-scale fields—we compare normalized 2D power spectra under an ℓ1\ell_1 distance.

Mean-centering. We first remove the spatial mean of each sample to suppress the dominant DC component and emphasize fluctuations:

x←x−Ei,j[x],y←y−Ei,j[y],(31)x \leftarrow x-\mathbb{E}_{i,j}[x], \qquad y \leftarrow y-\mathbb{E}_{i,j}[y], \tag*{(31)}

where Ei,j\mathbb{E}_{i,j} denotes averaging over spatial indices.

Centered Fourier transform. Let F{⋅}\mathcal{F}\{\cdot\} denote the 2D DFT over spatial axes, applied per sample and channel. We centre the zero-frequency component via fftshift:

F~x=fftshift⁡(F{x}),F~y=fftshift⁡(F{y}).(32)\widetilde{F}_x = \operatorname{fftshift}(\mathcal{F}\{x\}), \qquad\widetilde{F}_y = \operatorname{fftshift}(\mathcal{F}\{y\}). \tag*{(32)}

Overflow-safe normalized PSD. Naive squaring of F~\widetilde{F} can overflow in single precision on high-dynamic-range scientific fields. We therefore rescale the real and imaginary parts by a per-sample factor

s=max⁡i,j,cmax⁡(∣ℜ(F~i,j,c)∣,∣ℑ(F~i,j,c)∣)(33)s = \max_{i,j,c} \max\left(\left|\Re(\widetilde{F}_{i,j,c})\right|,\left|\Im(\widetilde{F}_{i,j,c})\right|\right) \tag*{(33)}

prior to squaring, giving a numerically stable power spectrum

P=(ℜ(F~)max⁡(s,ε))2+(ℑ(F~)max⁡(s,ε))2.(34)P = \left(\frac{\Re(\widetilde{F})}{\max(s,\varepsilon)}\right)^2 + \left(\frac{\Im(\widetilde{F})}{\max(s,\varepsilon)}\right)^2. \tag*{(34)}

We then normalize PP to a probability distribution over frequency bins,

P^=P∑i,j,cPi,j,c+ε,(35)\widehat{P} = \frac{P}{\sum_{i,j,c} P_{i,j,c} + \varepsilon}, \tag*{(35)}

which removes the residual amplitude dependence introduced by ss and isolates the relative allocation of spectral energy.

Discrepancy. Given normalized PSDs P^x(b)\widehat{P}_x^{(b)} and P^y(b)\widehat{P}_y^{(b)} for the bb-th sample, the PSDD is the per-bin ℓ1\ell_1 distance averaged over space, channels, and the batch:

LPSD(x,y)=1B∑b=1B1HWC∑i,j,c∣P^x,i,j,c(b)−P^y,i,j,c(b)∣.(36)\mathcal{L}_{\mathrm{PSD}}(x,y) = \frac{1}{B}\sum_{b=1}^{B}\frac{1}{HWC}\sum_{i,j,c}\left|\widehat{P}_{x,i,j,c}^{(b)}-\widehat{P}_{y,i,j,c}^{(b)}\right|. \tag*{(36)}

Lower values indicate closer agreement in spectral content between prediction and reference.

Hyperparameters

This appendix lists the full set of hyperparameters used to produce the results in § 4. All values correspond to the configurations of the FM teacher and sCM student checkpoints reported in the main paper, which we will release alongside the code.

All models are implemented in JAX/Flax NNX and trained on a single NVIDIA GPU per run, with Orbax checkpointing and an Adam optimiser (β1=0.9\beta_1=0.9, β2=0.999\beta_2=0.999, ε=10−8\varepsilon=10^{-8}). Inputs are normalised with per-dataset mean/standard deviation statistics computed from the training split; the train/validation split is 90%/10% across all datasets.

Network architecture

Both the FM teacher and the sCM student use the time-conditioned UNet of [6] with sinusoidal time embeddings and GroupNorm. The low-resolution observation is concatenated channel-wise with the noised field at every forward pass; no auxiliary encoder is used. Single attention block at the 16×1616\times16 feature scale, dropout disabled (p=0p=0), strided convolutions for resampling, single input/output channel for the scalar fields. Table 6 lists the role-specific shape parameters; the spatial resolution follows the dataset.

Table 6: UNet hyperparameters per role. CC: base channel width (latent_dims); MM: per-level channel multipliers; BB: residual blocks per level. The image resolution matches the target output of each dataset (Smoke Buoyancy: 128; Turbulent Channel Flow: 192; Kolmogorov Flow: 256).

RoleCCMMBBAttention scaleDropout
FM teacher96(1, 1, 1, 4)216×1616\times160.0
sCM student96(1, 1, 1, 2)316×1616\times160.0

Table 6.

Training

All runs use a fixed random seed of 42, gradient-norm clipping at 1.0, and the Adam optimiser. The two stages differ in their time sampler statistics, loss weighting and EMA usage. Table 7 lists the optimiser-side hyperparameters for each (method, dataset) combination, and Table 8 lists the method-specific objective hyperparameters.

MethodDatasetBatchEpochsLR (init)ScheduleWeight decay
FM teacherSmoke Buoyancy641501e−41\mathrm{e}{-4}cosine, 370,000 steps1e−41\mathrm{e}{-4}
Turbulent Channel Flow161501e−41\mathrm{e}{-4}constant—
Kolmogorov Flow321501e−41\mathrm{e}{-4}constant1e−41\mathrm{e}{-4}
sCM studentSmoke Buoyancy161502e−52\mathrm{e}{-5}constant—
Turbulent Channel Flow161502e−52\mathrm{e}{-5}constant—
Kolmogorov Flow161502e−52\mathrm{e}{-5}constant—

Table 7.

HyperparameterFM teachersCM student
PμP_{\mu} (logit-normal mean)−0.4-0.4−0.4-0.4
PσP_{\sigma} (logit-normal std.)1.01.00.80.8
tεt_{\varepsilon} (boundary clip)10−310^{-3}10−610^{-6}
Adaptive loss weightingyes (p=1p=1, ε=0.01\varepsilon=0.01)no
EMA rate αˉ\bar{\alpha}— (unused at training)0.980.98
Tangent clip—1.01.0
Tangent-warmup iterations—500500

Table 8. Method-specific objective hyperparameters. Time samples follow a logit-normal distribution log⁡(t/(1−t))∼N(Pμ,Pσ2)\log(t/(1-t)) \sim\mathcal{N}(P_{\mu}, P_{\sigma}^{2}), clipped at tεt_{\varepsilon} from the boundary. The FM teacher uses an adaptive loss weighting [44] with norm p=1p=1 and εnorm=0.01\varepsilon_{\mathrm{norm}}=0.01; the sCM student uses an unweighted loss with TrigFlow tangent clipping [20] and a linear warmup of the tangent term.

Table 7: Optimiser-side training hyperparameters. “Schedule” refers to the learning-rate schedule; constant keeps the listed value for the full run, cosine decays linearly between the listed initial value and 10−510^{-5} over the indicated number of steps. Weight decay is applied only when listed.

Inference

At inference time the conditioning input xLR↑\mathbf{x}^{\uparrow}_{\mathrm{LR}} (the upsampled low-resolution field defined in § 3.1) is mixed with Gaussian noise to construct the initial state ztstart\mathbf{z}_{t_{\mathrm{start}}} via Eq. (3), after which the FM teacher integrates the learned velocity field with a fifth-order Runge–Kutta solver (RK5), while the sCM student performs a single forward pass. Table 9 summarises the inference-time settings.

StageSolverNFEtstartt_{\mathrm{start}}
FM teacherRK5550.60.6
sCM student—110.30.3

Table 9. Inference settings. NFE: number of neural function evaluations per sample; these are dataset-independent.

References

  1. [1]Stephen B. Pope. Turbulent Flows. Cambridge University Press, 2000.DOI
  2. [2]Myoungkyu Lee and Robert D. Moser. Direct numerical simulation of turbulent channel flow up to Re_T ≈ 5200. Journal of Fluid Mechanics, 774:395–415, 2015. doi: 10.1017/jfm.2015.268.DOI
  3. [3]Giancarlo Alfonsi. Reynolds-Averaged Navier–Stokes Equations for Turbulence Modeling. Applied Mechanics Reviews, 62(4):040802, 2009. doi: 10.1115/1.3124648.DOI
  4. [4]Ugo Piomelli. Large-eddy simulation: achievements and challenges. Progress in Aerospace Sciences, 35(4):335–362, 1999. doi: 10.1016/S0376-0421(98)00014-1.DOI
  5. [5]Zhong Yi Wan, Ricardo Baptista, Anudhyan Boral, Yi-Fan Chen, John Anderson, Fei Sha, and Leonardo Zepeda-Nunez. Debias coarsely, sample conditionally: Statistical downscaling through optimal transport and probabilistic diffusion models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=5NxJuc0T1P.arxiv.org/abs/2305.15618
  6. [6]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020.arxiv.org/abs/2006.11239
  7. [7]Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.arxiv.org/abs/2011.13456
  8. [8]Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019.arxiv.org/abs/1907.05600
  9. [9]Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022.arxiv.org/abs/2210.02747
  10. [10]Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023.arxiv.org/abs/2303.08797
  11. [11]Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. In European conference on computer vision, pages 430–448. Springer, 2024.DOI
  12. [12]Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image super-resolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796–25805, 2024.arxiv.org/abs/2311.14760
  13. [13]Xiao Xue, Tianyue Yang, Mingyang Gao, Leyu Pan, Maida Wang, Kewei Zhu, Shuo Wang, Jiuling Li, Marco FP ten Eikelder, and Peter V Coveney. Uni-flow: a unified autoregressive-diffusion model for complex multiscale flows. arXiv preprint arXiv:2602.15592, 2026.arxiv.org/abs/2602.15592
  14. [14]Tianyue Yang and Xiao Xue. Meno: Meanflow-enhanced neural operators for dynamical systems. arXiv preprint arXiv:2604.06881, 2026.arxiv.org/abs/2604.06881
  15. [15]Dule Shu, Zijie Li, and Amir Barati Farimani. A physics-informed diffusion model for high-fidelity flow field reconstruction. Journal of Computational Physics, 478:111972, 2023.DOI
  16. [16]Vivek Oommen, Aniruddha Bora, Zhen Zhang, and George Em Karniadakis. Integrating neural operators with diffusion models improves spectral representation in turbulence modelling. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 481 (2309), 2025.
  17. [17]Thomas YL Lin, Jiachen Yao, Lufang Chiang, Julius Berner, and Anima Anandkumar. Decoupled diffusion sampling for inverse problems on function spaces. arXiv preprint arXiv:2601.23280, 2026.arxiv.org/abs/2601.23280
  18. [18]Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023.arxiv.org/abs/2303.01469
  19. [19]Yang Song and Prafula Dhariwal. Improved techniques for training consistency models. arXiv preprint arXiv:2310.14189, 2023.arxiv.org/abs/2310.14189
  20. [20]Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024.arxiv.org/abs/2410.11081
  21. [21]Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Hu, Enze Xie, Zhangyang Wang, Ping Luo, and Yujun Shi. Sana-sprint: One-step diffusion with continuous-time consistency distillation, 2025. URL https://arxiv.org/abs/2503.09641.
  22. [22]Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations, 2022. URL https://arxiv.org/abs/2108.01073.
  23. [23]Jiaqi Xu, Wenbo Li, Haoze Sun, Fan Li, Zhixin Wang, Long Peng, Jingjing Ren, Haoran Yang, Xiaowei Hu, Renjing Pei, et al. Fast image super-resolution via consistency rectified flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11755–11765, 2025.
  24. [24]Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in neural information processing systems, 35:23593–23606, 2022.arxiv.org/abs/2201.11793
  25. [25]Martin Schiødt, Nikolaj T Mücke, and Clara M Velte. Generative super-resolution of turbulent flows via stochastic interpolants. Scientific Reports, 2026.
  26. [26]Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. URL https://arxiv.org/abs/2010.02502.
  27. [27]Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19721–19730, 2025.arxiv.org/abs/2412.08629
  28. [28]Aymeric Delefosse, Anastase Charantonis, and Dominique Beréziat. Super-resolving coarse-resolution weather forecasts with flow matching. arXiv preprint arXiv:2604.00897, 2026.arxiv.org/abs/2604.00897
  29. [29]Arnela Hadzic, Franz Thaler, Lea Bogensperger, Simon Johannes Joham, and Martin Urschler. Restora-flow: Mask-guided image restoration with flow matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4943–4952, 2026.arxiv.org/abs/2511.20152
  30. [30]Forouzan Fallah, Wenwen Li, Chia-Yu Hsu, Hyunho Lee, and Yezhou Yang. Rareflow: Physics-aware flow-matching for cross-sensor super-resolution of rare-earth features. arXiv preprint arXiv:2510.23816, 2025.
  31. [31]Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024.arxiv.org/abs/2410.12557
  32. [32]Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613–6623, 2024.arxiv.org/abs/2311.18828
  33. [33]Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems, 37:47455–47487, 2024.arxiv.org/abs/2405.14867
  34. [34]Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023.arxiv.org/abs/2310.02279
  35. [35]Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024.arxiv.org/abs/2407.02398
  36. [36]Zhou Wang, Alan Bovik, Hamid Sheikh, Student Member, and Eero Simoncelli. Image quality assessment: From error measurement to structural similarity. IEEE Trans. Image Process., 13, 11 2003.
  37. [37]Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research, 22(4):730–751, 2025. ISSN 2731-5398. doi: 10.1007/s11633-025-1562-4. URL http://dx.doi.org/10.1007/s11633-025-1562-4.
  38. [38]Ernst Hairer, Syvert Norsett, and Gerhard Wanner. Solving Ordinary Differential Equations I: Nonstiff Problems, volume 8. 01 1993. ISBN 978-3-540-56670-0. doi: 10.1007/978-3-540-78862-1.DOI
  39. [39]Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2019.arxiv.org/abs/1907.10902
  40. [40]Xiao Xue, Hua-Dong Yao, and Lars Davidson. Synthetic turbulence generator for lattice Boltzmann method at the interface between RANS and LES. Physics of Fluids, 34(5):055118, 2022.DOI
  41. [41]Anthony Zhou, Zijie Li, Michael Schneier, John R Buchanan Jr, and Amir Barati Farimani. Text2pde: Latent diffusion models for accessible physics simulation. arXiv preprint arXiv:2410.01153, 2024.arxiv.org/abs/2410.01153
  42. [42]Jayesh K Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized pde modeling. arXiv preprint arXiv:2209.15616, 2022.arxiv.org/abs/2209.15616
  43. [43]Philip Hall and Nils Thuerey. ϕ_now: Differentiable simulations for pytorch, tensorflow and jax. In Forty-first international conference on machine learning, 2024.
  44. [44]Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24174–24184, 2024.arxiv.org/abs/2312.02696

Paper details

vlnc:W_649nge55
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents