acceptodds
Under review as a conference paper at ICLR 2027

LyapuFlow: Controlling Generative Flows with Lyapunov Feedback for Inverse Problems

Abstract

Pretrained flow models are now widely used as generative priors in science and vision, where inference-time guidance enables test-time constraints without retraining. Existing methods use projection, posterior sampling, or iterative optimization of the generative trajectory. We propose LyapuFlow, an alternative based on Lyapunov feedback control. At each sampling step, LyapuFlow predicts the terminal sample towards which the current flow is evolving, and evaluates the constraint violation on this prediction. Then, we compute the minimum-norm control that satisfies a prescribed Lyapunov decrease condition. The resulting control remains inactive when the uncontrolled dynamics already reduce the constraint violation at the prescribed rate. Otherwise, it provides a corrective update within a flow-relative trust region that prevents the control from dominating the pretrained dynamics. We demonstrate LyapuFlow in both data and latent spaces, outperforming alternatives spanning different mechanisms for test-time constraint enforcement in scientific machine learning and image inverse problems.

Primary area:
generative models
Keywords:
generative flow models, posterior guidance, Lyapunov control, constrained generation, inverse problems, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_H28oOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

LyapuFlow is a training-free method for guiding pretrained flow models using constraints evaluated at predicted endpoints. Its full controller computes a minimum-norm correction from a Lyapunov decrease condition, followed by velocity-relative clipping. The latent implementation uses a stop-gradient approximation and optionally restores full derivatives later. Experiments on four PDE tasks and image super-resolution and deblurring assess constraint satisfaction, sample fidelity, and computational cost.

Critical strengths

1.Clear formulation and practical relevance.

The paper clearly explains how endpoint predictions and a Lyapunov decrease condition yield adaptive feedback. Applying new observations or physical constraints without retraining the generative model is practically valuable.

2.Meaningful empirical support.

Experiments cover different PDE dynamics and multiple image restoration settings. Baseline comparisons, component ablations, and runtime measurements provide useful evidence supporting the complete sampler under the tested configurations.

Critical weaknesses

1.The contribution needs clearer positioning relative to prior work.

The control law follows a classical minimum-norm construction, so the contribution depends on its adaptation to generative sampling. MPC-Flow is particularly relevant because it also uses endpoint-based objectives for inference-time control of flow models. Comparing objectives, assumptions, derivative requirements, and computational costs would help qualify the introduction’s research-gap claim.

2.The benefits of the feedback rule are not fully isolated.

Baselines differ in correction location, scaling, derivatives, and clipping. During latent stop-gradient sampling, the controller becomes loss-normalized gradient guidance and no longer uses nominal-flow progress to gate intervention. As acknowledged by the authors, it does not generally guarantee decrease of the true Lyapunov function. A matched endpoint-guidance baseline would clarify its specific contribution. PDE errors before final projection and image measurement residuals would also strengthen the constraint-satisfaction evidence.

Questions for authors

1.How does LyapuFlow differ from prior endpoint-based control methods, particularly MPC-Flow, in its objective, assumptions, and computational requirements?

2.Could the authors provide a matched endpoint-guidance comparison and report PDE constraint errors before final projection and measurement residuals for images?

3.Can LyapuFlow be applied to other latent flow models, such as FLUX? What assumptions would be required, and would the stop-gradient approximation or gradient-switching schedule need to be adapted? Are results on another backbone available?

5 comments
AuthorsOpenReview ↗

Response to Reviewer H28o (1/3)

We thank the reviewer for the constructive and detailed feedback.

MPC baselines (Q1 and W1)

We agree that the minimum-norm control construction is classical. Our contribution is to apply this construction to the constraint violation of a predicted terminal sample, allowing us to derive a closed-form feedback rule for pretrained generative flows. The following table compares related methods by their correction target, velocity-model derivatives, and optimization requirements.

MethodKey approachCorrection targetBackprop through vθv_{\theta}Optimization loop
LyapuFlowLyapunov control with VV on x^1∣t\hat{x}_{1\mid t}xtx_tX (if stop-gradient), O (else)X
SAD-FlowerLyapunov control with VV on xtx_txtx_tXX
MPC-FlowReceding-horizon control with soft endpoint costxtx_tX (K=1), O (K>1, Δt)O (inner)
HardFlowReceding-horizon control with hard endpoint constraintx^1\hat{x}_1XO (inner)
OC-FlowFull-trajectory optimal controlx[0,1]x_{[0,1]}OO (outer)

LyapuFlow evaluates the constraint violation on a predicted terminal sample and computes the minimum-norm correction from a prescribed Lyapunov decrease condition. The correction is zero when the nominal flow already satisfies this condition. MPC-Flow instead optimizes a soft terminal objective and control effort at each sampling step, while HardFlow uses constrained trajectory optimization to enforce terminal feasibility.

LyapuFlow requires a differentiable constraint loss and an endpoint prediction, but does not require an optimization loop. Its full controller differentiates through the velocity model, while the stop-gradient approximation avoids this additional cost. We will clarify these distinctions and the corresponding computational requirements in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer H28o (2/3)

Feedback rule and constraint satisfaction (W2 and Q2)

(1) Matched endpoint-guidance ablation. The correction location and feedback rule are coupled: moving the same correction directly to the predicted endpoint would no longer implement the derived Lyapunov decrease condition. Endpoint-based methods are already included among our baselines. To isolate the contribution of the Lyapunov-based feedback gain, we ran an additional ablation comparing LyapuFlow against constant-coefficient gradient correction. Both methods use the same predicted terminal sample, constraint gradient, correction location, and trust-region clipping. The only difference is the feedback coefficient. Writing gt=∇ztVtg_t=\nabla_{z_t}V_t, the unbounded controls are

MethodUpdate rule
LyapuFlowut=−αVt∥∥gt∥∥22gtu_t=-\dfrac{\alpha V_t}{\|\|g_t\|\|_2^2}g_t
Constant-coefficientut=−β∥∥gt∥∥22gtu_t=-\dfrac{\beta}{\|\|g_t\|\|_2^2}g_t

We evaluated both methods with all-step stopgrad on the same five AFHQ dog images for Gaussian deblurring. All experiments use SD3.5-Medium, CFG 2, 28 NFEs, seed 0, measurement noise σn=0.03\sigma_n=0.03, and trust-region ratio ctr=1c_{\mathrm{tr}}=1.

MethodPSNR ↑SSIM ↑LPIPS ↓
LyapuFlow (α=5\alpha=5)28.9270.78810.1262
LyapuFlow (α=10\alpha=10)29.0070.79650.1220
LyapuFlow (α=15\alpha=15)29.0130.79520.1359
Constant-coefficient (β=1000\beta=1000)28.1820.75530.1178
Constant-coefficient (β=2000\beta=2000)28.7730.78860.1388
Constant-coefficient (β=3000\beta=3000)28.7960.78970.1388
Constant-coefficient (β=5000\beta=5000)28.7950.78980.1386
Constant-coefficient (β=10000\beta=10000)28.7960.78980.1386

LyapuFlow achieves higher PSNR than all tested constant-coefficient settings and the best SSIM at α=10\alpha=10. The constant-coefficient method obtains the lowest LPIPS at β=1000\beta=1000, but with lower PSNR and SSIM. Although this is a small ablation, these results suggest that adapting the correction magnitude to the current Lyapunov energy provides an advantage over the tested fixed-coefficient alternatives, even when velocity gradients are omitted.

(2) Image measurement residuals. We also study measurement consistency during sampling. The newly added control profiles in Appendix C.5 compare full gradients and stopgrad using the same image and measurement. The following table shows the results for one FFHQ motion-deblurring example.

Velocity gradientResidual norm, first → lastPSNR ↑SSIM ↑LPIPS ↓
Full gradients761.99 → 45.04 (−94.09%)21.740.4500.218
stopgrad⁡(vθ)\operatorname{stopgrad}(v_\theta)761.99 → 44.91 (−94.11%)27.860.7250.117

Both settings reduce the measurement residual by approximately 94% in this example. Despite omitting velocity gradients, stopgrad achieves comparable residual reduction and better reconstruction quality. This supports the practical effectiveness of the approximation. However, it does not provide the same formal decrease guarantee as the full controller.

(3) PDE constraint errors before final projection. We added a before-and-after projection comparison for all methods in Appendix B.6. The table below summarizes LyapuFlow's constraint errors across the four PDE tasks. Each entry reports before → after final projection.

TaskCE(IC) / 10−210^{-2}CE(CL/NCL) / 10−210^{-2}
Heat2.031 → 0.0005.040 → 0.001
Burgers3.253 → 0.0002.323 → 0.005
RD1.051 → 0.0001.442 → 0.000
NS15.597 → 0.0005.045 → 0.000

These results show that LyapuFlow already achieves low constraint errors before final projection. Projection further reduces the remaining violations. Importantly, as discussed in our response to Reviewer mfTc, LyapuFlow also achieves the lowest mean-field error across all four PDE tasks before projection, so its improvement in sample fidelity cannot be attributed to the final projection steps alone.

AuthorsOpenReview ↗

Response to Reviewer H28o (3/3)

Applicability to other latent flow models (Q3)

LyapuFlow can be applied to other latent flow models provided that the model supplies a velocity prediction and the constraint loss is differentiable through the decoder and measurement operator. The stop-gradient approximation does not require differentiating through the velocity network, while the gradient-switching schedule and feedback parameters may need adjustment for a different backbone.

To evaluate this, we ran additional experiments using FLUX.2, following the setup of MPC-Flow. We use 50 sampling steps and CFG scale 4, evaluating five images across three inverse-problem settings using 3 seeds (15 runs per method). LyapuFlow uses α=10\alpha=10, trust ratio 1, and stopgrad throughout. MPC-Flow uses MPC-RHC (K=1K=1) with 20 inner steps and learning rate 0.5, while FlowChef uses learning rate 0.5 and one inner step.

MethodPSNR (dB) ↑SSIM ↑Seconds/image ↓
LyapuFlow25.815 ± 2.5500.7572 ± 0.061126.97
MPC-Flow25.129 ± 1.8350.7307 ± 0.0691124.13
FlowChef23.061 ± 0.8570.6817 ± 0.073227.63

LyapuFlow achieves the highest mean PSNR and SSIM in this evaluation, while running approximately more than 4 times faster than MPC-Flow. It also outperforms FlowChef in reconstruction quality at similar runtime. These results provide initial evidence that LyapuFlow can be applied beyond the SD3.5-Medium backbone without changing its feedback formulation.

Reviewer_mfTcOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

LyapuFlow proposes a training-free guidance method for pretrained flow-based generative models that uses Lyapunov feedback control to enforce constraints during inference. The motivation is that existing guidance approaches, such as Diffusion Posterior Sampling (DPS), often require carefully tuned guidance strengths. Excessive guidance can improve constraint satisfaction while pushing generated samples away from the learned data distribution. LyapuFlow addresses this by calculating the minimum-norm correction needed to ensure a prescribed decrease in constraint violation.

Critical strengths

LypauFlow considers the constraint satisfaction problem and derives through the total derivative of the constraint energy, a minimum-norm feedback control problem under a Lyapunov decrease condition. The paper demonstrates substantial improvements over several established baselines, including DPS, FlowDPS, PCFM, and OC-Flow, across physics-constrained generation and image inverse problems. The authors also provide practical implementations through trust region clipping and discretization, and characterize their effects.

Critical weaknesses

Though LypauFlow gives guarantees exponential decrease in constraint violations, these guarantees still remain quite weak in terms of terminal constraint satisfaction. Consequently, achieving exact constraint satisfaction may depend on the underlying generative dynamics providing additional reductions beyond those enforced by the controller. Beyond the hyperparameter sensitivity/tuning justification, the motivations to use LypauFlow over standard guidance methods remains a bit weak. Specifically, the choice of exponential decay rate may fail to exploit opportunities for stronger interventions that could achieve comparable terminal constraint satisfaction with less cumulative deviation from the pretrained flow.

Questions for authors

How often is the Lyapunov decrease condition actually satisfied in practice? Since trust-region clipping and approximate gradients can violate the condition, it would be informative to measure the frequency and magnitude of these violations throughout sampling.

How much of the improvement comes from the Lyapunov formulation rather than the final projection steps? Can we see experiments that demonstrate results with/without the projection steps?

6 comments
AuthorsOpenReview ↗

Response to Reviewer mfTc

We thank the reviewer for the constructive and detailed feedback.

Control activity and terminal reduction (W and Q1)

To study the behavior of LyapuFlow during sampling, we added control profiles in Appendices B.5 and C.5, showing the Lyapunov energy, its exponential decay reference, and the raw and bounded controls. We also evaluated 10 trajectories per PDE task. The table below reports the fraction of inactive and trust-region-clipped steps, along with the median Lyapunov energy at selected flow times.

TaskInactive controlTrust-region clippedV0V_0V0.33V_{0.33}V0.66V_{0.66}V0.99V_{0.99}
Heat42.8%0.0%12.90000.01840.00190.0005
Burgers36.0%3.3%2.48000.01670.00600.0011
RD43.1%0.1%6.26000.01970.01260.0005
NS0.0%0.1%266.00006.09001.16000.0580

For Heat, Burgers, and RD, the controller remains inactive on 36.0–43.1% of sampling steps, indicating that the pretrained flow already satisfies the prescribed Lyapunov decrease condition. For NS, the controller remains active throughout sampling. These activation rates are measured along the controlled sampling trajectories and reflect how often the nominal flow violates the decrease condition at the states reached under feedback. Trust-region clipping occurs on at most 3.3% of steps. Importantly, LyapuFlow substantially reduces the constraint energy before final projection. At t=0.99t=0.99, the median Lyapunov energy is less than 0.05% of its initial median value across all four PDE tasks. While the prescribed decay condition does not guarantee exact terminal constraint satisfaction, these results show that the controller achieves substantial constraint reduction during sampling.

Contribution of final projection (Q2)

To separate the contribution of LyapuFlow from the final projection, we evaluated all methods before and after projection across the four PDE tasks. The table below summarizes the mean-field errors of LyapuFlow and the best competing method, together with the constraint errors of LyapuFlow. Each entry reports before → after final projection.

TaskLyapuFlow MMSE/CMMSE / 10⁻²Best competing MMSE/CMMSE / 10⁻²CE(IC) / 10⁻²CE(CL) / CE(NCL) / 10⁻²
Heat0.009 → 0.012PCFM (0.117 → 0.117)2.031 → 0.0005.040 → 0.001
Burgers0.161 → 0.163DPS (0.252 → 0.254)3.253 → 0.0002.323 → 0.005
RD0.019 → 0.019OC-Flow (0.085 → 0.083)1.051 → 0.0001.442 → 0.000
NS4.700 → 4.695DPS (5.205 → 5.200)15.597 → 0.0005.045 → 0.000

Burgers uses three final projection steps, while the other tasks use one. For PCFM, the internal projection at the last sampling step is included in this count. LyapuFlow achieves the lowest MMSE/CMMSE across all four tasks even before final projection, and this advantage is preserved afterward. Final projection further reduces the remaining constraint errors, but does not account for the improvements in mean-field accuracy. We have included the full before-and-after comparison for all methods in the revised appendix.

We thank the author for their thoughtful response. Would the authors also please address the point that the choice of exponential decay rate may fail to exploit opportunities for stronger interventions that could achieve comparable terminal constraint satisfaction with less cumulative deviation from the pretrained flow, clarifying why LyapuFlow should be used over standard guidance schemes.

AuthorsOpenReview ↗

Response to Follow-up Comment (1/2)

We thank the reviewer for reading our response and for continuing the discussion.

Our motivation for using exponential decay is to reduce constraint violations throughout sampling while preserving the pretrained generative dynamics as much as possible. LyapuFlow computes the minimum-norm control needed to satisfy the prescribed decrease condition at each step. When the pretrained flow already satisfies this condition, no correction is applied.

This minimizes control locally, rather than over the entire sampling trajectory. A different control schedule could potentially achieve the same terminal constraint satisfaction with less cumulative intervention, for example by applying stronger corrections early and fewer corrections later. Finding such a schedule would require optimizing control over the full trajectory, which is a more challenging problem. LyapuFlow instead provides a closed-form feedback rule that adapts to the current state of the flow.

To investigate whether more aggressive feedback can reduce cumulative control effort while achieving comparable terminal constraint satisfaction, we compared LyapuFlow with an aggressive configuration (α=200\alpha=200, without the trust region) and standard DPS guidance. We also report sample fidelity to assess how the different interventions affect the pretrained generative dynamics.

1. Exponential decay versus stronger interventions

We evaluate 64 multi-condition samples with 100 sampling steps, all before final projection. Under the controlled dynamics dxt=(vθ(xt,t)+ut) dtdx_t=(v_\theta(x_t,t)+u_t)\,dt, we measure cumulative control as Ucum=∑kΔt ∥utk∥2U_{\mathrm{cum}}=\sum_k\Delta t\,\|u_{t_k}\|_2. We also report mean-field error (MMSE/CMMSE) and terminal constraint error CE=∥h(x1)∥2\mathrm{CE}=\|h(x_1)\|_2.

TaskUcumU_{\mathrm{cum}} (LyapuFlow / Aggressive / DPS)MMSE/CMMSE /10−2/10^{-2} (LyapuFlow / Aggressive / DPS)CE /10−2/10^{-2} (LyapuFlow / Aggressive / DPS)
Heat14.70 / 33.90 / 23.070.040 / 0.971 / 0.5655.95 / 4.74 / 392.95
Burgers11.53 / 47.12 / 19.501.130 / 1.468 / 1.2065.62 / 4.53 / 69.65
RD7.20 / 27.05 / 20.380.048 / 0.249 / 0.1132.17 / 2.62 / 110.31
NS150.08 / 419.24 / 38.618.517 / 19.727 / 9.46016.74 / 19.90 / 11.46

The aggressive configuration and stronger interventions indeed require more cumulative control and produce worse mean-field accuracy across all four tasks. For Heat and Burgers, it achieves slightly lower terminal constraint errors, but at much greater control effort and reduced fidelity. For RD and NS, it also produces larger constraint errors. Thus, stronger feedback does not necessarily reduce the total intervention required during sampling.

One possible explanation is that large corrections push the sample away from the learned flow trajectory, making additional corrections necessary at later steps. Our finite-step analysis in Proposition 1 also shows that larger controlled velocities increase the discretization error, even when the instantaneous Lyapunov decrease condition is satisfied. This motivates the use of a moderate decay rate together with a flow-relative trust region to limit excessive corrections, particularly when early terminal predictions are less accurate.

Reviewer_SS9rOpenReview ↗
Rating
4(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces inference-time controller for imposing constraints on pre-trained generative flow models. It evaluates a quadratic constraint loss on a first-order prediction of the terminal sample and derives a closed-form, minimum-norm correction satisfying a instantaneous Lyapunov decrease condition.

Critical strengths
  1. A control-theoretic perspective. I particularly appreciate the idea of leveraging control theory, and specifically Lyapunov feedback, to enforce task-specific constraints on pre-trained generative flows at inference time.

  2. A practical feedback construction. Applying the Lyapunov decrease condition to a predicted terminal sample provides an interpretable and adaptive guidance rule. With full derivatives, the controller intervenes only when the nominal dynamics fail to achieve the prescribed decrease. Moreover, its closed-form solution avoids iterative trajectory optimisation and per-step constraint-Jacobian linear solves, making the approach computationally attractive.

  3. Convincing empirical results across diverse applications. LyapuFlow achieves the lowest mean-field error across all four PDE benchmarks and the best LPIPS across all nine main image inverse-problem comparisons. These improvements persist at 40 function evaluations (Tables 1, 2, and 12), providing encouraging evidence of the method's effectiveness across substantially different problem settings.

Critical weaknesses
  1. The experiments do not sufficiently isolate the contribution of the proposed Lyapunov feedback mechanism. Equation (22) shows that stopping gradients through the velocity model reduces the controller to normalised gradient guidance, removing the nominal-flow-dependent activation mechanism. This approximation is used throughout the super-resolution experiments, while Figure 13 suggests that full-gradient guidance can perform worse. Consequently, it remains unclear whether the reported improvements arise from the proposed Lyapunov formulation or from adaptive gradient normalisation and control clipping. This distinction is important because the Lyapunov feedback mechanism constitutes the paper's central methodological contribution.

  2. The evaluation omits closely related control-based approaches. In particular, MPC-Flow (Webber et al., ICML 2026) directly addresses flow-based image inverse problems through computationally efficient receding-horizon control, while HardFlow (Li et al., TPAMI 2026) considers terminal constraints through trajectory optimisation. Neither is discussed or evaluated.

Questions for authors
  1. Contribution of the Lyapunov feedback mechanism. Equation (22) shows that stopping gradients through the velocity model reduces the proposed controller to normalised gradient guidance, removing the nominal-flow-dependent activation mechanism. Since this approximation is used in several image experiments, including all super-resolution settings, it is unclear how much of the reported improvement can be attributed specifically to Lyapunov feedback rather than adaptive gradient normalisation. Could the authors provide an ablation comparing the full controller against this simplified alternative under otherwise identical settings, particularly for the deblurring tasks? This would clarify how effectively the proposed feedback enforces data consistency and whether the reported improvements primarily arise from constraint satisfaction or preservation of the generative prior. How frequently is the controller actually inactive because the pre-trained flow already satisfies the prescribed Lyapunov decrease condition?

  2. Comparison with MPC-based approaches and computational efficiency. Could the authors provide a direct experimental comparison with MPC-Flow (Webber et al., 2026) on the image inverse problems considered in this paper, using the same datasets, and comparable computational and hyperparameter-tuning budgets? Beyond reconstruction quality (PSNR, SSIM, LPIPS), it would be valuable to compare task-specific inference runtime and peak GPU memory. Could the authors discuss how LyapuFlow's terminal-prediction-based feedback differs from the constrained trajectory optimisation proposed in HardFlow (Li et al., 2025)? These comparisons would help establish the distinction and practical advantages of LyapuFlow relative to existing MPC-based approaches.

References:

  • Webber, G., Denker, A., Barbano, R., & Reader, A. J. (2026). Solving Inverse Problems with Flow-based Models via Model Predictive Control. ICML 2026. arXiv:2601.23231.
  • Li, Z., Alim, K., & Azizan, N. (2025). HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization. arXiv:2511.08425.
4 comments
AuthorsOpenReview ↗

Response to Reviewer SS9r (1/2)

We thank the reviewer for the constructive and detailed feedback.

Contribution of Lyapunov feedback (W1 and Q1)

We agree that stopping gradients through the velocity model simplifies the feedback rule in Eq. (22). In this setting, the nominal-flow-dependent activation condition cancels, and the controller reduces to gradient guidance with a gain determined by the current Lyapunov value. In contrast, with full gradients, the controller accounts for the nominal flow and intervenes only when the prescribed Lyapunov decrease is violated. With stopgrad, it retains the adaptive gain derived from the Lyapunov decrease condition but no longer has this flow-dependent activation mechanism. Figure 13 already compares full velocity gradients against different stop-gradient schedules on ten FFHQ images, including Gaussian and motion deblurring. Full gradients do not consistently give the best reconstruction quality. To further study this, we ran an additional ablation study.

(1) Full gradients versus stopgrad. The control profiles in Appendix C.5 compare both gradient settings using the same image and measurement. The table below reports a representative FFHQ motion-deblurring trajectory.

Velocity gradientResidual norm, first → lastPSNR ↑SSIM ↑LPIPS ↓
Full gradients761.99 → 45.04 (−94.09%)21.740.4500.218
stopgrad⁡(vθ)\operatorname{stopgrad}(v_\theta)761.99 → 44.91 (−94.11%)27.860.7250.117

Both settings reduce the measurement residual by approximately 94% in this example, while stopgrad yields better reconstruction quality. The other paired image examples in Appendix C.5 show a similar pattern: both settings substantially reduce the residual, while stopgrad consistently achieves better reconstruction quality.

(2) Ablation. We compare LyapuFlow against a constant-coefficient gradient correction under identical all-step stopgrad settings. Both methods follow the negative gradient of the constraint energy, but differ in how the correction magnitude is determined. Writing gt=∇ztVtg_t=\nabla_{z_t}V_t, the unbounded controls are

MethodUpdate rule
LyapuFlowut=−αVt∥∥gt∥∥22gtu_t=-\dfrac{\alpha V_t}{\|\|g_t\|\|_2^2}g_t
Constant-coefficientut=−β∥∥gt∥∥22gtu_t=-\dfrac{\beta}{\|\|g_t\|\|_2^2}g_t

Both methods use the same trust-region clipping with ctr=1c_{\mathrm{tr}}=1. We evaluated three decay rates α∈{5,10,15}\alpha\in\{5,10,15\} and five constant coefficients β∈{1000,2000,3000,5000,10000}\beta\in\{1000,2000,3000,5000,10000\}. The experiments use the same five AFHQ dog images for Gaussian deblurring, SD3.5-Medium, CFG 2, 28 NFEs, seed 0, and measurement noise σn=0.03\sigma_n=0.03.

MethodPSNR ↑SSIM ↑FID ↓LPIPS ↓
LyapuFlow (α=5\alpha=5)28.9270.788176.2990.1262
LyapuFlow (α=10\alpha=10)29.0070.796573.2370.1220
LyapuFlow (α=15\alpha=15)29.0130.795274.2370.1359
Constant-coefficient (β=1000\beta=1000)28.1820.755378.7000.1178
Constant-coefficient (β=2000\beta=2000)28.7730.788689.8220.1388
Constant-coefficient (β=3000\beta=3000)28.7960.789789.8640.1388
Constant-coefficient (β=5000\beta=5000)28.7950.789889.6450.1386
Constant-coefficient (β=10000\beta=10000)28.7960.789889.7280.1386

The results show that LyapuFlow performs better overall than the tested constant-coefficient guidance, whose performance saturates at β=3000\beta=3000 and higher. Even when stopgrad removes the velocity-gradient contribution to the activation check, the controller retains Lyapunov meaning through the prescribed decay condition and the coefficient based on αVt\alpha V_t, yielding meaningful reconstruction gains over a fixed coefficient.

(3) Activation. We also study how frequently the controller remains inactive or reaches the trust-region bound. The following table shows these fractions across the evaluated trajectories. We provide control profiles (for both the PDE benchmarks and image tasks) in the revised Appendices B.5 and C.5.

TaskInactive controlTrust-region clipped
Heat42.8%0.0%
Burgers36.0%3.3%
RD43.1%0.1%
NS0.0%0.1%
AFHQ Gaussian deblur (stopgrad)0.0%3.6%
AFHQ Motion deblur (stopgrad)0.0%3.6%

These results show that the pretrained flow often already satisfies the prescribed Lyapunov decrease condition, allowing the controller to leave the nominal dynamics unchanged. In the four representative full-gradient image trajectories in Appendix C.5, control is inactive at one of 28 steps (3.6%) in each; with stopgrad, it remains active while the constraint energy is nonzero. Trust-region clipping is infrequent across these settings and mainly limits unusually large corrections.

AuthorsOpenReview ↗

Response to Reviewer SS9r (2/2)

MPC baselines (W2 and Q2)

We thank the reviewer for pointing out MPC-Flow and HardFlow. To address this concern, we ran additional experiments with MPC-Flow and clarified the methodological differences in the revised paper.

(1) Comparison with MPC-Flow. We test MPC-Flow's MPC-RHC (K=1K=1) method on our image inverse-problem setting. We use the same SD3.5-Medium backbone, degradation operators, measurement noise, 28-NFE sampling schedule, and hyperparameter-search budget as LyapuFlow. MPC-Flow uses 10 inner Adam steps with learning rate 0.3 and penalty 2.5×10−72.5\times10^{-7}. LyapuFlow uses one closed-form correction per sampling step with α=5\alpha=5, ctr=1c_{\mathrm{tr}}=1, and stopgrad throughout. The table below shows reconstruction quality across 1k AFHQ images per task. Runtime is averaged over three images per task, and peak GPU memory is the maximum allocation across those images. Further results across datasets are provided in Table 1 of the revised manuscript.

TaskMethodPSNR ↑SSIM ↑FID ↓LPIPS ↓Seconds/image ↓Peak GPU memory (GiB)
SR ×4LyapuFlow30.050.8299.680.0458.511.24
MPC-Flow28.550.80517.100.11452.911.24
FlowDPS28.240.74616.050.12218.911.24
FlowChef29.430.79827.300.0958.411.24
Gaussian deblurLyapuFlow27.240.73320.030.1488.511.26
MPC-Flow26.380.73123.520.18553.411.28
FlowDPS25.410.68026.870.24419.311.27
FlowChef25.220.71740.810.2748.711.26
Motion deblurLyapuFlow27.180.74619.170.1579.411.25
MPC-Flow26.470.74424.130.19960.411.27
FlowDPS25.950.69322.780.21219.511.26
FlowChef26.330.73940.830.2308.611.25

LyapuFlow achieves the best LPIPS across all three tasks and the best PSNR on both deblurring tasks. Compared with MPC-Flow, it runs approximately 6 times faster, while using similar GPU memory. FlowChef is slightly faster than LyapuFlow, while FlowDPS requires roughly twice the runtime. These results show that LyapuFlow achieves strong reconstruction quality without the additional cost of iterative optimization at each sampling step.

(2) Distinction from MPC-Flow and HardFlow. The following table summarizes how each method uses endpoint information and computes its correction.

MethodKey approachCorrection targetBackprop through vθv_{\theta}Optimization loop
LyapuFlowLyapunov control with VV on x^1∣t\hat{x}_{1\mid t}xtx_tX (stopgrad), O (full)X
MPC-FlowReceding-horizon control with soft endpoint costxtx_tX (K=1K=1), O (K>1K>1, Δt\Delta t)O (inner)
HardFlowReceding-horizon control with hard endpoint constraintx^1\hat{x}_1XO (inner)

LyapuFlow and HardFlow both use a prediction of the terminal state to assess endpoint constraints. LyapuFlow turns the predicted constraint violation into a Lyapunov energy and imposes an instantaneous decay condition by correcting the current velocity. When feasible, this minimum-norm correction is closed-form. HardFlow instead solves a constrained receding-horizon subproblem to find a feasible predicted terminal state, then uses its surrogate trajectory to determine the next sampling state. MPC-Flow uses terminal prediction to optimize a soft task cost and control effort at each sampling step.

Thank you for the detailed response and the additional provided experiments and ablations. I enjoyed reading the work. As a minor note, I would add Distinction from MPC-Flow and HardFlow in the main manuscript.

We thank the reviewer for their positive feedback and for taking the time to read our response and additional experiments. We will add a discussion distinguishing LyapuFlow from MPC-Flow and HardFlow to the main manuscript, as suggested.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. We have uploaded a revised manuscript with additional experiments and ablations. Below, we summarize our responses and the main updates:

  • Ablation study. A matched ablation compares LyapuFlow's Lyapunov-based correction with fixed-coefficient gradient guidance under the same stop-gradient and trust-region settings.
  • MPC baselines (Table 1). We expanded the image comparison to include MPC-Flow and clarified how our method differs from MPC-Flow, HardFlow, and related methods.
  • Control profile analysis (Appendices B.5 and C.5). New control profiles show when feedback is inactive or clipped and how constraint violations decrease over sampling.
  • Final projection analysis (Appendix B.6). PDE results before and after final projection show that LyapuFlow's advantage is present before projection.
  • Transfer to FLUX.2. Preliminary FLUX.2 results on a small image set support the method's applicability beyond SD3.5-Medium.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.