acceptodds
Under review as a conference paper at ICLR 2027

ConicFlow: Constrained Inference-time Optimal Control for Flow Matching Models

38 views

Abstract

We propose ConicFlow, which frames constrained sampling from a flow matching model at inference time as an optimal control problem that balances deviation from the pretrained trajectory against reward, subject to hard terminal constraints. We discretize this problem and adapt receding-horizon control, optimizing only a short window of the trajectory at each step for computational tractability. We address the resulting sequence of receding-horizon subproblems with a receding-horizon primal–dual algorithm. Its dual multipliers capture the difficulty of satisfying each constraint and are refined throughout sampling, steering generation toward feasible final samples. We further propose a terminal-estimate variant that avoids accumulating local truncation errors by evaluating the reward and constraints at the predicted terminal point rather than at the end of the current horizon. Under standard regularity assumptions, we establish finite-horizon guarantees on suboptimality. Under convex-affine structure, we derive a sharper bound that scales linearly with omitted horizon length, and we show that the terminal-estimate variant is provably tighter when the flow dynamics change slowly. Experiments in both low- and high-dimensional settings with a recent large-scale flow matching model show that ConicFlow achieves high reward and strong constraint satisfaction while deviating less from the pretrained flow than recent baselines.

Primary area:
generative models
Keywords:
flow matching, optimal control, generative models, constrained optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MfB8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ConicFlow, an inference-time optimal control framework for guiding pretrained flow matching models toward high-reward samples while satisfying hard terminal constraints. The method introduces two receding-horizon strategies, RH-H and RH-T, and employs primal-dual optimization to balance reward maximization, constraint satisfaction, and control energy. The authors provide theoretical analyses of discretization, horizon approximation, and convergence, along with experiments on 2D synthetic data and FLUX.1-dev image editing.

Critical strengths
  1. The paper addresses an important problem in controllable generation by jointly considering reward optimization, hard terminal constraints, and minimal intervention in pretrained flow matching models.
  2. The proposed receding-horizon framework provides a structured approach to inference-time control, while the RH-T strategy leverages terminal flow-map predictions to incorporate information beyond the truncated optimization horizon.
  3. The experiments demonstrate improvements in reward–control energy trade-offs and report high feasibility rates under the specified constraints, suggesting the potential of optimal control for guiding pretrained generative models.
Critical weaknesses
  1. The theoretical guarantees rely on strong assumptions that are not established in practice:
    • Assumption 2 requires affine flow dynamics, whereas FLUX.1-dev employs a nonlinear neural velocity field.
    • Assumption 3 requires concave rewards and convex constraints, but these properties are not guaranteed for CLIP Score, ImageReward, or PickScore-based constraints.
      While such assumptions do not necessarily invalidate the method, more extensive experiments and failure-case analyses are needed to demonstrate its practical effectiveness. The paper devotes substantial attention to idealized theoretical guarantees but provides comparatively limited empirical validation.
  2. Much of the analysis builds on established techniques from numerical optimal control, model predictive control, and convex optimization. The paper would benefit from deeper theoretical insights specific to the proposed framework, such as:
    • Does RH-T retain its advantages under nonlinear flow dynamics?
    • How do learned flow-map approximation errors affect feasibility?
    • Does convergence hold under inexact primal optimization?
    • How does constraint feasibility affect dual convergence?
  3. High inference-time computational cost. Backpropagation through a large generative model such as FLUX is computationally expensive. ConicFlow requires over five minutes per image, approximately 7–14× slower than HardFlow, limiting its practical scalability.
  4. The experiments focus on generated-image editing without evaluating real-image editing. The authors could extend the framework through image inversion and compare against FlowEdit [1], FlowAlign [2], and ChordEdit [3] on MagicBrush [4] and PIE-Bench [5], considering editing accuracy, content preservation, and computational cost.
  5. Limited experimental scale. The unconstrained image-editing experiments evaluate only four prompt–seed pairs per reward, which is insufficient to establish robust performance across diverse prompts and generation settings.

Suggestion: Improve the balance between theory and experiments. Consider moving detailed theoretical derivations to the appendix, retaining concise theorem statements and key insights in the main paper, and allocating more space to empirical validation.

[1] https://arxiv.org/abs/2412.08629
[2] https://arxiv.org/abs/2505.23145
[3] https://arxiv.org/abs/2602.19083
[4] https://arxiv.org/abs/2306.10012
[5] https://arxiv.org/abs/2310.01506

Questions for authors
  1. Can the authors provide independent evaluation metrics to distinguish genuine improvements in generation quality and semantic constraint satisfaction from potential reward overoptimization?
  2. Could the authors consider additional applications, such as image restoration with comparisons to gradient-based methods (e.g., PnP-Flow [1], FlowDPS [2]), or robotic control tasks, where the proposed framework may offer practical advantages?

[1] https://arxiv.org/abs/2410.02423
[2] https://arxiv.org/abs/2503.08136

3 comments
AuthorsOpenReview ↗

Response to Reviewer MfB8

We thank the reviewer for recognizing the paper's strengths. We expanded our results to over 3× the original experimental instances, harder constraints, additional baselines, and independent evaluation. All new results appear on rebuttal p. 10 (Figs. 4–6), with protocols and examples at the beginning of Appendix A. Tables R1–R5 refer to our general response to all reviewers and the AC.


W1/W5. Experimental scale and difficult constraints

We expanded to 54 constrained prompt pairs, jointly imposing PickScore, CLIP, and ImageReward constraints at three increasing thresholds (Table R1; Fig. 5). ConicFlow's advantage persists across this substantially larger and harder evaluation: higher joint feasibility, lower violation of each constraint, and 4.5×–54.5× less mean control energy than HardFlow. Appendix A.1 reports lower violation in all 162 paired comparisons. At the hardest threshold ConicFlow's feasibility drops, but it keeps lower violation at far less energy.

We further evaluate 24 unconstrained PIE-Bench prompt pairs against five baselines (Table R2; Fig. 6), with ConicFlow leading in both optimized reward and independent evaluation by Skywork-VL-Reward-7B. The trends from the original submission results thus persist at a greater scale.

Q1. Independent verification of edit quality

Every edit in both studies is now evaluated by Skywork-VL-Reward-7B, which none of the methods optimizes. ConicFlow achieves the highest independent Skywork gain at every constrained threshold (Table R1) and among all unconstrained baselines (Table R2) supporting gains beyond the optimized metrics.


W2. Theory and nonlinear implementation

We agree the guarantees should not be read as certificates for FLUX. Lemma 1 permits nonlinear dynamics under regularity and repairability; sharper horizon and primal–dual bounds assume convex-affine structure and exact primal responses. The analysis offers terminal-feasibility-specific insights: Lemmas 2–3 separate RH-H's nominal-drift dependence from RH-T's perturbation amplification; Slater-based repair quantifies dependence on the feasibility margin; and replanning bounds terminal violation from the actual final-window state, without an explicit accumulated trajectory-amplification factor.

Importantly, the new nonlinear FLUX horizon ablation (Table R4; Fig. 4b) tests the design insight directly. RH-T wins 94/96 and 91/96 equal-penalty comparisons to RH-H at H=1,2H=1,2, with the gap closing near the full horizon. This supports RH-T empirically outside the sharper assumptions.

The practical method uses 24 Adam steps, one dual update per window, detached drift gradients, and a learned terminal map. Under additional local strong convexity, primal suboptimality δj\delta_j gives response error at most 2δj/σj\sqrt{2\delta_j/\sigma_j} and a Lipschitz constraint-error bound; terminal-map error similarly affects constraint evaluation. These conditional bounds do not certify detached nonlinear Adam. Our new constrained optimizer ablation (Table R5; Fig. 4c) shows constraint violation falling markedly through 24 steps, achieving 6/6 feasible pairs, with much lower energy than HardFlow on the same problem.


W3. Inference-time computational cost

We agree runtime matters. A new matched wall-clock study (Table R3; Fig. 4a) compares methods at ~35, 120, and 230 s. ConicFlow leads at every budget: relative reward is 0.64 versus HardFlow's 0.34 at ~35 s, and 0.81 versus 0.40 at ~120 s. Thus its advantage persists at matched compute. ConicFlow, and to a lesser extent FMRG, improves with more budget, while HardFlow stays flat. In all the expanded experiments the runtime of ConicFlow never exceeded ~300 seconds and could be cut in half with taking a small hit in performance (See Table R5).


W4/Q2. Recommended baselines and additional applications

We appreciate the suggested references and have added a paragraph discussing them in Related Work. We also expanded the baselines to include the gradient-based approaches gradient guidance, FlowChef, and FlowDPS, alongside HardFlow and FMRG (Table R2), addressing the request for stronger comparisons with guidance methods. ConicFlow leads all five in optimized reward and independent Skywork gain. Real-image inversion approaches target a different source-preservation protocol from our generated-image, terminal-constraint setting; Adapting ConicFlow to these settings is a future direction we are exploring.

We will also foreground the theoretical assumptions and main insights, moving derivations to the appendix and bolstering the empirical section with the new and expanded results we have provided for the rebuttal. We hope the larger evaluation, independent judging, new baselines, and compute-controlled comparisons address the reviewer's key concerns.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts in addressing my concerns. I will adjust my score.

We thank the reviewer for reading and responding to our rebuttal. We hope that our responses have addressed the main concerns and would appreciate it if the reviewer could reconsider their assessment accordingly.

Reviewer_X8GtOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ConicFlow, an inference-time optimal control framework for steering pretrained flow matching models toward high-reward outputs subject to terminal constraints. It penalizes control energy to limit deviations from the pretrained dynamics and uses receding-horizon optimization to avoid optimizing the entire sampling trajectory at once. A terminal-prediction variant evaluates rewards and constraints at the predicted final sample rather than at the end of the short planning window. The method alternates primal optimization with projected dual updates and transfers constraint multipliers across sampling steps.

Critical strengths
  • A coherent formulation of constrained generation. The optimal control objective clearly separates reward maximization, terminal feasibility, and intervention cost. The receding-horizon formulation offers a practical way to optimize multiple future transitions while limiting the size of each optimization problem.

  • Useful terminal prediction and persistent constraint multipliers. Evaluating the objective at a predicted terminal sample addresses the mismatch between noisy intermediate states and final-output requirements. Carrying dual multipliers across windows is also a sensible mechanism for adapting the relative importance of constraints throughout sampling. The experiments and ablations provide encouraging evidence, including comparisons at approximately matched wall-clock time in Appendix D.2.

Critical weaknesses
  • The experimental evidence is too limited to support broad conclusions about constrained image generation. Table 1 averages over only four prompt–seed instances per reward, and the constrained evaluation uses a small set of manually selected editing prompts. These experiments demonstrate promising behavior on specific examples, but provide limited evidence of robustness across more difficult combinations of constraints.

  • The connection between the theoretical guarantees and the implemented algorithm is insufficiently established. The convergence analysis assumes affine dynamics, concave rewards, convex constraints, and exact primal responses at each dual iteration. The image experiments instead involve nonlinear neural dynamics and scoring functions, approximate primal optimization with a finite number of Adam steps, and detached drift terms within the planning window. Terminal prediction also differs across experimental settings, using either a rollout or a learned single-call flow map. These approximations are understandable computational choices, but the current guarantees do not directly cover their combined effect. In particular, finite dual-iteration bounds should not be interpreted as accounting for an inexact primal solve or the modified gradients used in practice. The paper should clearly distinguish the idealized method covered by the analysis from the implemented method, and quantify how primal optimization error, gradient detachment, and terminal-prediction error affect final constraint satisfaction.

  • Minor issues The experimental count at L2281–2283 needs clarification: nine instances, two seeds, and five thresholds imply 90 comparisons, whereas 86 are reported. Finally, please revisit the stability explanation at L2272–2274: a contraction factor of zero does not explain divergence under the stated recurrence.

  • Typo :“pretraned flow matching models”(line 478)

Questions for authors

See weakness 1 and 2.

2 comments
AuthorsOpenReview ↗

Response to Reviewer X8Gt

We sincerely thank the reviewer for the careful assessment and for highlighting our coherent constrained-control formulation, terminal prediction, persistent multipliers, and matched-compute evidence. The two concerns motivated substantial additions. All new results are on the additional rebuttal page (p. 10; Figs. 4–6), with instances and protocols at the beginning of Appendix A. Tables R1–R5 refer to our general response to all reviewers and the AC.


W1. Experimental breadth and challenging constraints

Our expanded evaluation now includes more than 3× the instances of the original submission, testing whether the observed trends persist across broader edits and harder constraints.

The new constrained study covers 54 source-edit prompt pairs from the PIE-Bench benchmark[1] that represent edits of varying difficulty, jointly enforcing PickScore, CLIP, and ImageReward at three difficulty thresholds (Table R1; Fig. 5). ConicFlow consistently surpasses HardFlow in joint feasibility at every threshold, while requiring 4.5×–54.5× less control energy and yielding lower violations of all three constraint metrics. Notably, the advantage persists as constraints become harder, although absolute feasibility decreases somewhat at the strictest threshold. This strengthens the evidence beyond the original examples.

We also evaluate 24 unconstrained PIE-Bench prompt pairs against five baselines (Table R2; Fig. 6), where ConicFlow leads in reward and the reward–energy trade-off. Furthermore, every edit in both studies is independently evaluated by Skywork-VL-Reward-7B, which no method optimizes. ConicFlow yields the highest Skywork gain in all reported comparisons, providing independent, model-based support for the observed gains rather than relying only on the optimized scores. We distinguish this evaluation from human verification of semantic correctness.


W2. Theory–implementation correspondence

We have clarified the distinction between analysis and implementation. Our analysis characterizes how discretization, horizon truncation, and replanning affect reward and terminal feasibility, not a convergence guarantee for nonlinear FLUX. Lemma 1 accommodates nonlinear dynamics under regularity and repairability, while the stronger bounds assume convex-affine structure and exact primal responses. More specifically, Lemmas 2–3 reveal distinct horizon-error mechanisms: RH-H depends on nominal drift, whereas RH-T depends on perturbation amplification. The replanning analysis bounds terminal feasibility using the actual reached state in the last window, without an explicit accumulated trajectory-amplification factor. These results identify when terminal prediction is advantageous and which assumptions support that conclusion.

In practice, constrained FLUX employs 24 Adam steps, one dual update per window, detached drift gradients, and a learned single-call terminal map. We explicitly distinguish theoretical dual iterations with exact primal responses from practical Adam steps. Under additional local strong convexity, a primal suboptimality gap δj\delta_j entails response error at most 2δj/σj\sqrt{2\delta_j/\sigma_j} and a corresponding Lipschitz constraint-error bound; terminal-prediction error likewise affects constraint evaluation through its Lipschitz constant. These conditional relations are not certificates for detached nonlinear optimization.

Crucially, we have now completed optimization-sensitivity experiments. In the constrained Adam-step ablation (Table R5; Fig. 4c), constrant violation on decreases markedly through 24 steps, where ConicFlow achieves 6/6 feasible pairs, with substantially less energy compared to HardFlow. Furthermore, Table R4 shows the behavior the analysis predicts: RH-T at a short horizon already matches its full-horizon performance, whereas RH-H only approaches it as the horizon grows. These ablations quantify practical sensitivity, though detachment bias and learned-map error are not individually isolated; we will make that limitation explicit in the text.


W3. Minor corrections

We thank the reviewer for these careful observations. The 86 comparisons are 8×5×2=808\times5\times2=80 plus 1×3×2=61\times3\times2=6 for the three-constraint case. We have clarified the count, corrected the stability discussion, and fixed the typo.


We hope the expanded results, independent quality evaluation, and completed ablations address the reviewer's central concerns.

[1] https://arxiv.org/abs/2310.01506

Reviewer_X8GtOpenReview ↗

Response

Thank you for the detailed response and additional experiments. The expanded PIE-Bench evaluation and independent reward assessment strengthen the empirical evidence, while the clarification of the theoretical scope and optimization ablations substantially address my main concerns. I am raising my score.

Reviewer_UwpSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ConicFlow, an inference-time constrained optimal control framework for flow matching models that balances terminal reward maximization, hard constraint satisfaction, and deviation from pretrained generative dynamics. To reduce the computational cost of full-trajectory optimization, the authors introduce a receding-horizon primal-dual algorithm that optimizes short trajectory segments while updating constraint multipliers across sampling steps. A key contribution is the terminal-prediction variant (RH-T), which uses a pretrained flow map to evaluate rewards and constraints at the predicted final state rather than at the end of the planning horizon. The paper establishes bounds on discretization error, horizon truncation error, and primal-dual convergence under regularity and convex-affine assumptions. Experiments on synthetic data and FLUX-based image editing demonstrate improved reward–control trade-offs and constraint satisfaction compared with existing methods.

Critical strengths
  1. The optimal control framework provides a unified treatment of reward optimization, hard terminal constraints, and regularization toward pretrained flow dynamics.

  2. The RH-T formulation uses flow maps to preserve the terminal objective while optimizing only a short horizon. The associated analysis provides an intuitive explanation for why terminal prediction can reduce horizon approximation error.

  3. ConicFlow demonstrates improved reward–control trade-offs and high empirical feasibility on image editing tasks. The additional ablations on planning horizon, computational budget, and terminal prediction help support the proposed design.

Critical weaknesses

No critical weakness that support rejection.

Minor weakess include the assumptions and increased computations, but these are understandable.

Questions for authors
  • The main theoretical guarantees require affine flow dynamics and convex terminal objectives, whereas practical flow matching models and PickScore-based constraints are highly nonlinear and nonconvex. Can the authors clarify which theoretical insights extend to the practical setting? In particular, under what conditions does RH-T retain its advantage over RH-H for nonlinear flows?

  • Theoretical novelty and guarantees: The convergence analysis appears to build largely on standard MPC and convex primal-dual optimization results. Could the authors clarify the main technical challenges and novel theoretical insights beyond existing analyses? Additionally, how informative are the horizon-dependent bounds when using short planning horizons such as H=2H=2 and N=8N=8?

  • The theoretical analysis assumes exact primal minimization, whereas the implementation uses a finite number of Adam steps and approximate gradients. How sensitive are constraint satisfaction and solution quality to these approximations? Furthermore, since feasibility is measured using PickScore thresholds, could the authors evaluate semantic constraint satisfaction using independent metrics or human evaluation?

2 comments
AuthorsOpenReview ↗

Response to Reviewer UwpS

We thank the reviewer for recognizing ConicFlow's unified formulation, RH-T's terminal prediction, and its reward–energy and feasibility results, and for finding no rejection-level weakness. All updated experimental results are on the additional rebuttal page (p. 10); experimental instances and details are at the beginning of Appendix A. Tables R1–R5 refer to our general response to all reviewers and the AC.


Q1. Nonlinear flows and when RH-T is preferable

We begin by highlighting that Lemma 1 permits nonlinear dynamics and nonconvex terminal objectives under regularity, feasibility, and repairability; later bounds require stronger assumptions. RH-T truncates optimized control, not terminal evaluation: it evaluates reward and constraints after predicted pretrained continuation, whereas RH-H evaluates the intermediate endpoint. The nonlinear-compatible nesting argument (Appendix C.4, Step 1) preserves the terminal state when zero controls are appended under exact continuation. The benefit requires accurate continuation, bounded amplification, and feasible repair.

The horizon ablation (Table R4, (page 10, Fig. 4b)) isolates this choice: RH-T and RH-H differ only in where the reward is evaluated. RH-T's advantage is largest at short horizons, where the window endpoint is farthest from the final sample, and vanishes as H→NH\to N, where the two objectives coincide. On nonlinear FLUX with which the ablation was done, RH-T also reaches higher reward with lower energy, so the benefit extends beyond the affine-convex setting.


Q2. Theoretical novelty and short-horizon bounds

While MPC and primal–dual optimization are established, our analysis addresses the effect of horizon truncation on terminal reward and feasibility in constrained generative dynamics. Specifically, Lemmas 2–3 distinguish the error mechanisms of RH-H, governed by nominal drift MM, and RH-T, governed by perturbation amplification κA\kappa_A. Remark 1 shows that κA≤2\kappa_A\leq\sqrt{2} suffices for a tighter RH-T bound. Appendix C.4 combines zero-control nesting with Slater-based repair to preserve terminal feasibility after truncation. The constant Gh=Lr+Lh(J‾D−J‾D)/γG_h=L_r+L_h(\overline J_D-\underline J_D)/\gamma makes dependence on the feasibility margin explicit. Lemmas 9–10 and Theorem 1 further account for errors across successive replanning steps. Since the final window starts at the actual reached state, terminal violation is bounded by LhεN−1L_h\varepsilon_{N-1}, without an explicit accumulated trajectory-amplification factor. This provides feasibility-aware analysis for terminally constrained generation, rather than new MPC updates.

These bounds give sufficient regimes, not general guarantees for the nonlinear case. Still, Table R4 shows the behavior the analysis predicts: RH-T at a short horizon already matches its full-horizon performance, whereas RH-H only approaches it as the horizon grows.


Q3. Approximate optimization and semantic feasibility

The analysis assumes exact primal responses, whereas constrained FLUX uses 24 Adam steps, one dual update per window, detached drift gradients, and a learned single-call terminal map. Under local σj\sigma_j-strong convexity, a primal gap δj\delta_j implies solution error at most 2δj/σj\sqrt{2\delta_j/\sigma_j} and additional violation at most Lg,j2δj/σjL_{g,j}\sqrt{2\delta_j/\sigma_j}. This does not certify nonlinear Adam, and more steps may not remove detachment bias.

Empirically, we observe that constraint satisfaction is sensitive to the number of adam optimization steps for the inner loop. In the constrained Adam-steps ablation (Table R5; page 10, Fig. 4c), constraint violation decreases steadily with more inner steps and vanishes at our default of 24, with energy well below tuned HardFlow. This suggests that the inner loop needs enough optimization steps to get sufficiently close to exact primal minimization.

We agree with the reviewer that reward-defined feasibility is not independent semantic verification. The rewards defining the constraints are the ones being optimized. We therefore add an independent judge, Skywork-VL-Reward-7B, a vision-language reward model that is trained to judge whether a given prompt is a good descriptor of an image. We use it to score whether the target prompt correctly describes each edited image and report the gain over the source unedited image. No method optimizes or queries it, and it is typically a stronger judge of semantic edits than PickScore, CLIP, or ImageReward, so it tests whether reward gains reflect real edits. ConicFlow has the highest Skywork gain in every comparison, in both the 54-pair constrained (Table R1; page 10, Fig. 5) and 24-pair unconstrained (Table R2) studies.


We hope these results resolve the reviewer's remaining concerns.

Thank the author for the detailed reply. All my concerns have been resolved. I will keep my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of new results

We thank the reviewers for their constructive feedback. We kindly ask reviewers to see p. 10 of the revised PDF for all new figures and results, and the newly added Appendix A for qualitative comparisons. We have expanded the constrained and unconstrained evaluations, added a strong independent judge that no method optimizes (Skywork-VL-Reward-7B), and added ablations on compute fairness, terminal prediction, and optimizer sensitivity.


1. Constrained editing: 54 prompt pairs

All 54 pairs are from PIE-Bench: 3–4× the prompt instances of our original submission. We jointly enforce PickScore, CLIP, and ImageReward constraints. Each threshold lies a fraction α of the way from the source image's score to a fixed ceiling, so larger α is harder. Feasibility is the fraction satisfying all three.

Table R1. Constraint feasibility, energy, and independent quality.

αMethodFeasible ↑Energy ↓Skywork ↑
0.25ConicFlow98.1%7.7×10²0.33
HardFlow18.5%3.5×10³0.07
0.50ConicFlow94.4%1.9×10³0.54
HardFlow3.7%1.2×10⁴0.33
0.75ConicFlow50.0%3.3×10³0.59
HardFlow3.7%1.8×10⁵0.49

ConicFlow improves feasibility and independent quality at every threshold with 4.5×, 6.3×, and 54.5× lower energy, respectively. Fig. 5 also shows lower individual PickScore, CLIP, and ImageReward violations.


2. Unconstrained editing: 24 prompt pairs

All 24 pairs are from PIE-Bench: 3–4× the prompt instances of our original submission.

Table R2. Mean ± standard error (higher is better).

MethodReward gain ↑Skywork gain ↑
ConicFlow2.66 ± 0.180.36 ± 0.08
HardFlow1.53 ± 0.200.11 ± 0.05
FMRG1.38 ± 0.170.11 ± 0.05
Gradient guidance0.81 ± 0.140.06 ± 0.05
FlowChef0.89 ± 0.160.05 ± 0.03
FlowDPS0.70 ± 0.130.03 ± 0.02

ConicFlow improves mean reward by ~74% over HardFlow, the strongest baseline, and achieves over 3× its independent Skywork gain, whereas competing methods generally need much higher energy for comparable gain (Fig. 6).


3. A new, strong, independent judge of edit quality: Skywork-VL-Reward-7B

We re-judge every edit in both studies with Skywork-VL-Reward-7B, a vision-language reward model that no method optimizes and that is typically a stronger judge of semantic edits than the rewards themselves. ConicFlow has the highest Skywork gain in every comparison (Tables R1, R2), so its gains are confirmed by a judge that cannot be gamed.


4. Matched-compute and design ablations

Table R3. Matched compute (8 portrait PickScore edits; Fig. 4a). Reward gain is normalized by default ConicFlow; time is median wall-clock per run.

MethodStepsTimeRelative gain ↑
ConicFlow337s0.64
FMRG333s0.54
HardFlow4035s0.34
ConicFlow12122s0.81
FMRG13126s0.76
HardFlow167120s0.40
ConicFlow24233s1.00
FMRG25232s0.77
HardFlow336230s0.35

ConicFlow leads at all three matched runtime levels, supporting that its advantage is not due to larger compute budgets.

Table R4. Terminal rollout (RH-T) vs. window-end objective (RH-H), 16 problems (Fig. 4b). Median normalized gain; “ahead” counts reward wins at equal λ over 96 operating points.

HRH-T ↑RH-H ↑RH-T ahead
10.850.5494/96
21.000.6791/96
41.020.8070/96
60.960.8855/96
80.920.9253/96

RH-T is most beneficial at short horizons (94/96 and 91/96 wins at H=1,2). As expected, both variants converge at H=N=8.

Table R5. Adam steps per control decision (6 constrained prompt pairs, α=0.5; Fig. 4c). Time is wall-clock per edit.

MethodStepsViolation ↓Feasible ↑Energy ↓Time
ConicFlow30.1240/676448s
ConicFlow60.0393/6173182s
ConicFlow120.0145/62061152s
ConicFlow240.0006/61715289s
ConicFlow480.0006/61327566s
HardFlow400.0630/68916129s

From 24 steps on, ConicFlow has zero violation and satisfies all constraints on all six pairs, with ~5.2× less energy than tuned HardFlow. More steps bring no further gain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.