- Rating
- 3(+2)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces FINO, an inference-time optimization framework for constrained inverse design using pretrained diffusion and flow models. FINO optimizes the initial noise by alternating between reward optimization and constraint-restoration steps, aiming to satisfy design constraints without manually balancing them against the reward. The method is evaluated on photonic inverse design and molecular property optimization.
Critical strengths
-
The problem is relevant to scientific inverse design, where optimizing target properties while respecting physical or chemical constraints is essential.
-
The method is relatively simple, applicable to pretrained diffusion and flow models, and requires no additional training.
-
The photonics experiments demonstrate promising results, particularly for out-of-distribution target properties. The use of independent FDTD simulations for evaluation strengthens these results.
Critical weaknesses
-
Limited methodological novelty: Initial noise optimization for reward-guided generation is well established, and separating objective optimization from constraint satisfaction is a classical idea in constrained optimization. The main contribution appears to be the particular combination of feasibility restoration, perturbation-based candidate selection, and reward optimization. It is unclear whether this constitutes a sufficiently substantial methodological advancement.
-
Missing relevant literature: Two particularly relevant papers are not discussed: Constrained Flow Optimization via Sequential Fine-Tuning for Molecular Design (Gutjahr et al., ICML 2026) and Source-Guided Flow Matching (Wang et al., ICLR 2026). These works overlap substantially with FINO's motivation and methodological ingredients. A clearer comparison is necessary to establish the novelty.
-
Reliance on differentiable rewards and constraints: The proposed method requires differentiable reward and constraint functions, which considerably limits its applicability to realistic scientific inverse-design problems. Many relevant properties and constraints are only accessible through non-differentiable simulations, expensive numerical evaluations, or experiments. Consequently, FINO often requires differentiable surrogate predictors, introducing additional modeling effort and potentially substantial approximation errors. This is especially concerning for the claimed out-of-distribution design capabilities, where surrogate reliability is inherently uncertain. The molecular experiments already demonstrate this limitation. Despite achieving zero constraint violations according to the surrogate, 12.8–26.6% of generated molecules violate the actual RDKit synthetic-accessibility constraint.
-
Insufficient baseline comparisons and ablations: The baselines use a common weighted reward–constraint objective, with weights selected based on InitNO, raising concerns about whether they are appropriately tuned for constraint satisfaction. Comparisons against standard constrained optimization procedures in the initial-noise space would be particularly informative. Furthermore, the contributions of individual components, especially perturbation-based candidate selection and feasibility restoration, are not adequately isolated.
-
Computational efficiency: The procedure requires repeated generator evaluations and gradient calculations, including additional rollouts for perturbation-based candidate selection. A comparison under matched computational budgets, including inference-time costs, would be valuable.
Questions for authors
-
How does FINO compare to standard constrained optimization algorithms applied directly to the initial-noise space, particularly under matched computational budgets?
-
How much of the improvement arises from separating reward and constraint updates, versus perturbation-based candidate selection?
-
How reliable are the differentiable surrogate predictors for out-of-distribution designs? Can the authors provide independent evaluations of the optimized molecular properties?
8 comments
W2: Missing relevant literature
A: We agree that both papers share important motivations with FINO, and incorporate them into the revised related work section to clarify our methodological contributions.
First, Constrained Flow Optimization (CFO) (Gutjahr et al., 2026) addresses a closely related problem of maximizing rewards while satisfying constraints, further highlighting the importance of multi-objective problem, also the target problem in this work. However, CFO approaches this problem through sequential fine-tuning of pretrained generative models, requiring task-specific model adaptation. In contrast, FINO operates entirely at inference time by optimizing the initial noise without updating the pretrained model parameters. Furthermore, while CFO is experimentally evaluated on synthetic benchmarks and molecular design, FINO demonstrates its applicability to both molecular generation and topological photonic inverse design.
Second, Source-Guided Flow Matching (SGFM) (Wang et al., 2026) shares our motivation of performing guidance through the source distribution while preserving the pretrained flow dynamics. SGFM provides a general framework for sampling from a modified source distribution, including optimization-based sampling methods closely related to existing initial noise optimization approaches such as D-Flow(Ben-Hamu et al. 2024). However, our methodological contribution goes beyond the standard approaches of initial noise optimization as shown in all these previous works. Rather, FINO specifically addresses constrained reward optimization by factorizing reward maximization and constraint satisfaction into separate updates, incorporating feasibility restoration and perturbation-based candidate selection to maintain constraint satisfaction while improving target-property accuracy.
We have cited both works in the related works of revised manuscript.
W4: Insufficient baseline comparisons and ablations
A: We thank the reviewer for raising concerns regarding baseline tuning and comparison fairness. Our original intention in using a shared weight combination was to highlight the practical difficulty of balancing reward maximization and constraint satisfaction through weighted objectives. However, we agree that method-specific hyperparameter tuning is necessary for a more comprehensive comparison. We therefore conducted additional weight sweeps and reran the experiments using independently selected baseline configurations.
First, we investigated the sensitivity of the fabrication constraint weight. We performed hyperparameter searches at a target bandgap of 0 meV, generating 20 designs for each configuration. As weight searching for 3 objectives costs we fix and and sweep through only. Initially, we observed that classifier guidance (CG) required a substantially larger fabrication constraint weight to reduce constraint violations, as shown below.
Table R2. Fabrication constraint weight sweep of classifier guidance on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%).
| 1 | 10 | ||||||
|---|---|---|---|---|---|---|---|
| CG | 0.32/54 | 0.28/54 | 0.33/44 | 0.34/14 | 0.54/8 | 0.68/4 | 0.71/4 |
Here, and are fixed. Increasing from 1 to reduces the violation rate from 54% to 4%, illustrating the sensitivity of constraint satisfaction to weight selection.
To improve the numerical scaling of the fabrication constraint, we reformulated its evaluator using the sum of pixel values with a threshold of 16, instead of the mean pixel value with a threshold of 0.25. These two formulations define the same feasibility condition for an patch, but differ in gradient magnitude. Using the rescaled evaluator, we heuristically search for the and combination first and then sweep through
Table R3. Method-specific fabrication constraint weight sweep for BoN, CG, and SMC on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%).
| 1 | 5 | 10 | 50 | 100 | 500 | |
|---|---|---|---|---|---|---|
| BoN | 2.24/0 | 2.93/0 | 3.46/0 | 3.46/0 | 3.46/0 | 3.46/0 |
| CG | 0.26/35 | 0.35/10 | 0.44/5 | 0.40/0 | 0.38/0 | — |
| SMC | 1.69/45 | 1.80/40 | 1.86/35 | 2.95/15 | 3.33/5 | 2.96/5 |
The results demonstrate that different baseline methods exhibit different trade-offs between target-property accuracy and constraint satisfaction as the fabrication constraint weight increases. To prioritize constraint satisfaction, we selected the smallest fabrication weight at which the violation rate reached its minimum observed value and showed no further improvement with larger weights in our search. Based on this criterion, we selected for BoN, for CG, and for SMC.
For ALM, we independently investigated the number of inner iterations and the frequency of constraint projections during diffusion sampling.
Table R4. ALM hyperparameter sweep over inner iterations and constraint projections on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%)
| Iterations / # of Projections | 1000 | 100 | 10 |
|---|---|---|---|
| 1 | 0.35 / 0 | 0.34 / 0 | 0.34 / 10 |
| 2 | 0.21 / 0 | 0.23 / 0 | 0.41 / 0 |
| 3 | 0.28 / 0 | 0.31 / 0 | 0.39 / 0 |
We selected two inner iterations and ten equally spaced projection steps for the subsequent experiments.
For InitNO, we repeated the weight sweep using the updated fabrication constraint calculation.
Table R5. InitNO weight sweep across reward and constraint coefficients on target bandgap value of 0meV. Each entry reports MAE / constraint violation rate (%).
| (1,1) | (1,5) | (1,10) | (5,1) | (5,5) | (5,10) | (10,1) | (10,5) | (10,10) | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 50.61/0 | 36.30/5 | 47.76/0 | 13.48/25 | 24.23/10 | 19.97/0 | 6.96/50 | 7.99/30 | 16.16/25 |
| 5 | 30.38/10 | 50.39/0 | 27.60/0 | 13.68/5 | 13.51/30 | 27.74/5 | 4.18/50 | 0.41/40 | 22.45/35 |
| 10 | 35.87/5 | 41.67/0 | 42.15/0 | 9.31/15 | 17.88/15 | 12.29/20 | 11.73/50 | 12.01/15 | 5.20/25 |
We selected for InitNO.
Q1 Comparison across standard constrained optimization algorithms
A: We thank the reviewer for suggesting a comparison with standard constrained optimization methods applied directly to the initial-noise space.
To address this concern, we implemented an ALM-based optimization method in the initial-noise space. Similar to our CG+ALM baseline, the method first performs a reward update, followed by constraint projection using ALM. We conducted additional experiments by varying the reward-update parameter η and the number of projection steps while fixing the inner and outer iteration counts on target bandgap value of 198meV.
Table R1. InitNO+ALM result on different experiment settings. Each entry reports MAE / constraint violation rate (%).
The results demonstrate a trade-off between target-property accuracy and constraint satisfaction. Increasing η substantially improves MAE but generally increases constraint violations. Conversely, increasing the number of projection steps can improve constraint satisfaction at the expense of target-property accuracy.
These results suggest that directly applying ALM-based constraint projection in the initial-noise space does not necessarily achieve both high target-property accuracy and reliable constraint satisfaction. This further motivates FINO's feasibility-aware optimization strategy.
W1: Limited methodological novelty
A: First, we agree that initial noise optimization for reward-guided generation is well established ([1-2]). However, as demonstrated in InitNO results, directly extending initial noise optimization by combining the reward and constraints into a weighted objective does not reliably achieve both target-property accuracy and constraint satisfaction.
Second, we acknowledge that separating objective optimization from constraint satisfaction is a classical idea in constrained optimization. However, directly applying conventional constraint optimization methods to diffusion-based generation presents additional challenges. Applying constraint enforcement within the reverse diffusion trajectory modifies the pretrained sampling process, while applying conventional constrained optimization directly in the initial-noise space does not necessarily resolve the trade-off between reward maximization and constraint satisfaction. As demonstrated in Table 1 of the main paper, directly applying ALM-based projection in the initial-noise space improves feasibility under certain configurations but still exhibits substantial constraint violations or large target-property errors.
Our contribution is to bridge these two approaches by introducing factorized optimization in the initial-noise space, enabling reward maximization and constraint restoration to be handled separately without modifying the pretrained generative model or its sampling trajectory. Specifically, FINO integrates feasibility restoration, perturbation-based candidate selection, and reward optimization into a unified inference-time framework. Our experiments demonstrate that this formulation achieves competitive target-property accuracy while consistently satisfying the constraints across different scientific inverse-design tasks.
[1] Eyring et al. (2024). "ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimization." NeurIPS 2024. [2] Tang et al. (2025). "Inference-Time Alignment of Diffusion Models with Direct Noise Optimization." ICML 2025.