acceptodds
Under review as a conference paper at ICLR 2027

FlowBender: Feedback-Aware Training for Self-Correcting Conditional Flows

Abstract

Conditional diffusion and flow models routinely fail to satisfy the very constraints that define their task. For instance, a depth-conditioned model often produces images whose re-extracted depth disagrees with the input, even though the forward operator—the depth predictor defining the constraint—is available during both training and inference. Existing approaches generally fall into two categories: supervised models that treat the conditioning signal as a static cue and ignore alignment information at inference, and guidance-based methods that consult it through hand-tuned linear updates, typically trading fidelity to the condition against the plausibility of the generated sample. We argue that the fundamental gap in both paradigms is that the model is never trained to utilize its own alignment error. We introduce FlowBender, a closed-loop framework that treats this error as a first-class input, training the network to learn a correction policy conditioned on inference-time feedback. At each step, an unguided look-ahead pass estimates the clean signal, a task-specific deviation is computed via the forward operator, and a refinement pass consumes this signal to produce a corrected velocity. We propose several variants of FlowBender, including a gradient-based formulation for differentiable operators and a zero-order variant for non-differentiable settings such as JPEG compression. For efficient sampling, we introduce a prior-step shortcut that enables closed-loop correction at a minimal additional computational cost. Across image-to-image translation, restoration, and 3D mesh texturing, FlowBender consistently outperforms standard supervised baselines, alignment-loss-augmented training, and state-of-the-art inference-time guidance, improving fidelity and plausibility simultaneously rather than trading them against each other.

Primary area:
generative models
Keywords:
flow matching, diffusion models, generative computer vision, guidance

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_KrMzOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FlowBender, a feedback-aware framework for conditional flow models. A shared network first predicts a clean-sample estimate, from which a forward operator computes an alignment-error signal. A second pass conditions on this signal to predict a refined velocity. The framework supports gradient-based feedback and derivative-free measurement residuals, while retaining the standard flow-matching training objective. An optional prior-step shortcut reuses cached predictions to reduce denoiser evaluations. The method is evaluated on super-resolution, depth/edge-conditioned generation, JPEG restoration, and 3D mesh texturing, demonstrating favorable measurement-consistency and plausibility trade-offs against supervised and inference-time guidance baselines.

Critical strengths
  1. Naturally motivated, clearly presented, and well supported experimentally. The paper identifies an intuitive limitation of conditional generation: models may violate constraints even when the forward operator is available to quantify the mismatch. The proposed feedback-aware formulation follows naturally from this observation, making the motivation and method easy to follow. The extensive evaluation across image and 3D tasks, together with informative ablations and compute-budget controls, provides substantial support for the approach. This coherence between motivation, methodology, and empirical evidence makes the central claims convincing.
  2. A flexible and practically relevant approach to learned correction. FlowBender learns how to incorporate alignment feedback into velocity prediction while retaining the standard flow-matching objective. Its support for both gradient-based and derivative-free feedback makes the framework applicable to different operator types, including non-differentiable JPEG compression. The integration with both ControlNet and LoRA further demonstrates its compatibility with existing generative models. This flexibility gives the contribution relevance beyond a single task or a particular feedback formulation.
Critical weaknesses

None, please see questions.

Questions for authors
  1. Clarification — Task-specific configuration recommendations. Based on the experiments, what configurations would the authors recommend for different tasks, particularly regarding the feedback variant and shortcut threshold? Have the authors observed any consistent patterns that could guide these choices, for example depending on the operator type or task characteristics?
  2. Jointly conditioning on the initial prediction and its error. Table 3 compares self-conditioning with error-based feedback. Have the authors considered providing both the initial clean-sample estimate x^1\hat{x}_1 and its feedback signal sts_t to the refinement pass? Since the feedback may discard information from the initial prediction, could these two inputs provide complementary information?
  3. Optional exploration — Multiple feedback-refinement rounds. It would be interesting to explore whether recomputing feedback from the refined prediction and performing additional refinement passes within the same timestep offers further benefits. This differs from the multi-step look-ahead studied in Table 15. Since FlowBender already performs feedback-based correction throughout sampling, I view this as a potential extension rather than a necessary addition to the current submission.
3 comments

We thank the reviewer for the encouraging review. We are glad that the motivation was clear, and that the experiments, ablations, and compute-matched controls were found to support the approach convincingly. We also appreciate the recognition of the framework’s flexibility.

1. Recommended configurations. When applicable, we generally recommend the zero-order variant: it is cheaper, often close to the best variant, and the only option for non-differentiable operators. The shortcut retains much of the gain at roughly half the denoiser cost, and should be the go-to in time-constrained settings. For the most reliable performance, we recommend full two-pass sampling with the combined (x̂₁) variant, which is the most consistent across tasks. We will add this discussion to the final version.

2. Jointly conditioning on x̂₁ and s_t. Thank you for the suggestion. This is an appealing variant, as x̂₁ is already computed in the look-ahead pass and adds negligible cost, and we tested it as requested. On SR, providing both inputs performs on par with s_t (zero-order variant) alone (differences within seed variation), suggesting the feedback already carries the relevant information. We will add this result to the appendix.

We provide x̂₁ through the latent input slot (used for first-order feedback) and s_t through the zero-order slot, both from the same look-ahead pass, using the zero-order training protocol:

SRPSNR vs. LR input ↑FID ↓
Standard FT34.353.93
x̂₁ only (self-conditioning)33.454.28
s_t only (zero-order)39.253.36
s_t + x̂₁39.833.46

For reference, in additional multi-seed runs (3 seeds), the zero-order variant varies by ±0.88 dB PSNR and ±0.11 FID across seeds.

3. Multiple feedback-refinement rounds. We agree this is a promising direction and consider it for future work.

Thank you again for the constructive suggestions.

The rebuttal has addressed my concern. Overall, I find the paper straightforward and easy to follow, with notable performance improvements. The additional experiments suggest that there is still room for further improvement, but I do not think this detracts from the overall quality of the work. I will maintain my positive recommendation.

Thank you for taking the time to read our rebuttal and for your continued support of the paper.

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

The paper proposes FlowBender, a training scheme for conditional flow matching models where the network receives its own alignment errors as additional input. At each sampling step, the lookahead channel with null feedback produces a clean estimate x̂1. Calculate the task-specific feedback signal s_t based on H(x̂1) and y. The second refinement pass then consumes s_t and outputs the velocity used for integration.

Three feedback types are studied: first order, gradient with respect to measurement loss. x_t or x̂1; zeroth order, requiring no differentiable measurement space residuals; and combinations of the two. Training uses standard CFM regression loss with stopping gradient on s_t and feedback dropout (p_un = 0.1). At inference time, two optional components are added: a CFG-like interpolation between the refinement speed and the lookahead speed, and a previous-step shortcut that reuses the refinement x̂1 from the previous step to approximate an N+1 NFE.

The experiments use SD3.5-Large and ControlNet on Unsplash to perform four tasks: 8x super-resolution, depth-to-RGB, edge-to-RGB, and JPEG restoration. Further experiments fine-tuned TRELLIS-2 using LoRA for image-adjusted 3D textures on Objaverse and Toys4K. Baselines include standard fine-tuning, ControlNet++-style alignment loss, FlowChef, and in the appendix FlowDPS, LatentDAPS, FlowLPS, DDPO, DRaFT, AlignProp, higher-order solvers, double stride, CFG, and double training epochs.

The paper also reports on some analyses: ablation of fast thresholds, self-tuning, feedback windows, multi-step lookaheads, and p_un; operator shift testing; and a decomposition demonstration that approximately 80% of the correction energies are orthogonal to the gradient.

Critical strengths

The idea is simple, architecture-agnostic and practical. Treating measurement errors as learning inputs rather than manual weighted updates is a clean way to introduce the idea of ​​iterative refinement of learning into large latent flow models. The zeroth-order variant extends it to nondifferentiable operators, such as JPEG, where gradient bootstrapping does not apply. The method is demonstrated at real scale (SD3.5-Large, TRELLIS-2) and across closed form, neural networks, non-differentiable and rendering operators. The 3D texture gain is large and consistent, with roughly +4-5 dB higher masked PSNR and better FID compared to the fine-tuned baseline.

This empirical study is unusually thorough for the baseline it covers. It includes computational matching training (Table 13), equivalent NFE test-time alternatives (2x steps, RF/ABM solver, CFG), reward fine-tuning via explicit checkpoint selection analysis (Table 8), several inference-time posterior samplers, reimplementation in the spirit of Holzschuh & Thuerey, wall-clock decomposition, and operator shift testing. Together these make it quite convincing that the gains are not just from additional sampling calculations.

Critical weaknesses

Conceptual and analytical issues. (1) Since y is already in c and H is a fixed known function, s_t is a deterministic function of (x_t, c) for the given network. The CFM minimizer E[u_t | x_t, c, s_t] is therefore equal to E[u_t | x_t, c]: the feedback adds no information at the overall optimum. So the statement "theoretical guarantees are intact" is only true in a trivial sense, and the gain must come from approximation, optimization, or effective depth effects (two weights sharing a channel are deeper computations). This article does not discuss this. A shortcut to the previous step is another case: s_t depends on the trajectory history, creating an unanalyzed train-test mismatch. (2) Securities and Exchange Commission. 5.4 Orthogonality analysis does not support the claim that “it is not guidance in disguise”. In the high-dimensional latent, cos = 0.42 indicates strong alignment, well above chance. Any learned linear preconditioner (still bootstrap-like) will also produce large orthogonal components. A portion of Δv can simply reflect the network's response to any non-null input. Controls required: shuffled or random feedback, or fitted linear or preprocessed bootstrapping plots. (3) 2D toys are compared to small MLPs that fail on the 4-category conditioning task, which says more about underfitting than open-loop training.

A matter of experience. (1) The most natural competitor, the combination of standard FT with inference time guidance, is missing. For texturing, FlowChef was applied to pretrained, non-finetuned TRELLIS-2. For depth and edge, FID values ​​223 and 97 indicate a configuration mismatch. (2) The fidelity of recovery is measured against conditions, i.e., measurement consistency. Recovery metrics that reference ground truth are largely absent from the main results, and the metric scales are uninterpretable (e.g., LPIPS of 0.83 vs. 22.24). (3) Claims of uniform simultaneous gains are overstated. In terms of SR, IT Guidance has the best PSNR/SSIM. On edge to RGB, several variants have worse FID than standard FT (14.29–14.97 vs. 13.98). Under SR operator shifts, FlowBender's FID is worse than the baseline. The best variant was different for each task, and there were no seeds, confidence intervals, or repeated training runs, although the FID difference of about 0.3 was within reasonable noise. (4) Even with shortcuts, the wall-clock cost in the latent space due to VAE decoding is about 2-2.4 times that of standard FT. Computational matching training is shown only for SR. (5) The absolute value residual discards the sign of the error, and this design choice is unreasonable.

Questions for authors

Given that s_t is a deterministic function of (x_t, c), how do you interpret the gain? Compare this to a two-channel model where the second channel receives uninformative input of matching shape (e.g., zeros or scrambled feedback from another sample). This separates the effect of feedback from the effect of extra depth.

Can you please report on standard FT plus FlowChef or DPS style guidance on fine tuning the model for all tasks including texturing?

For seconds. 5.4, ​​what is the explained variance of Δv under a fitted linear plot of s_grad or under a diagonal preconditioner per time step? What is the orthogonality score when feedback is randomized?

Please report reference ground truth PSNR/LPIPS for SR and JPEG and clarify LPIPS scaling across tables.

Can you provide the differences between seeds or training runs, at least for SR and depth results as well as FID numbers?

How does the shortcut's train-test mismatch behave when t_thresh goes to 0? Have you tried training with cached feedback from previous steps?

Why use |y − H(x̂1)| instead of the signed residual of the zero-order variable?

5 comments

We thank the reviewer for the detailed and rigorous review, and for recognizing the practicality of the approach, the breadth of operators and scales it covers, and the thoroughness of the baselines and compute controls. The suggested controls have strengthened the paper; we address each question below.

1. Interpreting the gain given that s_t is a deterministic function of (x_t, c). We agree the CFM minimizer is unchanged; we will state this explicitly and revise the "guarantees intact" wording. The gain is an approximation effect: H is an external computation (e.g., a full depth estimator) that a finite-capacity network would otherwise have to learn implicitly from training pairs. In the limit of capacity and data this could in principle be distilled into the network; the controls below suggest it is not at the scale we study. The gain requires correct, sample-specific feedback, not extra depth or compute:

SRNFEPSNR vs. LR input ↑LPIPS vs. GT ↓FID ↓
Standard FT4034.350.2233.93
Ours, trained with shuffled (other-sample) feedback8033.830.2474.21
Ours (zero-order)8039.250.2133.36

Shuffled feedback keeps the second pass, compute, and input statistics, yet loses the entire gain; zeroing its feedback at inference barely changes it (34.28 PSNR), i.e., the model learns to ignore an uninformative channel. Self-conditioning (Tab. 3) has the same extra pass, yet also trails Standard FT (33.45 vs. 34.35 PSNR). This indicates that the gain comes from the feedback computed through H, rather than from the extra pass alone.

2. Standard FT + inference-time guidance. We applied FlowChef to both Standard FT and FlowBender on all tasks. FlowChef improves fidelity for either model, and the gains are complementary: FlowBender + FlowChef outperforms Standard FT + FlowChef on both fidelity and plausibility in every task (on par in edge FID). We will add these results to the final version.

Task (fidelity / FID)Standard FTOursStandard FT + FlowChefOurs + FlowChef
SR (PSNR vs. LR input ↑ / FID ↓)34.35 / 3.9339.25 / 3.3642.90 / 3.9746.35 / 3.39
Depth (δ₁ ↑ / FID ↓)0.789 / 18.210.819 / 15.700.899 / 20.560.915 / 18.77
Edge (MAE ↓ / FID ↓)0.0533 / 13.980.0456 / 14.810.0232 / 17.370.0217 / 17.32
Texturing (M.PSNR ↑ / FID ↓)21.91 / 8.7426.39 / 6.6426.68 / 7.4728.19 / 6.46
Ours: zero-order for image tasks, first-order (x̂₁) for texturing (Objaverse).

3. Explained variance and randomized feedback for §5.4. Thank you for the suggested controls. Scalar and diagonally preconditioned guidance explain under 20% of the learned correction (held-out R² = 0.16 and 0.17, respectively). We agree that orthogonality alone is not diagnostic: randomized feedback yields similar orthogonality, but only the true feedback reduces the render loss. We will replace the orthogonality argument in §5.4 with these two controls and refine the claim accordingly. Setup as in §5.4 (∇x̂₁ texturing model, 20 assets × 9 timesteps); fits are per timestep, pooled across assets, with R² measured on held-out assets.

FeedbackOrthogonal energyRender-loss change of x̂₁
True gradient0.82−15% (−29% at high noise; improves in 98% of cases)
Shuffled across voxels0.85−1%
Matched Gaussian noise0.74+1%
The model therefore exploits the content of the feedback, and the correction it applies is largely not expressible as scalar or diagonally preconditioned guidance.

4. Ground-truth metrics and LPIPS scaling. GT-referenced metrics improve as well, by smaller margins, as expected: in ill-posed tasks such as 8× SR, many outputs are consistent with the input, so GT improvement is bounded. We will add GT metrics to the main tables and clarify LPIPS scaling in the captions.

GT-referenced (single run)PSNR ↑SSIM ↑LPIPS ↓
SR: Standard FT25.020.6610.223
SR: Ours (zero-order)25.760.6640.213
JPEG: Standard FT26.000.7200.196
JPEG: Ours (zero-order)27.620.7430.187

SSIM and LPIPS above are raw values, while the image-to-image tables in the paper report them ×100; we will make this consistent. The consistency metrics differ in scale across tasks (LPIPS 0.83 vs. 22.24 for Standard FT) because they are computed on different images: for SR, the 8×-downsampled output against the LR input, where fine detail is removed and images agree closely once coarse structure matches; for JPEG, the re-compressed full-resolution output against the compressed input, where small differences in detail visibly change the quantized result.

The answers to the rest of the reviewer's questions are in the next comment.

5. Variance across seeds. We retrained the three main SR variants with two additional seeds (3 total). Every FlowBender seed beats every Standard FT seed on all three metrics, and the FID gap (0.70) is several times the seed standard deviation. We will add these results to the final version.

SR (3 seeds)PSNR vs. LR input ↑PSNR vs. GT ↑FID ↓
Standard FT33.67 ± 1.2424.78 ± 0.274.11 ± 0.19
Ours (zero-order)39.29 ± 0.8825.88 ± 0.173.41 ± 0.11
Ours (combined-x̂₁)38.83 ± 1.5225.79 ± 0.303.41 ± 0.03

These runs required over 500 A100 hours including evaluation, so we prioritized SR within the rebuttal period. We will try to add multi-seed depth results before the end of the discussion period, and will include them in the final version in any case.

6. Shortcut train-test mismatch as t_thresh → 0; training with cached feedback. The mismatch is small on depth and edge and larger on SR, where the shortcut still retains about half the fidelity gain over Standard FT. We will add the t_thresh sweep for depth and edge to the appendix (SR is shown in Fig. 6).

Task (fidelity / FID)Standard FT (N NFE)Ours, two-pass (2N NFE)Ours, shortcut t_thresh = 0 (N+1 NFE)
SR (PSNR vs. LR input ↑ / FID ↓)34.35 / 3.9339.25 / 3.3636.9 / 3.76
Depth (δ₁ ↑ / FID ↓)0.789 / 18.210.819 / 15.700.815 / 15.87
Edge (MAE ↓ / FID ↓)0.0533 / 13.980.0456 / 14.810.0470 / 14.86

All FlowBender results use the zero-order variant.

Training with cached feedback is an elegant way to close the mismatch, and we thank the reviewer for the suggestion. We have not tried it, as it requires the previous step's refined prediction during training, i.e., simulating partial trajectories, which requires a non-trivial change to the flow-matching training scheme. We agree it is an exciting direction for future work, possibly via cheaper approximations of the cached estimate.

7. Absolute vs. signed residual. Thank you for catching this; it is a typo, which we will fix in the final version. The implementation uses the signed residual H(x̂₁) − y. This is also reflected in the normalization s_t = s̃_t / max(|s̃_t|) (App. A.2), whose absolute value would be redundant for a non-negative residual.

We will also revise the claim of uniform simultaneous gains to match the tables, as discussed with Reviewer Wk6G.

We hope these responses and new experiments address your concerns; if so, we would be grateful if you would consider raising your score.

I appreciate the authors' detailed and constructive rebuttals. The new experiments resolved most of my major concerns, and I am pleased that some of them directly answered the specific control questions I requested.

I believe the following issues have been resolved or substantially resolved:

  • Regarding the interpretation of gain, I appreciate the authors' acknowledgment that the CFM minimizer is unaffected by feedback and their commitment to revising the wording "guaranteed intact." The stochastic feedback control experiment is exactly what I envisioned. The model trained using feedback from other samples completely lost its gain during the same second training and computation, actually performing even slightly worse than the standard FT. Combined with the self-conditional ablation experiment, this strongly demonstrates that the improvement comes from correct, sample-specific feedback computed through H, rather than additional depth. The explanation for the approximation effect is reasonable.

  • Regarding §5.4, I appreciate the authors' acknowledgment that orthogonality alone is insufficient for diagnosis, and the stochastic feedback control experiment clearly demonstrates this. Rendering the loss results (true gradient, stochastic gradient feedback, and Gaussian feedback) further strongly demonstrates that the model utilizes the content of the feedback. The retained R² values ​​of 0.16 and 0.17 indicate that the time-step scalar and diagonal preprocessing guidance explains only a small portion of the correction. I suggest the revised text limit the conclusions to this range. These fitting results exclude fitted scalar and diagonal families, but do not exclude general linear mappings, such as spatially variable or sample-dependent preprocessors. Furthermore, these results alone do not demonstrate that the correction is nonlinear.

  • The comparison of standard FT + FlowChef fills a major gap in my proposal across all four tasks (including texture processing on fine-tuned models). FlowBender + FlowChef outperforms standard FT + FlowChef in fidelity across all tasks, and outperforms standard FT + FlowChef in FID across all tasks except edge processing (where performance is comparable). This is a useful result indicating that the two methods are complementary.

  • The GT reference metrics for SR and JPEG, the explanation of the different evaluation spaces behind ×100 scaling and LPIPS values, the results for the three sub-SRs, the depth and edge results with t_thresh = 0, and the explanation of the use of signed residuals in the implementation all answered my questions. Regarding shortcuts, the explanation of why it was reasonable not to attempt caching feedback training is reasonable; I agree that this should be left for future research, rather than being a necessary part of this paper.

The remaining concerns, which I consider minor, mainly concern the presentation of the results:

  • The claims of consistently outperforming current best inference time guidance and simultaneously improving fidelity and plausibility still need further elaboration. The new tables show that standard FT + FlowChef typically achieves higher fidelity than using FlowBender alone (e.g., δ₁ for depth is 0.899 vs. 0.819, and MAE for edges is 0.0232 vs. 0.0456), but at the cost of higher FID values. Furthermore, several FlowBender variants, including the zero-order model used in the new tables, exhibit higher edge FID values ​​than the standard FT and also higher FID values ​​under SR operator offsets. I welcome your commitment to revise these conclusions. A more accurate statement would be that FlowBender significantly outperforms supervised fine-tuning and is a better bootstrap base model, rather than one that dominates bootstrap on its own.

  • The current ANOVA only covers SR. The SR results convincingly suggest that the main gain does not come from noise. The SR seed study did not determine the variability of depth or edge results, so the robustness of small differences between variants on these tasks remains uncertain. I look forward to results on multiple sub-depths (if possible).

  • In terms of non-functional efficiency (NFE), this shortcut approaches the cost of open-loop algorithms, but not in terms of actual runtime. When using VAE decoding, the measurement time per image reported in the appendix is ​​still approximately 2–2.4 times that of the standard FT, while on SR, the shortcut retains only about half the fidelity gain. Using it in conjunction with FlowChef would increase costs further. I appreciate the authors' transparency in reporting the time decomposition and hope the paper will also describe the efficiency claims in terms of actual runtime rather than just non-functional efficiency.

Reviewer_Wk6GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FlowBender: a conditional flow-matching model that takes its own alignment error as an input. At each step, an unguided look-ahead pass gives a clean estimate x̂₁; a forward operator H computes a deviation from the condition y (as a gradient, a raw residual, or both); and a second pass consumes that signal to produce a corrected velocity. Experiments cover SR, depth/edge-to-RGB, JPEG restoration (SD3.5 + ControlNet), and 3D mesh texturing (TRELLIS-2 + LoRA), with comparisons to standard fine-tuning, ControlNet++-style alignment loss, reward fine-tuning (DDPO/DRaFT/AlignProp), and four inference-time guidance methods.

Critical strengths
  • The core idea is simple, well-motivated, and architecture-agnostic. Keeping the exact CFM objective (feedback enters only as input) is a clean design and distinguishes it from reward fine-tuning. The benchmark and ablations are comprehensive, including applications across multiple domains.
  • The zero-order variant is a useful contribution: it extends learned correction to non-differentiable operators (JPEG), where gradient guidance doesn’t apply, and it’s also the cheapest variant at inference.
Critical weaknesses

No weakness rises to the level of supporting rejection. It is a solid, broadly validated execution of a known idea, with claims that need trimming to match the tables, which are explained in the questions section.

Questions for authors
  • Headline claim overstated on neural-operator tasks. At matched inference compute (80 NFEs), Standard FT + CFG or + RF-Solver reaches equal or better FID than every FlowBender variant on edge-to-RGB (12.41/12.84 vs. 13.68) and depth-to-RGB (14.13/14.90 vs. 14.57). FlowBender still wins fidelity on both. The “simultaneous fidelity and plausibility” claim therefore holds for SR, JPEG, and 3D, but not for depth and edge; the abstract and Sec. 5.1 should say so, and some rows in Tables 10 and 12 belong in Table 1.
  • The paper proposes combined w.r.t. x̂₁ as the most consistent default, but gives no guidance on when the cheaper zero-order variant suffices, no explanation for why the preferred gradient target differs across tasks, and does not test zero-order feedback on 3D texturing. A short decision rule, or a hypothesis linking variant choice to operator properties, would make the method considerably easier to adopt.
  • How stable is the training? It is recommended to report either multi-seed means or at least a PSNR/FID-vs-iteration curve for one task so readers can judge stability better.
1 comment

We thank the reviewer for the careful reading and for recognizing the simplicity of the design, the zero-order variant, and the breadth of the evaluation.

1. Fidelity–plausibility claim on depth/edge at matched compute. We agree for the configurations in Table 1, and will revise the claim and tables as suggested in the final version. However, we note that when we add standard CFG to FlowBender at matched NFEs, it matches (edge) or exceeds (depth) the best baseline FID while keeping a large fidelity advantage; we will add these results to the revision.

Method (~80 NFE)Edge MAE ↓Edge FID ↓Depth MAE ↓Depth FID ↓
Standard FT + CFG 3.00.050012.410.081414.90
Standard FT + RF-Solver0.051612.840.082514.13
Ours (combined-x̂₁, shortcut + text CFG 3.0)0.031412.460.074613.79

Our row uses 81 NFE: 41 for the shortcut sampler plus 40 unconditional-text passes; we do note that our wall clock generation time is still approximately 1.5x slower.

2. Choosing a variant and gradient target; zero-order on 3D. No single variant dominates across tasks, and it is difficult to know in advance which will work best for a given problem. As a general guideline, we recommend trying zero-order first, especially with heavy operators: it is cheaper, is often close to the best variant, and is the only option for non-differentiable operators. Adding the gradient can help further in some tasks (e.g., edge). We will add this discussion to the final version.

The gradient target is task-dependent; we treat it as a hyperparameter and recommend x̂₁ as the default, since it avoids the network Jacobian and is best or comparable in most of our settings.

Zero-order is not a natural fit for texturing with TRELLIS-2: the residual is a 2D render error, but the model accepts images only as DINOv3 cross-attention features, which are unlikely to preserve pixel-precise error, or to indicate which 3D voxels it came from. The first-order gradient provides exactly this, by backpropagating the render error through the differentiable renderer onto the 3D latent. We will add this discussion to the final version as well.

3. Training stability. Training is stable across seeds: we retrained the main SR variants with two additional seeds (3 total), and every FlowBender seed beats every Standard FT seed on all three metrics. We will report multi-seed means and add validation curves to the appendix.

Method (3 seeds)PSNR vs. LR input ↑PSNR vs. GT ↑FID ↓
Standard FT33.67 ± 1.2424.78 ± 0.274.11 ± 0.19
Ours (zero-order)39.29 ± 0.8825.88 ± 0.173.41 ± 0.11
Ours (combined-x̂₁)38.83 ± 1.5225.79 ± 0.303.41 ± 0.03

Validation curves show FlowBender leading Standard FT by 0.7–1.3 dB in GT PSNR throughout training, with all runs plateauing by 2–3k steps.

We hope these responses address your concerns; if so, we would be grateful if you would consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.