acceptodds
Under review as a conference paper at ICLR 2027

FlowTool: Controlling Tool Parameter in Image Retouching via Flow Matching

Abstract

Tool-based image editing (image retouching) is commonly formulated with autoregressive multimodal large language models (MLLMs) that sequentially generate reasoning, tool selections, and parameter values. In this work, we present a novel approach to tool-based image editing by framing the task as a flow matching problem. We introduce FlowTool, a framework that directly models the distribution of high-quality tool parameters conditioned on the input image and user instruction using conditional rectified flow. FlowTool combines a vision-language model backbone for multimodal understanding with a Diffusion Transformer parameter generator that transforms Gaussian noise into an editing plan. We train FlowTool with a two-stage supervised flow-matching curriculum, followed by reward-based post-training. Across MMArt-Bench, FlowTool-Eval, ArtEdit-Bench, and MIT-Adobe5K, FlowTool achieves significantly stronger reference-based performance than specialized MLLM editing agents and proprietary MLLMs, while remaining competitive with proprietary models under reference-free evaluation. Moreover, FlowTool significantly improves inference efficiency, reducing latency by at least while requiring nearly less memory than the compared baselines. These results demonstrate that tool-based image editing can be effectively modeled as conditional generation over structured continuous editing parameters, without autoregressive reasoning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image editing, photo retouching, flow matching

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_mnZoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FlowTool presents a practical framework for tool-based image editing, with clear benefits in structured parameter prediction and inference efficiency. However, the methodological novelty is relatively limited, and the paper does not sufficiently establish that Flow Matching is preferable to simpler or alternative multimodal prediction methods for this task. The lack of controlled comparisons against strong continuous-output baselines is particularly important because it leaves the central methodological motivation insufficiently validated. Together with the remaining questions concerning synthetic instruction quality, reward robustness, and the interpretation of the efficiency results, these limitations weaken the case for acceptance.

Critical strengths
  1. Clear presentation. The paper is generally well-structured and clearly presents the proposed framework, training pipeline, and experimental results, making the method relatively easy to follow.

  2. Reasonable evaluation metrics. The evaluation uses a combination of reference-based metrics and reference-free assessments, providing complementary perspectives on image-editing quality and instruction following.

Critical weaknesses
  1. Insufficient evidence for the necessity of Flow Matching. The central methodological claim is that conditional Flow Matching is a better formulation for continuous image-editing parameters than autoregressive generation. However, the current experiments do not isolate the contribution of Flow Matching from that of the structured output representation and task-specific training. In particular, the comparison between different ODE integration steps does not establish superiority over a properly tuned direct-regression model or other multimodal prediction approaches. The one-step variant is not necessarily a strong representative of these alternatives. Without such controlled comparisons, it remains unclear whether the observed gains arise from Flow Matching itself or primarily from replacing text-based outputs with a dedicated continuous prediction architecture.

  2. Limited methodological novelty beyond existing components. The proposed framework combines a pretrained VLM, a DiT-based parameter generator, a tool-presence head, a two-stage supervised training curriculum, and reward-based post-training using an existing DiffusionNFT-style objective. While this integration is reasonable, the paper does not introduce a substantially new flow-matching objective or reinforcement-learning algorithm. The main contribution therefore appears to lie in task formulation and system integration. The paper would benefit from a clearer explanation of what generalizable methodological insight it provides beyond applying existing generative modeling and post-training techniques to a new output space.

  3. Potential limitations in data construction and generalization. The training corpus contains approximately 497K examples derived from internally curated expert edits, with user instructions synthesized by a pretrained VLM. The paper does not sufficiently characterize the fidelity of these synthetic instructions, the distribution of tool activations, or the diversity of editing styles. Moreover, the 300-example internal evaluation set provides limited evidence of generalization to real user instructions and out-of-distribution editing requests. It would be useful to evaluate instruction quality, report more detailed dataset statistics, and assess generalization under independently collected user prompts.

  4. Reward design and efficiency claims require further validation. The post-training reward combines pixel-level similarity to an expert-edited target with a VLM-based instruction-following score. These objectives can conflict, and the reported gains from RL are not uniform across all metrics. Further reward ablations and independent evaluation would help establish the robustness of the improvement. In addition, the reported latency comparison mixes local inference time for FlowTool with end-to-end API response times for proprietary models. Although this distinction is disclosed, it prevents attributing the full latency gap to the proposed generation paradigm. Comparisons against controlled local baselines with matched hardware and inference boundaries would provide stronger evidence for the claimed efficiency advantage.

Questions for authors

Please refer to "Weaknesses" above.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer mnZo (1/2)

We thank the reviewer for the thoughtful comments and suggestions. We have significantly revised and updated our paper with additional results and discussion. We address the reviewer's concerns as follows:

W1

W1. Insufficient evidence for the necessity of Flow Matching.

We agree that evaluating different ODE integration steps does not replace a comparison with a separately trained direct-regression model. To address this concern, we train a deterministic regression model and an autoregressive model (both with Qwen3.5-4B backbone) with our curated training dataset (the same data used to train FlowTool). The regression comparison evaluates the benefit of flow matching over deterministic prediction, while the autoregressive comparison assesses whether the gains stem from our flow-based formulation rather than differences in training data.

MethodMMArt L1 ↓MMArt L2 ↓MMArt SC ↑MMArt PQ ↑MMArt O ↑FT-Eval L1 ↓FT-Eval L2 ↓FT-Eval SC ↑FT-Eval PQ ↑FT-Eval O ↑ArtEdit L1 ↓ArtEdit L2 ↓ArtEdit SC ↑ArtEdit PQ ↑ArtEdit O ↑
FlowTool9.4217.988.169.598.817.8713.218.049.648.767.6512.958.349.758.99
Qwen3.5-4B-Regression10.2221.396.679.507.8010.2121.816.649.617.809.2618.247.279.728.25
Qwen3.5-4B-Autoregressive10.1620.736.039.587.419.3818.765.669.527.088.5416.256.379.767.67

MMArt = MMArt-Bench, FT-Eval = FlowTool-Eval, ArtEdit = ArtEdit-Bench.

The results show that FlowTool significantly outperforms the regression baseline despite both methods producing continuous editing parameters. This supports the benefit of flow-based parameter generation over the deterministic predictor, beyond the choice of a continuous output representation. FlowTool also substantially outperforms the autoregressive baseline, further supporting the contribution of our flow-based formulation beyond access to task-specific training data. These controlled ablations strengthen our empirical validation relative to JarvisArt and RetouchIQ, which, to our knowledge, do not report comparable experiments that hold training data fixed to distinguish the contribution of their modeling approaches from the benefits of their datasets. More details are available in Appendix C of the revised manuscript.

W2

W2. Limited methodological novelty beyond existing components.

We agree that FlowTool builds on established flow-matching and reinforcement-learning objectives; our contribution is not a new objective for either framework. Rather, it is a formulation of parameter-based tool use as conditional generative modeling over continuous editing parameters, together with the architectural and training choices needed to make this formulation effective. FlowTool can produce an editing plan through flow matching without autoregressive reasoning. We are the first to do this for image retouching, to the best of our knowledge, which position our work completely different from existing state-of-the-art works like JarvisArt, RetouchIQ, where they all use the same paradigm of finetuning MLLM for autoregressive parameter generation.

The central methodological insight is that multimodal understanding and executable parameter generation can be learned together without expressing the editing plan as an autoregressive language sequence. FlowTool models dependencies among continuous parameters during joint generation, while explicitly separating tool activation from parameter-value prediction. This distinction matters because predicting an appropriate value and deciding whether to activate the corresponding tool are different decisions. The tool-presence head and masked flow-matching objective encode this structure, and our ablation demonstrates the benefit of explicitly modeling activation.

The two-stage curriculum addresses a complementary optimization issue: adapting a pretrained VLM to a randomly initialized continuous generator. Our ablation supports first learning parameter generation under fixed multimodal representations, followed by joint adaptation. Together, these results provide evidence for the design choices underlying the formulation.

So overall, our novelty lies in distinctive problem formulation, adaptation of multiple flow-based techniques to image editing with tools (being the first to do so and successfully), and strong empirical effectiveness and efficiency compared to existing methods. In short, our approach is new, different and it performs much better than specialized MLLMs and competitively with frontier MLLMs, while remains much more efficient in model size and inference time.

AuthorsOpenReview ↗

Official Response to Reviewer mnZo (2/2)

W3

W3. Potential limitations in data construction and generalization.

We would like to clarify that our evaluation already covers both in-distribution and out-of-distribution settings. FlowTool-Eval is the only in-distribution benchmark, consisting of held-out examples from our internal data collection. MMArt-Bench, ArtEdit-Bench, and MIT-Adobe5K are out-of-distribution benchmarks relative to FlowTool’s training dataset, constructed independently of our internal corpus. Overall, our whole evaluation set across 4 benchmarks consists of 1135 samples, in which 300 samples are in-distribution witht the training data, and the remaining 835 samples are entirely out of distribution. Our evidence of generalization therefore extends beyond the 300-example internal evaluation set: performance on these three external benchmarks supports transfer beyond the training-data distribution. In addition, we also added the user preference study in Appendix D of the revised manuscript, showing that FlowTool remains strong on human evaluation, in addition to reference-based metrics and VLM-as-a-judge evaluation.

We agree that a more detailed assessment of synthetic instruction quality and an evaluation using independently collected user prompts would strengthen the paper. However, manually verifying instructions and collecting and annotating new user requests require substantial annotation effort beyond the short rebuttal period. We plan to add these analyses in a subsequent revision, together with a more detailed discussion of instruction quality, subjecting to internal approval. Thank you for your suggestion.

W4

W4. Reward design and efficiency claims require further validation.

We used SC and L1 as reward to train our model with RL. Considering the 15 metrics across the three instruction-driven benchmarks evaluated with L1+SC rewards in Table 1, SC and the overall score improve across all three benchmarks, while both L1 and L2 improve significantly on FlowTool-Eval and ArtEdit-Bench. These results demonstrate that reference similarity and instruction alignment can improve jointly under our combined reward.

On MMArt-Bench, RL improves SC and the overall score, with only very small increases in L1 and L2. These minor declines in reference similarity accompany significant gains on the other two benchmarks. The decreases in PQ on FlowTool-Eval and ArtEdit-Bench are also negligible. Since the base model already achieves strong perceptual quality, we place greater emphasis on improving SC through RL, and the consistent SC gains demonstrate the effectiveness of this choice. Overall, these results support improved instruction alignment and overall editing quality while largely preserving perceptual quality.

Regarding latency, since the proprietary models cannot be deployed locally and we cannot access their server-side timing, we report end-to-end API response time as the accessible measure of their deployment latency, which we believe is the best alternative. Moreover, for locally deployable baselines (JarvisArt and RetouchIQ), we control the hardware, evaluation inputs, and timing scope: all models run on an NVIDIA A100, using the same 200 examples, with single-sample inference and latency measured for editing-plan generation. Under these conditions, FlowTool takes 0.19 seconds, compared with 32.11 seconds for JarvisArt and 30.36 seconds for RetouchIQ. Thus, our significant efficiency advantage is supported by a fair, controlled local comparisons and does not depend on API-related overheads. We believe provided a sufficiently strong enough evidence for our claimed efficiency advantage.

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

We really appreciate your constructive feedback, which has helped us significantly strengthen the manuscript.

As the rebuttal period is about to end, we would appreciate your thoughts on whether our responses and revisions have adequately addressed your concerns. If any questions remain, we would be happy to discuss them and provide further clarification.

Thank you again for your time and consideration.

Reviewer_mmz9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

FlowTool reformulates tool-based image retouching as conditional generation over a structured continuous tool-parameter space.

  1. A VLM encodes the image and instruction;
  2. A DiT-like generator uses conditional rectified flow to generate continuous parameters;
  3. A separate tool-presence head predicts which tools are active.

The method uses a two-stage supervised flow-matching curriculum followed by reward-based post-training using rendered-image rewards. Experiments cover MMArt-Bench, FlowTool-Eval, ArtEdit-Bench, and MIT-Adobe5K, with reference-based quality metrics, VLM-judged instruction alignment, latency/memory measurements, and ablations for ODE steps, tool-presence modeling, curriculum and VLM scale.

Critical strengths

S1: The paper makes a clear case that normalized retouching parameters form a structured continuous space and that jointly generating them can avoid token-level numerical decoding and error accumulation.

S2: The paper has road and relevant validation. The method is evaluated on multiple benchmarks with both reference-based and reference-free metrics, efficiency measurements, and targeted ablations. The tool-presence and two-stage-curriculum ablations support the proposed decomposition.

Critical weaknesses

W1: Novelty is not yet convincing. The paper combines several familiar components, and RetouchIQ (CVPR 2026) already covers the same retouching setting with executable tool parameters and generalist reward. The paper should say more clearly what flow-based joint generation adds, beyond changing the decoder, and should include a matched comparison with RetouchIQ if possible.

W2: Some of the main comparisons need tighter protocols. The reward is based on a VLM judge, and the reference-free results are also judged by a VLM. The paper should state whether these judges are independent and check for evaluator-specific gains. The 50× latency number also mixes local inference with proprietary API response time, so it is not a clean model-speed comparison. The internal training data and FlowTool-Eval split need more detail on image and near-duplicate removal.

W3: In the paper, the order of the tools is not considered, which limits the practical usage in real world.

Questions for authors

Q1: What happens in a matched comparison with RetouchIQ?

Q2: Does flow matching beat direct regression or conditional diffusion when the backbone and parameter budget are fixed?

Q3: Is the reward judge independent of the evaluation judges?

Q4: How were image-level and near-duplicate overlaps removed?

Q5: Can the latency results separate local model time from API and network time?

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer mmz9 (1/2)

We thank the reviewer for the thoughtful comments and suggestions. We have significantly revised and updated our paper with additional results and discussion. We address the reviewer's concerns as follows:

W1, Q1, Q2

W1: Novelty is not yet convincing. The paper combines several familiar components, and RetouchIQ (CVPR 2026) already covers the same retouching setting with executable tool parameters and generalist reward. The paper should say more clearly what flow-based joint generation adds, beyond changing the decoder, and should include a matched comparison with RetouchIQ if possible.

Q1: What happens in a matched comparison with RetouchIQ?

Q2: Does flow matching beat direct regression or conditional diffusion when the backbone and parameter budget are fixed?

Regarding the novelty, image retouching is an established research direction that predates RetouchIQ, which represents one recent contribution to this broader area. Our contribution concerns how editing plans are represented, learned, and generated within this setting. RetouchIQ and JarvisArt follow the standard autoregressive paradigm, generating reasoning traces and numerical tool parameters as token sequences. On the contrary FlowTool formulates the task as conditional generative modeling over continuous editing parameters, producing an editing plan through flow matching without autoregressive reasoning. We are the first to do this for image retouching, to the best of our knowledge.

This distinction extends beyond replacing the decoder: FlowTool changes the output representation, training objective, and generation process. Combined with explicit tool-presence prediction, this produces a coordinated editing plan without serializing numerical values or requiring reasoning-trace supervision. Our novelty therefore lies in this problem formulation and the integration of multimodal understanding, continuous parameter generation, and tool selection into a unified framework, while RetouchIQ's focus is on training generalist reward model and training autorgressive model with that reward. Moreover, FlowTool is also the first to propose adaptation of multiple flow-based techniques to image editing with tools, and strong empirical effectiveness and efficiency compared to existing methods. In short, our approach is new, different and it performs much better than specialized MLLMs and competitively with frontier MLLMs, while remains much more efficient in model size and inference time.

Regarding a matched comparison, fully reproducing RetouchIQ's training pipeline is currently infeasible because the its artifacts, including checkpoints, training data and trained reward model, are not publicly released. We also note that our evaluation follows the standard practice of JarvisArt and RetouchIQ: training on a dedicated task-specific dataset and comparing against existing state-of-the-art systems.

Nevertheless, we appreciate the reviewer's request for a controlled comparison that more directly assesses the contribution of our formulation. To address the reviewer's concern, we train a deterministic regression model and an autoregressive model (both with Qwen3.5-4B backbone) with our curated training dataset (same data used to train FlowTool). The regression comparison evaluates the benefit of flow matching over deterministic prediction, while the autoregressive comparison assesses whether the gains stem from our flow-based formulation rather than differences in training data.

MethodMMArt L1 ↓MMArt L2 ↓MMArt SC ↑MMArt PQ ↑MMArt O ↑FT-Eval L1 ↓FT-Eval L2 ↓FT-Eval SC ↑FT-Eval PQ ↑FT-Eval O ↑ArtEdit L1 ↓ArtEdit L2 ↓ArtEdit SC ↑ArtEdit PQ ↑ArtEdit O ↑
FlowTool9.4217.988.169.598.817.8713.218.049.648.767.6512.958.349.758.99
Qwen3.5-4B-Regression10.2221.396.679.507.8010.2121.816.649.617.809.2618.247.279.728.25
Qwen3.5-4B-Autoregressive10.1620.736.039.587.419.3818.765.669.527.088.5416.256.379.767.67

The results show that FlowTool significantly outperforms the regression baseline, justifying the benefit of flow-based parameter generation over deterministic generation. FlowTool also substantially outperforms autoregressive baseline, indicating that our flow-based formulation contributes to the imrpovement, beyond the benefit of the training dataset alone. These controlled ablations strengthen our empirical validation relative to JarvisArt and RetouchIQ, which do not report comparable experiments that hold training data fixed to distinguish the contribution of their modeling approaches from the benefits of their datasets. More details in Appendix C of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer mmz9 (2/2)

W2, Q3, Q4, Q5

W2: Some of the main comparisons need tighter protocols. The reward is based on a VLM judge, and the reference-free results are also judged by a VLM. The paper should state whether these judges are independent and check for evaluator-specific gains. The 50× latency number also mixes local inference with proprietary API response time, so it is not a clean model-speed comparison. The internal training data and FlowTool-Eval split need more detail on image and near-duplicate removal.

Q3: Is the reward judge independent of the evaluation judges?

Q4: How were image-level and near-duplicate overlaps removed?

Q5: Can the latency results separate local model time from API and network time?

As stated in Section 5.1, we use Qwen3.8-27B for reward computation during RL and GPT-5 for reference-free evaluation. The reward and evaluation judges are totally different and independent, so we believe no evaluator-specific gain was induced

Regarding latency, since the proprietary models cannot be deployed locally and we cannot access their server-side timing, we report end-to-end API response time as the accessible measure of their deployment latency, which we believe is the best alternative. Moreover, for locally deployable baselines (JarvisArt and RetouchIQ), we control the hardware, evaluation inputs, and timing scope: all models run on an NVIDIA A100, using the same 200 examples, with single-sample inference and latency measured for editing-plan generation. Under these conditions, FlowTool takes 0.19 seconds, compared with 32.11 seconds for JarvisArt and 30.36 seconds for RetouchIQ. Thus, our significant efficiency advantage is supported by a fair, controlled local comparisons and does not depend on API-related overheads.

Regarding data separation, each source-edited image pair is assigned a unique identifier, and the training and FlowTool-Eval splits contain mutually exclusive sets of these identifiers. This ensures that no image-pair record is shared between the two splits. Moreover, FlowTool-Eval is intended to measure in-distribution performance on held-out examples from our internal data collection. We additionally evaluate on MMArt-Bench, ArtEdit-Bench, and MIT-Adobe5K, which were constructed independently of our internal dataset. These external benchmarks provide clear evidence of generalization beyond the internal evaluation distribution. Our evaluation therefore covers both held-out internal data and independently curated external benchmarks. In addition, we also added the user preference study in Appendix D of the revised manuscript, showing that FlowTool remains strong on human evaluation, in addition to reference-based metrics and VLM-as-a-judge evaluation.

W3

W3: In the paper, the order of the tools is not considered, which limits the practical usage in real world.

Firstly, in terms of practical usage in real world, we like to refer the reviewer to Appendix D in the revised manuscript, where we conducted a comprehensive user preference study and points out many advantages of FlowTool compared to the current systems on real human evaluation.

Secondly, for your concern regarding the order of tools, we would like to clarify that, as specified in Section 4.1, our formulation considers parameter-based retouching in which the rendering engine determines the execution order. For the Lightroom tools, the render engine applies them in a predefined order and processing pipeline. Changing the order of these tool parameters does not change the underlying rendering order.

JarvisArt and RetouchIQ likewise use Lightroom tools, so their autoregressive output order should not be confused with the engine's internal processing order. So FlowTool and all baselines actually share the same tool execution order. We agree that learning tool order is important for broader editing workflows in which sequential operations can produce different outcomes. Such workflows extend beyond the standardized parameter-based setting studied here and represent an interesting direction for future work.

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

We really appreciate your constructive feedback, which has helped us significantly strengthen the manuscript.

As the rebuttal period is about to end, we would appreciate your thoughts on whether our responses and revisions have adequately addressed your concerns. If any questions remain, we would be happy to discuss them and provide further clarification.

Thank you again for your time and consideration.

Reviewer_GBK8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a framework, FlowTool, to retouch images, which is different from the commonly used autoregressive-based editing. The proposed method combines MLLM, DiT, and a tool-presence prediction head. The model is trained through a two-stage supervised curriculum followed by reward-based post-training. Extensive experimental results demonstrate its effectiveness.

Critical strengths
  1. This paper is well-organized and easy to follow.
  2. The paper explores an interesting problem that replaces MLLM-based image editing with flow matching tool prediction.
  3. The proposed method shows impressive inference speed-up.
Critical weaknesses
  1. One of my major concerns is that several important baselines are missing, including directly applying a MLP regression head, a transforemer regression head, or other alternative methods to predict the continuous editing parameters.
  2. Unfair comparison. FlowTool is trained on 497K task-specific examples with additional reward-based post-training, while general-purpose MLLMs are evaluated in a zero-shot setting. Fairer comparisons with autoregressive prediction and direct regression baselines trained under similar conditions are needed to isolate the benefits of flow matching.
  3. Evaluations are mainly based on L1 or L2-related metrics, which may not reflect the real editing quality. Human evaluations are required to better examine the advantages and improvements.
Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer GBK8 (1/2)

We thank the reviewer for the thoughtful comments and suggestions. We have significantly revised and updated our paper with additional results and discussion. We address the reviewer's concerns as follows:

W1, W2

W1. One of my major concerns is that several important baselines are missing, including directly applying a MLP regression head, a transforemer regression head, or other alternative methods to predict the continuous editing parameters.

W2. Unfair comparison. FlowTool is trained on 497K task-specific examples with additional reward-based post-training, while general-purpose MLLMs are evaluated in a zero-shot setting. Fairer comparisons with autoregressive prediction and direct regression baselines trained under similar conditions are needed to isolate the benefits of flow matching.

Firstly, we would like to note that our evaluation follows the standard practice of existing works in this field such as JarvisArt and RetouchIQ: training on a dedicated task-specific dataset and comparing against existing state-of-the-art systems.

However, we believe the reviewer's concern is absolutely valid. Therefore, to address these two concerns, we train a deterministic regression model and an autoregressive model (both with Qwen3.5-4B backbone) with our curated training dataset (the same data used to train FlowTool). The regression comparison evaluates the benefit of flow matching over deterministic prediction, while the autoregressive comparison assesses whether the gains stem from our flow-based formulation rather than differences in training data.

MethodMMArt L1 ↓MMArt L2 ↓MMArt SC ↑MMArt PQ ↑MMArt O ↑FT-Eval L1 ↓FT-Eval L2 ↓FT-Eval SC ↑FT-Eval PQ ↑FT-Eval O ↑ArtEdit L1 ↓ArtEdit L2 ↓ArtEdit SC ↑ArtEdit PQ ↑ArtEdit O ↑
FlowTool9.4217.988.169.598.817.8713.218.049.648.767.6512.958.349.758.99
Qwen3.5-4B-Regression10.2221.396.679.507.8010.2121.816.649.617.809.2618.247.279.728.25
Qwen3.5-4B-Autoregressive10.1620.736.039.587.419.3818.765.669.527.088.5416.256.379.767.67

MMArt = MMArt-Bench, FT-Eval = FlowTool-Eval, ArtEdit = ArtEdit-Bench.

The results show that FlowTool significantly outperforms the regression baseline, justifying the benefit of flow-based parameter generation over deterministic generation. FlowTool also substantially outperforms autoregressive baseline, indicating that our flow-based formulation contributes to the observed gains, beyond the benefit of the training dataset alone. Together, these comparisons provide stronger empirical support for our formulation. These controlled ablations strengthen our empirical validation relative to JarvisArt and RetouchIQ, which, to our knowledge, do not report comparable experiments that hold training data fixed to distinguish the contribution of their modeling approaches from the benefits of their datasets. More details in Appendix C of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer GBK8 (2/2)

W3

W3. Evaluations are mainly based on L1 or L2-related metrics, which may not reflect the real editing quality. Human evaluations are required to better examine the advantages and improvements.

Firstly, we want to emphasize that we are not testing only on L1 and L2, but we also use reference-free metrics with GPT-5 as the judge.

Secondly, we thank the reviewer for suggesting a human preference study. In response, we conducted a study with 22 participants across 24 editing cases, comparing FlowTool with GPT-5.6 Sol and JarvisArt. For each comparison, participants viewed the editing instruction, input image, and two anonymized outputs presented in randomized order, then assessed instruction following and overall visual quality separately. The comprehensive results and analysis are provided in Section 5.3 and Appendix D of the revised manuscript. Below is the summary of our user study result:

User study results. Judgments: number (share) of the 528 judgments per competitor and criterion choosing FlowTool (Win), neither (Tie), or the competitor (Loss); Gain = win share − loss share. By case / By user: number of the 24 cases or 22 participants in which FlowTool receives more (W), as many (T), or fewer (L) judgments than the competitor; p from a two-sided sign test on the non-tied outcomes.

CompetitorCriterionWinTieLossGainBy case W / T / LpBy user W / T / Lp
GPT-5.6 SolInstruction Following246 (46.6%)27 (5.1%)255 (48.3%)−1.7%9 / 2 / 130.527 / 1 / 140.19
GPT-5.6 SolOverall Look289 (54.7%)40 (7.6%)199 (37.7%)+17.0%16 / 1 / 70.0914 / 1 / 70.19
JarvisArtInstruction Following367 (69.5%)16 (3.0%)145 (27.5%)+42.0%20 / 0 / 40.00221 / 0 / 1<0.001
JarvisArtOverall Look369 (69.9%)25 (4.7%)134 (25.4%)+44.5%19 / 0 / 50.00722 / 0 / 0<0.001

Overall, the results show a clear preference for FlowTool over JarvisArt on both criteria. This advantage holds across individual judgments and remains statistically significant when preferences are aggregated by case or by user. Against GPT-5.6 Sol, instruction-following preferences are closely matched, while overall look preferences favor FlowTool. These findings provide human evidence of FlowTool’s strong editing quality, outperforming specialized MLLMs and achieving competitive performance against a leading general MLLM.

AuthorsOpenReview ↗

A Gentle Reminder

Dear Reviewer,

We really appreciate your constructive feedback, which has helped us significantly strengthen the manuscript.

As the rebuttal period is about to end, we would appreciate your thoughts on whether our responses and revisions have adequately addressed your concerns. If any questions remain, we would be happy to discuss them and provide further clarification.

Thank you again for your time and consideration.

Reviewer_23xROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FlowTool recasts tool-based image retouching as conditional generation of continuous editing parameters instead of text-based reasoning and tool calls. It combines a vision-language backbone, a DiT parameter generator using rectified flow (3 ODE steps by default) and a tool-presence head, with supervised training followed by reward-based post-training. It is evaluated on four benchmarks with reference-based and judge-based metrics, plus latency and memory.

Critical strengths
  1. A well-motivated reformulation. With a fixed tool set, fixed execution order and low-dimensional continuous parameters, directly generating final control values is a natural fit and has clear value for interactive retouching.

  2. Strong reference-based quality and speed. L2 improves over the best general model by about 21% on MMArt and 22% on ArtEdit (Table 1). Latency is 0.19 s versus 9.66 s for the fastest API baseline and over 30 s for the local specialized agents (Fig. 3), with 9.89 GB versus 16.13 to 17.09 GB of memory.

Critical weaknesses
  1. The paper shows that the whole system works, but not that flow matching is the key. Avoiding numeric text output does not require flow, since MLP, mixture-density or continuous regression heads also avoid it. Training mostly pairs one image with one expert parameter set plus synthesized instructions, and multimodality of valid solutions is not shown. The one-step ODE result is also not an independently tuned regression baseline, despite being described as "a simple regression model".

  2. Key advantages are confounded. FlowTool uses about 497K internal expert examples, while general models receive no task-specific training. The 50x latency figure compares a local model with API response time, which includes network, queueing and service effects. Reference similarity also does not establish user preference: semantic consistency is below the best proprietary model on MMArt (8.16 vs 8.34) and FlowTool-Eval (8.04 vs 8.30), and there is no human preference study.

Questions for authors
  1. With the same 497K examples and VLM conditioning, how does flow compare with a deterministic regressor and with a text parameter generator without reasoning?
  2. Can you show evidence that valid parameter solutions are diverse or multimodal under the same condition?
  3. Can you provide a human preference evaluation?
  4. Can you report the tool catalog, parameter dimensionality, full DiT configuration and training hyperparameters? They are not specified in the appendix.
6 comments
AuthorsOpenReview ↗

Official Response to Reviewer 23xR (1/3)

We thank the reviewer for the thoughtful comments and suggestions. We have significantly revised and updated our paper with additional results and discussion. We address the reviewer's concerns as follows:

W1, Q1

W1. The paper shows that the whole system works, but not that flow matching is the key. Avoiding numeric text output does not require flow, since MLP, mixture-density or continuous regression heads also avoid it. Training mostly pairs one image with one expert parameter set plus synthesized instructions, and multimodality of valid solutions is not shown. The one-step ODE result is also not an independently tuned regression baseline, despite being described as "a simple regression model".

Q1. With the same 497K examples and VLM conditioning, how does flow compare with a deterministic regressor and with a text parameter generator without reasoning?

We agree that flow matching is not required to avoid numerical text generation. Our motivation for adopting it is that a single image–instruction pair can admit multiple editing configurations that satisfy the request. A deterministic regression head predicts a single parameter configuration, while flow matching allows us to model a conditional distribution over editing configurations. DiffusionNFT also relies on within-group reward variation, which flow matching enables by sampling multiple editing plans for the same input, whereas a deterministic regression head produces identical outputs and therefore cannot support this group-relative RL setup without additional stochasticity. We therefore consider flow matching a suitable choice for both modeling the potentially one-to-many nature of the task and optimizing rendered outcomes through VLM feedback.

Following the reviewer's suggestion, we train a deterministic regression model and an autoregressive model (both with Qwen3.5-4B backbone) with our curated training dataset (the same data used to train FlowTool). The regression comparison evaluates the benefit of flow matching over deterministic prediction, while the autoregressive comparison assesses whether the gains stem from our flow-based formulation rather than differences in training data.

MethodMMArt L1 ↓MMArt L2 ↓MMArt SC ↑MMArt PQ ↑MMArt O ↑FT-Eval L1 ↓FT-Eval L2 ↓FT-Eval SC ↑FT-Eval PQ ↑FT-Eval O ↑ArtEdit L1 ↓ArtEdit L2 ↓ArtEdit SC ↑ArtEdit PQ ↑ArtEdit O ↑
FlowTool9.4217.988.169.598.817.8713.218.049.648.767.6512.958.349.758.99
Qwen3.5-4B-Regression10.2221.396.679.507.8010.2121.816.649.617.809.2618.247.279.728.25
Qwen3.5-4B-Autoregressive10.1620.736.039.587.419.3818.765.669.527.088.5416.256.379.767.67

The results show that FlowTool significantly outperforms the regression baseline, justifying the benefit of flow-based parameter generation over deterministic generation. FlowTool also substantially outperforms autoregressive baseline, indicating that our flow-based formulation contributes to the observed gains, beyond the benefit of the training dataset alone. Together, these comparisons provide stronger empirical support for our formulation. Also, these controlled ablations strengthen our empirical validation relative to JarvisArt and RetouchIQ, which, to our knowledge, do not report comparable experiments that hold training data fixed to distinguish the contribution of their modeling approaches from the benefits of their datasets. More details are provided in Appendix C of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer 23xR (2/3)

Q2

Q2. Can you show evidence that valid parameter solutions are diverse or multimodal under the same condition?

We thank the reviewer for this suggestion. We conducted an additional sampling study that provides evidence of diverse parameter configurations achieving comparable reference similarity under the same image–instruction condition.

We select 50 held-out conditions from each of the four benchmarks. For each condition, we let FlowTool generate 10 editing plans from independent initial noise samples. We record the sampled parameters and a PSNR/SSIM-based similarity score to the expert-edited target, scaled to [0,1][0,1] with higher values indicating greater similarity. We measure parameter variability over 13 global tone and color controls, expressing standard deviations and spans as percentages of each tool’s range. Results is in below table:

BenchmarkMean similaritySimilarity STDParameter STD (% range)Samples within 10% of bestSamples within 0.02 of best, per condition
MMArt-Bench0.2750.0156.0%71%5.1 / 10
FlowTool-Eval0.6360.0385.4%67%2.9 / 10
ArtEdit-Bench0.6580.0455.6%64%2.3 / 10
MIT-Adobe5K0.6020.0304.6%70%2.6 / 10
All0.5430.0325.4%68%3.2 / 10

(1) Parameter diversity is evident. For the same image and instruction, the average parameter standard deviation across seeds is 5.4% of the tool range. Specifically, across tonal controls: the ten samples span, on average, 50.4% of the Whites range, 41.0% of the Blacks range, and 38.1% of the Highlights range (span means difference between the maximum and minimum parameter values). These results demonstrate substantial parameter variation while the input image and instruction remain unchanged.

ToolSTD across seeds (% range)Mean span (% range)
Whites14.450.4
Blacks11.641.0
Highlights10.538.1
Shadows5.919.2
Vibrance4.718.5
Contrast4.516.7

(2) Multiple sampled plans achieve comparable reference similarity. Despite this parameter variation, the average within-condition standard deviation in similarity is 0.032. Moreover, 68% of all samples are within 10% of their condition’s best score, and an average of 3.2 samples per condition fall within an absolute margin of 0.02. These findings show that multiple sampled plans can retain comparable reference fidelity.

Below, we provide a concrete example from an image-instruction pair from FlowTool-Eval benchmark: four sampled plans achieve similar reference scores while using markedly different combinations of tonal and detail-enhancement parameters. Their Whites settings range from −96 to +23, and their Blacks settings from −30 to +28, illustrating the extent of parameter variation compatible with similar reference fidelity.

SeedSimilarityExposureContrastHighlightsShadowsWhitesBlacksTextureClarityVibranceSaturation
70.621+0.04+14−73+18−17−17+38+14+30+9
90.6210.00+10−58+26+17−27+61+20+17+15
50.611+0.10+35−38+17−96+28+67+35+1+4
60.611+0.03+15−91+44+23−30+17+14+24+7

These results support modeling a conditional distribution over editing configurations, since different parameter combinations can achieve comparable reference fidelity for the same request. Together with our controlled regression comparison, they strengthen the empirical motivation for flow-based parameter generation. We will include these results in the next revision of the paper.

AuthorsOpenReview ↗

Official Response to Reviewer 23xR (3/3)

W2, Q3

W2. Key advantages are confounded. FlowTool uses about 497K internal expert examples, while general models receive no task-specific training. The 50x latency figure compares a local model with API response time, which includes network, queueing and service effects. Reference similarity also does not establish user preference: semantic consistency is below the best proprietary model on MMArt (8.16 vs 8.34) and FlowTool-Eval (8.04 vs 8.30), and there is no human preference study.

Q3. Can you provide a human preference evaluation?

Our evaluation of proprietary MLLMs follows the practice in JarvisArt and RetouchIQ of using general-purpose models without task-specific fine-tuning. Fine-tuning frontier proprietary models like GPT or Claude on nearly 497K examples is beyond our available resources. However, our comparisons also include JarvisArt and RetouchIQ, both of which receive substantial task-specific training on their respective datasets. Thus, our results include comparisons against specialized editing agents, alongside general-purpose models. The additional autoregressive baseline trained on our own dataset, described above, further isolate the contribution of training data.

Regarding latency, since the proprietary models cannot be deployed locally, we report end-to-end API response time as the accessible measure of their deployment latency, which we believe is the best alternative. Nevertheless, our efficiency advantage also holds against locally deployed baselines under the same hardware and evaluation protocol: on an NVIDIA A100, FlowTool generates an editing plan in 0.19 seconds, compared with 32.11 seconds for JarvisArt and 30.36 seconds for RetouchIQ.

We acknowledge that FlowTool's SC scores are slightly below the strongest proprietary models on MMArt-Bench and FlowTool-Eval. However, SC and PQ capture complementary aspects of editing quality: instruction alignment alone does not establish that an output is visually natural or free of artifacts. FlowTool achieves higher PQ than all compared proprietary models across the three instruction-driven benchmarks. Considering both dimensions, it achieves the highest overall score on MMArt-Bench and ArtEdit-Bench, while remaining competitive on FlowTool-Eval. These results support a strong balance between instruction alignment and overall look.

Finally, we thank the reviewer for suggesting a human preference study. In response, we conducted a study with 22 participants across 24 editing cases, comparing FlowTool with GPT-5.6 Sol and JarvisArt. For each comparison, participants viewed the editing instruction, input image, and two anonymized outputs presented in randomized order, then assessed instruction following and overall visual quality separately. The comprehensive results and analysis are provided in Section 5.3 and Appendix D of the revised manuscript. Below is the summary of our user study result:

User study results. Judgments: number (share) of the 528 judgments per competitor and criterion choosing FlowTool (Win), neither (Tie), or the competitor (Loss); Gain = win share − loss share. By case / By user: number of the 24 cases or 22 participants in which FlowTool receives more (W), as many (T), or fewer (L) judgments than the competitor; p from a two-sided sign test on the non-tied outcomes.

CompetitorCriterionWinTieLossGainBy case W / T / LpBy user W / T / Lp
GPT-5.6 SolInstruction Following246 (46.6%)27 (5.1%)255 (48.3%)−1.7%9 / 2 / 130.527 / 1 / 140.19
GPT-5.6 SolOverall Look289 (54.7%)40 (7.6%)199 (37.7%)+17.0%16 / 1 / 70.0914 / 1 / 70.19
JarvisArtInstruction Following367 (69.5%)16 (3.0%)145 (27.5%)+42.0%20 / 0 / 40.00221 / 0 / 1<0.001
JarvisArtOverall Look369 (69.9%)25 (4.7%)134 (25.4%)+44.5%19 / 0 / 50.00722 / 0 / 0<0.001

Overall, the results show a clear preference for FlowTool over JarvisArt on both criteria. This advantage holds across individual judgments and remains statistically significant when preferences are aggregated by case or by user. Against GPT-5.6 Sol, instruction-following preferences are closely matched, while overall look preferences favor FlowTool. These findings provide human evidence of FlowTool’s strong editing quality, outperforming specialized MLLMs and achieving competitive performance against a leading general MLLM.

Q4

Q4. Can you report the tool catalog, parameter dimensionality, full DiT configuration and training hyperparameters? They are not specified in the appendix.

We have revised the manuscript and added Appendix B for implementation details (training hyperparameters, DiT configuration, etc.) as well as Appendix E for the tool catalogs that we used.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response Regarding the Revised Manuscript

We thank all reviewers for their careful assessment and constructive feedback. In addition to the individual responses, we have revised the manuscript to incorporate reviewers' suggestions and provide further experimental evidence and clarification. Additions and revisions are highlighted in blue. The main updates are as follows:

  • Human evaluation in Section 5.3 of the main text. We added a user preference study with 22 participants across 24 editing cases, comparing FlowTool with GPT-5.6 Sol and JarvisArt on instruction following and overall look.
  • Expanded implementation details in Appendix B. We provide additional architectural details, model configurations, and training settings to improve transparency and reproducibility. Disclosure of certain proprietary details remains subject to legal constraints and internal approval, so we include the information we are currently permitted to share and will add more in the revised manuscript.
  • Additional ablation experiments in Appendix C. We compare FlowTool with deterministic regression and autoregressive models trained on the same dataset. These comparisons assess the benefit of flow matching over deterministic parameter prediction and examine whether our gains extend beyond the contribution of task-specific training data.
  • Comprehensive human evaluation analysis in Appendix D. We provide the study design and evaluation protocol, detailed preference results by case and by user, analyses across editing experience levels, and a qualitative analysis of user rationales.
  • Tool catalog in Appendix E. We provide the catalog of the editing tools used in our framework.

Together, these revisions strengthen the empirical support for FlowTool through controlled comparisons and human evaluation, while making the methodology and implementation more transparent. We hope to engage in further discussion with the reviewers and address any remaining concern.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.