acceptodds
Withdrawn from ICLR 2027

Debate-Driven Generation: Multimodal Agent Debate for Iterative Text-to-Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in text-to-image generation have placed growing emphasis on reasoning for better instruction following and refinement. Nevertheless, the need to transform sparse language into dense visual content makes contextual disambiguation and elaboration a central bottleneck beyond cross-modal alignment. Therefore, we introduce the Multimodal Agent Debate (MAD) framework that adapts multi-agent debate to iterative text-to-image refinement by addressing three gaps, namely evolving debate topics, the mismatch between semantic precision and execution feasibility, and the need for factual rather than reflective moderation. MAD maintains (i) an explicit roadmap of verifiable criteria grounded in the current image, (ii) uses a Regulator to decide whether debated instructions are executable by a downstream editor or require re-debate with pragmatic constraints, and (iii) employs a cross-modal Moderator to verify prompt satisfaction and produce verifiable success conditions. Experiments on GenEval, T2I-CompBench, and TIIF-Bench show consistent overall gains in instruction following across diverse image generators. In addition, MAD also improves instruction executability and reduces repetitive refinements, supporting MAD as a training-free refinement method for T2I. The source code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal, debate, image generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3dyWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how multimodal agent reasoning can improve instruction following, attribute binding, and compositional relation errors in text-to-image (T2I) generation. The authors propose a training-free framework, Multimodal Agent Debate (MAD), in which a proponent and an opponent debate the current image against a set of dynamic verification criteria. A Moderator then determines whether the prompt is satisfied and derives verifiable diagnostic conditions. A Regulator, drawing on these diagnostics and the accumulated history, decides between executing an image edit and adding constraints to trigger re-debate, thereby linking language-level corrections to the downstream image editor. The authors report that the framework improves the Overall scores of multiple generation systems on GenEval, T2I-CompBench, and TIIF-Bench.

Critical strengths

1.The MAD framework cleanly separates "diagnosing what is wrong" from "making edits that are executable," which carries clear engineering value. Rather than feeding raw critique text directly to the editor, the framework introduces a control branch that chooses between executing an edit and adding constraints to trigger re-debate. This design explicitly acknowledges, at the system level, that identifying a problem and effectively fixing it are not the same task. It also provides an observable interface for analyzing ineffective edits and repeated operations.

2.The experiments are thorough and cover multiple generation systems. The evaluation (Table 1) reports paired Base/MAD results for Kontext, SDXL, ReasonGen-R1, GoT, MILR, and TwiG, and CLIP-FID decreases across all three reported datasets. This provides preliminary evidence that the method offers a degree of transferability across different generation systems.

Critical weaknesses

1.The novelty of this work is somewhat limited. The authors extend traditional text-based debate into multimodal agent debate to enhance the semantic fidelity of generated images, but this approach leans more toward incremental methodological improvement than an original theoretical breakthrough. Moreover, the idea of leveraging multimodal agents to strengthen image generation and editing has already appeared in existing works such as T2I-Copilot[1], GenArtist[2], and Crea[3]. The authors therefore need to more clearly articulate the essential distinction between their contribution and these existing methods.

2.The final-image selection rule shown in part 5 of Figure 3 — arg⁡max⁡0≤t≤T(m1t−∣m2t∣)\arg\max_{0\le t\le T}(m_1^t-|m_2^t|), i.e., selecting the best frame from history — contradicts the rule stated in Eq. (13) of Section 3.4, which terminates early by returning ItsI^{t_s} upon the first satisfaction and otherwise returns ITI^T once the budget is exhausted, without additional evaluation. The former requires running all rounds and then picking the optimal frame from history, whereas the latter requires terminating immediately and returning the current frame. These two strategies cannot both hold. Please clarify which implementation is actually adopted and reconcile the two statements.

[1]. T2i-copilot: A training-free multi-agent text-to-image system for enhanced prompt interpretation and interactive generation[C]//2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025

[2]. Genartist: Multimodal llm as an agent for unified image generation and editing[J]. Advances in Neural Information Processing Systems, 2024.

[3].A collaborative multi-agent framework for creative image editing and generation[J]. Advances in Neural Information Processing Systems, 2026

Questions for authors

1.The dynamic criteria (the criteria roadmap) in MAD may introduce requirements beyond the original prompt, yet the authors do not clearly specify which newly added criteria constitute necessary semantic completion (e.g., common-sense details about objects or scenes) and which constitute unauthorized goal expansion (e.g., additional attributes or preferences not requested by the user). We recommend that the authors explicitly distinguish three categories of constraints — original hard constraints, necessary object/scene common sense, and optional visual-quality preferences — and further clarify the priority rules for handling conflicts between newly added criteria and the original prompt.

2.The authors introduce adversarial proponent–opponent debate in MAD, but the ablation study does not adequately examine the independent contribution of the adversarial design itself. The current comparison only contrasts two debaters against a single critic (Single Critic), which cannot isolate the respective effects of "multi-agent" versus "adversarial interaction." We recommend adding control experiments such as two cooperative critics and two independent checkers with aggregated results, so as to independently assess the contribution of the adversarial interaction itself.

2 comments

Weakness 1. Novelty and distinctions from existing agentic generation frameworks.

Thanks for your comment. We would like to clarify that MAD differs from existing frameworks in targeting editing needs that are dynamic, relative and incremental. Each edit changes the image. The next need is thus relative to the current result rather than to the prompt alone. T2I-Copilot, GenArtist and CREA coordinate agents cooperatively toward overall semantic fidelity. Once this goal is met, the agents tend to agree and stop. For example, a prompt asks for a cat but the initial image shows a dog. After the first edit turns the dog into a cat, cooperative agents accept the image. The cat may still lack whiskers. This need appears only after the first edit. In MAD, the opponent must challenge the current image. It raises the missing whiskers as the next increment. MAD then updates its criteria to include them. Fig. 6 shows a similar case. In it, single-agent CoT accepts an image with missing toothbrush bristles. MAD identifies and corrects this omission.

Our results support this distinction. Under the same settings with Qwen-Image-Edit, MAD achieves 85.55% on GenEval. T2I-Copilot and CREA reach 68.50% and 52.99% (Table R1). On T2I-CompBench, Two Cooperative Critics reaches 45.69% Overall, against 46.75% for full MAD (Table R3). Fixing the criteria lowers T2I-CompBench Overall from 0.4675 to 0.4624 (Table 2 of the main paper).

Table R1. GenEval results for external frameworks and MAD.

MethodSingle ObjectTwo ObjectsCountingColorsPositionColor AttributionOverall ↑
CREA92.50%67.68%56.25%58.51%14.00%29.00%52.99%
T2I-Copilot96.15%89.13%56.96%78.49%67.37%22.92%68.50%
Qwen-Image-Edit+MAD97.50%91.92%87.50%89.36%72.00%75.00%85.55%

Weakness 2. Inconsistent final-image selection rules in Fig. 3 and Eq. (13).

Thank you for carefully checking Fig. 3 against Eq. (13) and pointing out this inconsistency. We have corrected Fig. 3 to follow Eq. (13), which is the rule our implementation adopts. The argmax rule in part 5 of Fig. 3 was an error in the figure. MAD returns the current image as soon as the Moderator judges that it satisfies the prompt. If the budget is exhausted, MAD returns the final image without further evaluation. No selection over historical images is performed. Appendix B states the same rule in its discussion of the maximum debate budget.

Question 1. Scope and priority of dynamically added criteria.

Thank you for your valuable suggestions. We have classified the added criteria into the three suggested categories. Most are original requirements or necessary semantic completion, and the original prompt always takes priority. We used Qwen and Mistral as annotators (Table R2). Both annotators identify necessary object or scene common sense as the largest category. Original hard constraints and necessary common sense together account for 70.94% to 80.89% of the criteria.

Optional preferences cannot override the original prompt. The Moderator detects a violation only when the image fails the original prompt requirements (Appendix C). An added preference therefore does not trigger an edit by itself. The proponent also defends regions that already satisfy the prompt. We have defined the three categories and this priority rule explicitly in the revised manuscript.

Table R2. Distribution of analyzed criteria by annotator and dataset.

AnnotatorDatasetOriginal hard constraintsNecessary object/scene common senseOptional visual-quality preferences
QwenT2I-CompBench25.70%55.19%19.12%
QwenGenEval17.94%53.00%29.07%
MistralT2I-CompBench28.24%48.68%23.08%
MistralGenEval14.83%61.56%23.61%

Question 2. Independent contribution of adversarial interaction versus multiple agents.

Thank you for your valuable suggestions. We have added the two suggested controls, and they show that the gain comes from adversarial interaction rather than from the number of agents. Both controls are run on T2I-CompBench and use two agents without opposing roles. All other components are retained. Two Cooperative Critics reaches 45.69% Overall, and Two Independent Checkers reaches 45.37% (Table R3). Neither clearly exceeds the single critic at 45.50%. Full MAD reaches 46.75%. Making the two agents oppose each other thus adds 1.06 to 1.38 points over the two-agent controls.

The advantage is consistent across categories. Full MAD outperforms Two Independent Checkers in all six categories. It also outperforms Two Cooperative Critics in five of the six categories. The largest gap is on Color (+1.63 and +2.51 points, respectively). We will add both controls to Table 2 of the revised manuscript.

Table R3. Controls for agent number and adversarial interaction on T2I-CompBench.

MethodColorShapeTextureSpatialNon-spatialComplexOverall ↑
MAD w/ Single-Critic Reasoning72.58%51.26%62.87%29.85%25.63%30.81%45.50%
Two Cooperative Critics73.65%49.68%63.80%30.20%25.56%31.23%45.69%
Two Independent Checkers74.53%50.15%63.52%30.04%24.43%29.54%45.37%
MAD76.16%51.84%64.55%30.76%26.08%31.09%46.75%
Reviewer_6Jc5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MAD is a training-free framework for iterative text-to-image refinement. Two debaters examine the current image against evolving criteria, a visually grounded Moderator decides whether the prompt is satisfied, and a history-aware Regulator authorizes an edit or adds a criterion for re-debate. The generator and editor remain frozen. The default budget is two outer rounds, including rounds without editing. Results suggest improvements across several generation pipelines and benchmarks.

Critical strengths
  1. The distinction between identifying semantic discrepancies and producing executable corrections is meaningful.
  2. The paired improvements are encouraging and remove the Moderator's image input reduces performance.
Critical weaknesses
  1. Essential protocols such as actual sample counts, generator–editor mappings, the Table 2 backbone, and the evaluation are insufficiently specified.
  2. Compute attribution such as additional reasoning and editing in pipelines remains unclear.
  3. Execution and termination claims lack outcome-level validation.
  4. The mechanistic contribution needs sharper positioning. Related works such as T2I-Copilot already combine ambiguity resolution, structured interpretation, visual evaluation, and iterative generation.
Questions for authors
  1. What is matched in the single-critic comparison, what costs are actually incurred, and does its 1.25-point gap persist under paired uncertainty analysis?
  2. Is MAD inherits controller and editor errors? May dynamic criteria add unnecessary requirements, and later edits damage correct content?
2 comments

Weakness 1. Insufficiently specified experimental protocols.

Thanks for your comment. All reasoning roles in MAD, including debate, moderation, and regulation, use Qwen3-VL-8B. The T2I-CompBench ablation in Table 2 uses 1,800 prompts, with 300 per category, a fixed random seed, and Kontext as both the generator and editor. Unless stated otherwise, Kontext is the default editor. In the T2I-CompBench comparisons across generators, all generators use Kontext as the editor.

GenEval uses 553 prompts with one image per prompt and a fixed random seed. T2I-CompBench uses 1,800 prompts, TIIF-Bench uses more than 2,500 long prompts, and COCO-5K uses the 5,000-example COCO validation set. Within each Base/MAD pair, the prompt set, random seed, and evaluator are held fixed. GenEval is scored using its official evaluator, T2I-CompBench using its benchmark-specific evaluators for attribute binding, relations, and complex composition, and TIIF-Bench using Qwen2.5-VL-72B. Full details are provided in Appendix A.

Weakness 2 and Question 1. Additional computation and the single-critic comparison.

Thanks for your comment. The two configurations are matched in everything except the debaters. They use the same Qwen3-VL-8B instance for all reasoning roles, the same Kontext editor, the same maximum number of rounds, and the same limit of two edits per sample. MAD w/ Single-Critic Reasoning replaces only the two debaters with a single self-refining critic.

MAD incurs extra reasoning and about one extra edit per sample, but no extra model memory. All roles share one VLM instance, so the model weights are loaded only once. MAD averages 1.73 edits per sample versus 0.82 for the control (Table R1). Among edited samples, the averages are nearly the same (1.84 vs. 1.83). MAD therefore refines more samples rather than editing each sample more deeply. At the system level, Kontext+MAD takes 69.9 seconds per sample, compared with 98.1 for T2I-Copilot, and uses about 46G of GPU memory, compared with 63G (supplementary Table 6 in Appendix A).

Yes, the gap persists, and it grows to 6.25 points on TIIF-Bench. With SD3 on TIIF-Bench, MAD exceeds the same control by 6.25 points Overall and scores higher in eight of the nine individual dimensions (Table R2). On T2I-CompBench, MAD also outperforms the control in all six categories (Table 2 of the main paper).

Table R1. Image-editing cost comparison.

MetricMAD w/ Single-Critic ReasoningMAD
Maximum edits per sample22
Average edits per sample0.821.73
Average edits among edited samples1.831.84

Table R2. SD3 results on TIIF-Bench.

MethodOverallBasic AvgAtt.Rel.Reas.Advanced AvgAtt.+Rel.Att.+Reas.Rel.+Reas.StyleTextReal World
SD3 Baseline63.3067.8173.9573.5055.9862.8868.6159.0463.1676.6731.8067.00
SD3 + MAD w/ Single-Critic Reasoning65.1670.0875.1575.5959.4964.5670.1161.6863.9877.6734.7767.99
SD3 + MAD71.4175.6477.0077.6972.2469.5473.8365.8970.3474.0053.5969.11

The SD3 Baseline and SD3 + MAD rows are reproduced from supplementary Table 5 in Appendix A. The MAD w/ Single-Critic Reasoning row is an additional rebuttal experiment. All scores are percentages. Basic Avg and Advanced Avg denote the Basic Following and Advanced Following averages, respectively. Att., Rel., and Reas. denote Attribute, Relation, and Reasoning, following the notation in the supplementary material.

Weakness 3. Outcome-level validation of execution and termination.

Thanks for your comment. Both mechanisms are validated at the outcome level, since weakening either one lowers the final benchmark scores. MAD stops when the Moderator judges from the current image that the prompt is satisfied. Removing the Moderator's image input causes the largest drop among all controlled variants. Overall falls from 0.4675 to 0.4417 on T2I-CompBench (Table 2 of the main paper) and from 64.05 to 60.15 on GenEval (supplementary Table 9 in Appendix B). Replacing the Regulator with a fixed instruction template also lowers Overall, to 0.4616 on T2I-CompBench and 63.15 on GenEval.

These outcome-level results are consistent with the process-level diagnostics. In Fig. 8, MAD produces the highest share of executable instructions. It reaches 81.0% on GenEval and 75.2% on T2I-CompBench, compared with 74.6% and 65.8% for Single Critic. It also reduces repetitive instructions from 42.2% to 10.9% on GenEval relative to Vanilla Debate. Fig. 6 shows a complete case. CoT and Vanilla Debate terminate despite missing toothbrush bristles, whereas MAD identifies the omission and produces an executable correction.

Weakness 4. Positioning of the mechanistic contribution.

Thanks for your comment. Unlike T2I-Copilot, MAD drives refinement through opposing debate over criteria that evolve with the current image. T2I-Copilot fixes its prompt interpretation before generation and relies on a single evaluator (Chen et al., 2025). A single evaluator can accept its own output too early. As Fig. 6 shows, CoT misses the absent toothbrush bristles. In MAD, the opponent must challenge the image, and the criteria are updated after each edit.

Controlled ablations support both mechanisms. Replacing the debaters with a single critic lowers all six T2I-CompBench categories (Table 2 of the main paper). With SD3 on TIIF-Bench, it also reduces Overall by 6.25 points (Table R2). Fixing the criteria lowers T2I-CompBench Overall from 0.4675 to 0.4624 (Table 2 of the main paper). MAD also cuts repetitive edit instructions from 42.2% to 10.9% on GenEval (Fig. 8).

Question 2. Controller and editor errors, unnecessary criteria, and preservation of correct content.

Thanks for your comment. MAD can inherit errors from its controllers and editor, but its checks limit both the errors and their side effects. The Moderator verifies the debaters' claims against the current image. The Regulator checks each correction against the execution history. This reduces repetitive edit instructions from 42.2% to 10.9% on GenEval and from 39.7% to 10.5% on T2I-CompBench (Fig. 8).

These checks also prevent unnecessary edits and protect correct content. A new criterion does not trigger an edit by itself. The Moderator detects a violation only when the image fails the original prompt (Appendix C). The proponent also defends regions that already satisfy the prompt. MAD stops as soon as the Moderator judges the prompt satisfied. Each edit follows an [Action] + [Object] + [Target State] format, which localizes the correction.

Reviewer_KTDvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Multimodal Agent Debate (MAD), a training-free framework for iterative text-to-image (T2I) refinement. Given an initial image generated by a base model, a proponent and an opponent VLM agent debate whether the image satisfies the prompt based on dynamically maintained criteria. A Moderator then renders a binary satisfaction judgment alongside diagnostic feedback. If unsatisfied, a history-conditioned Regulator either dispatches an executable edit instruction to a downstream image editor or injects an updated criterion to trigger a re-debate round. The authors evaluate MAD across multiple diffusion backbones (SDXL, SD3, FLUX-Kontext, FLUX.2) and reasoning-enhanced pipelines (GoT, ReasonGen-R1, MILR, TwiG) on GenEval, T2I-CompBench, TIIF-Bench, and COCO-5K. Experiments demonstrate consistent improvements in overall alignment scores, lower CLIP-FID, and higher instruction executability with reduced repetitive edits across all base models.

Critical strengths
  • Training-free and model-agnostic formulation. MAD operates as a modular, training-free plug-in. Its compatibility is demonstrated across diverse generation paradigms, spanning diffusion, autoregressive, and unified architectures.

  • Insightful process-level diagnostics. Beyond standard terminal benchmark metrics, the paper provides fine-grained operational diagnostics, tracking instruction executability and cross-round edit repetition (e.g., reducing repetition from 42.2% to 10.9% on GenEval). Such mechanistic analyses are rare in agentic generation literature and offer clear visibility into the dynamics of the refinement loop.

Critical weaknesses

1. The necessity of multi-agent debate is not convincingly established.
The central premise of the paper is that adversarial multi-agent debate drives refinement quality. However, the ablation study fails to isolate this effect:

  • Redundant Debate Context: Because the Moderator already receives the input prompt and current image directly (Eq. 11), the debate transcript DtD^t serves merely as auxiliary context for a judge that is already fully visual-grounded.
  • Marginal Empirical Delta: In Table 2, Vanilla Debate (0.4489) does not outperform a Single Critic (0.4504). Furthermore, replacing both debaters with a single self-refining critic under a matched compute budget yields an Overall score of 0.4550 vs. MAD's 0.4675—a marginal difference of ~0.012.
  • Task Simplicity: Inspecting the qualitative examples, tasks such as counting one vs. two cats (Fig. 3), reading an "OPEN/CLOSED" sign, or checking container containment (Fig. 1) are well within the zero-shot perception capability of a single modern VLM (e.g., Qwen3-VL-8B). These results validate the utility of a standard closed-loop "verify-and-edit" pipeline, but do not substantiate the claim that multi-agent debate is essential.

2. Evaluated generators do not reflect competitive contemporary baselines.
Most evaluated base models (SDXL, SD3/SD3.5-M, Janus-Pro, Kontext, GoT, ReasonGen-R1) exhibit lower GenEval baselines in Fig. 4(a) (ranging from ~0.50 to ~0.80) compared to current open weights (e.g., Z-Image at 0.84, Qwen-Image at 0.87). While FLUX.2 and Qwen-IE are mentioned, they appear only in the T2I-CompBench chart (Fig. 4(b)) without concrete numeric scores, and are omitted from Table 1, GenEval, and TIIF-Bench. Those base models leave substantial room for easy gains, which may sharply diminish on stronger generators with fewer trivial prompt-following errors.

3. Benchmarks are mismatched with the stated core motivation.
The paper motivates MAD by highlighting the need for contextual disambiguation and semantic elaboration in underspecified prompts. However, GenEval and T2I-CompBench rely primarily on explicit, compositional template prompts (e.g., "a black apple and a green backpack"), which require precise spatial/attribute binding rather than contextual disambiguation. To validate the claimed disambiguation capabilities, I recommend that the framework can be evaluated on knowledge-intensive benchmarks (e.g., WISE) or dense-caption benchmarks (e.g., DPG-Bench, OneIG-Bench).

4. Missing direct comparisons with prior agentic/iterative refinement methods.
The empirical comparison primarily evaluates MAD stacked on top of existing generators or contrasts it against internal ablations (Single Critic, Vanilla Debate). Although external iterative frameworks like T2I-Copilot and MCCD are discussed in Related Work, they are not empirically benchmarked under identical testbeds. Without direct comparison against established external iterative refinement baselines, it is difficult to determine MAD's relative positioning in the literature.

Minor Issue

  • Several citations do not support the statements they are attached to. In Sec. 2.2, Wei & Yang (2011) and Xiao-Yong Wei (2013), both on active learning for interactive video search and retrieval, are cited for ImageCoT/T2I-R1-style reasoning and iterative verification (the author field of the 2013 entry is also malformed). In Sec. 2.3, Sheng et al. (2025), a contrastive-decoding method for LLM hallucination mitigation that involves no debate, is cited as an example of text-only debate.
  • The header reads “Under review as a conference paper at ICLR 2026”; the ICLR 2027 style file should be used.
Questions for authors
  • What is the reference set for CLIP-FID on GenEval and T2I-CompBench?
  • MILR reports a GenEval Overall of 0.95 on Janus-Pro-7B, but the MILR baseline in Fig. 4(a) is clearly lower. Which MILR configuration is used?
3 comments

Weakness 1(a). Redundant debate context

Thanks for your comment. We would like to clarify that DtD^t is not redundant. Direct access to the image does not necessarily mean that the Moderator has already formed specific judgments about the relevant visual requirements. DtD^t highlights the requirements that warrant attention. It also provides explicit assessments and disagreements about whether they are satisfied. Our MAD w/ Single-Critic Reasoning control corresponds to the setting you describe. It retains the image-aware Moderator and Regulator, replacing only the two opposing debaters with a single self-refining critic. Table R1 reproduces the relevant T2I-CompBench results from Table 2 of the main paper.

Table R1. Controlled comparison on T2I-CompBench, reproduced from Table 2 of the main paper.

MethodColorShapeTextureSpatialNon-spatialComplexOverall
MAD w/ Single-Critic Reasoning72.58%51.26%62.87%29.85%25.63%30.81%45.50%
MAD76.16%51.84%64.55%30.76%26.08%31.09%46.75%

Removing opposing debate lowers scores in all six categories and reduces Overall by 1.25 percentage points, despite the Moderator retaining direct access to the image. A possible explanation is that a single agent can overlook fine-grained violations, while opposing debate helps surface and examine these details before the Moderator makes its decision. For example, in Fig. 6, single-agent CoT accepts an image with missing toothbrush bristles, whereas MAD identifies and corrects the omission. The debate transcript therefore gives the Moderator specific disputed details to verify, helping it detect issues that a single-agent assessment may miss.

Weakness 1(b). Marginal empirical delta

Thanks for your comment. We would like to clarify that Vanilla Debate is an unadapted baseline rather than our method. Its parity with Single Critic suggests that conventional multimodal debate is insufficient for image refinement. By contrast, our method yields substantial gains. On T2I-CompBench, full MAD reaches 0.4675, exceeding Single Critic (0.4504) and Vanilla Debate (0.4489) by 1.71 and 1.86 points, respectively (Table 2 of the main paper). Importantly, on TIIF-Bench, which contains long prompts, MAD with SD3 exceeds MAD w/ Single-Critic Reasoning by 6.25 points Overall, including +12.75 on Reasoning and +18.82 on Text (Table R2).

On T2I-CompBench, Overall averages the scores of all categories, so the 1.25-point Overall gain understates the effect of MAD. MAD improves Color by 3.58 points over the same control (Table R1).

Table R2. SD3 results on TIIF-Bench.

MethodOverallBasic AvgAtt.Rel.Reas.Advanced AvgAtt.+Rel.Att.+Reas.Rel.+Reas.StyleTextReal World
SD3 Baseline63.3067.8173.9573.5055.9862.8868.6159.0463.1676.6731.8067.00
SD3 + MAD w/ Single-Critic Reasoning65.1670.0875.1575.5959.4964.5670.1161.6863.9877.6734.7767.99
SD3 + MAD71.4175.6477.0077.6972.2469.5473.8365.8970.3474.0053.5969.11

The SD3 Baseline and SD3 + MAD rows are reproduced from supplementary Table 5 in Appendix A. The MAD w/ Single-Critic Reasoning row is an additional rebuttal experiment. All scores are percentages. Basic Avg and Advanced Avg denote the Basic Following and Advanced Following averages, respectively. Att., Rel., and Reas. denote Attribute, Relation, and Reasoning, following the notation in the supplementary material.

Weakness 1(c). Task simplicity

Thanks for your comment. We would like to note that the examples in Figs. 1 and 3 were kept simple on purpose to illustrate the refinement process. MAD is not limited to these tasks, and even a strong single VLM can fail on simple prompts. For example, in Fig. 6, CoT uses the same Qwen3-VL-8B as a single-agent reasoner and accepts an image with missing toothbrush bristles, whereas MAD identifies and corrects the omission. On TIIF-Bench, which uses long, constraint-dense prompts, MAD improves Overall for all four generators by 3.21 to 18.98 percentage points (supplementary Table 5 in Appendix A). Fig. 11 in Appendix B and Figs. 12 and 13 in the supplementary material show more complex cases, including long prompts and texture changes.

Opposing debate brings gains beyond a standard closed-loop "verify-and-edit" pipeline. MAD w/ Single-Critic Reasoning is such a pipeline, which differs from MAD only in replacing the two debaters with a single self-refining critic. Full MAD outperforms this control in all six T2I-CompBench categories (Table R1), and its Overall is 6.25 percentage points higher on TIIF-Bench with SD3 (Table R2).

Weakness 2. Evaluated generators do not reflect competitive contemporary baselines.

Thanks for your comment. We respectfully note that MAD still yields substantial gains on strong contemporary generators. Its improvements are therefore not limited to weak base models. On T2I-CompBench, MAD improves FLUX.2 by 13.0 points on Color and 9.9 on Spatial (supplementary Table 3 in Appendix A). These gains are larger than those for FLUX.1-Kontext, which has a lower baseline in every category. Stronger generators therefore do not necessarily leave less room for improvement. On GenEval, MAD improves Overall from 77.81% to 85.55% for Qwen-Image-Edit and from 83.06% to 83.52% for FLUX.2 (Table R3).

Numeric scores for FLUX.2 and Qwen-Image-Edit are reported in supplementary Table 3 in Appendix A, and we have added GenEval results for both models. Across the ten Base/MAD pairs in this table, MAD improves 51 of the 60 category-level scores, including five of the six categories for both FLUX.2 and Qwen-Image-Edit. Given the limited time, the TIIF-Bench experiments with these two generators are still in progress.

Table R3. GenEval results for FLUX.2 and Qwen-Image-Edit before and after MAD.

MethodSingle Object ↑Two Objects ↑Counting ↑Colors ↑Position ↑Color Attribution ↑Overall ↑
FLUX.2 Baseline98.75%94.95%70.00%94.68%67.00%73.00%83.06%
FLUX.2 + MAD98.75%91.92%75.00%90.43%73.00%72.00%83.52%
Qwen-Image-Edit Baseline100.00%88.89%75.00%82.98%55.00%65.00%77.81%
Qwen-Image-Edit + MAD97.50%91.92%87.50%89.36%72.00%75.00%85.55%

Weakness 3. Benchmarks are mismatched with the stated core motivation.

Thanks for your comment. We would like to clarify that our evaluation goes beyond template prompts. We also evaluate MAD on TIIF-Bench, which more directly tests the capabilities in our motivation. TIIF-Bench uses long, constraint-dense prompts and includes categories for reasoning and real-world knowledge. MAD improves Overall for all four generators by 3.21 to 18.98 percentage points (supplementary Table 5 in Appendix A). It also improves all four generators on the categories most relevant to disambiguation and elaboration (Table R4).

Table R4. TIIF-Bench category gains across four generators, calculated from supplementary Table 5 in Appendix A.

CategoryGain (points)
Reasoning+6.83 to +27.70
Attribute + Reasoning+1.70 to +20.00
Relation + Reasoning+3.54 to +18.82
Real World+2.11 to +6.53

GenEval and T2I-CompBench complement this evaluation by showing that MAD also preserves general compositional ability. MAD improves GenEval Overall for all seven generators (supplementary Table 4 in Appendix A) and improves 51 of the 60 category-level scores on T2I-CompBench (supplementary Table 3 in Appendix A).

Thank you for the excellent suggestion of WISE, DPG-Bench, and OneIG-Bench. Adding these comparisons would make our evaluation more comprehensive. In the revised manuscript, we will move the TIIF-Bench results into the main paper and clarify how they relate to our motivation.

Weakness 4. Missing direct comparisons with prior agentic/iterative refinement methods.

Thank you for your valuable suggestions. We have added direct comparisons with T2I-Copilot and CREA, and MAD achieves the highest scores among the compared systems. Under the same experimental settings on GenEval, MAD achieves 85.55% Overall, compared with 68.50% for T2I-Copilot and 52.99% for CREA (Table R5). On TIIF-Bench, SD3+MAD achieves the highest Overall score of 71.41 among all reported systems, exceeding T2I-Copilot at 70.05 (supplementary Table 5 in Appendix A).

MAD is also more efficient than T2I-Copilot. Kontext+MAD requires 69.9 seconds per sample, 28.7% less than T2I-Copilot. It uses two VLM/LLM models instead of four and about 46G of GPU memory instead of 63G (supplementary Table 6 in Appendix A). Its Overall score of 69.29 on TIIF-Bench remains close to that of T2I-Copilot.

We did not include MCCD, as we could not find a publicly available official implementation. We are therefore unable to reproduce its results under our evaluation conditions.

Table R5. GenEval results for external frameworks and MAD.

MethodSingle ObjectTwo ObjectsCountingColorsPositionColor AttributionOverall ↑
CREA92.50%67.68%56.25%58.51%14.00%29.00%52.99%
T2I-Copilot96.15%89.13%56.96%78.49%67.37%22.92%68.50%
Qwen-Image-Edit+MAD97.50%91.92%87.50%89.36%72.00%75.00%85.55%

Minor Issue 1. Citation accuracy

Thank you for carefully checking our citations and identifying these errors. We have removed the inaccurate citations. Wei & Yang (2011) and Wei (2013) have been removed from the sentence on ImageCoT and T2I-R1 in Sec. 2.2, while the references to those two methods are retained. Since both entries are cited only in this sentence, the malformed 2013 entry has also been removed from the reference list. We have also removed Sheng et al. (2025) from the text-only debate citations in Sec. 2.3 and rechecked the remaining citations for accuracy.

Minor Issue 2. Conference template and header

Thank you for pointing out this oversight. We have switched to the ICLR 2027 style file and corrected the conference year in the header.

Question 1. Reference set for CLIP-FID

Thanks for your comment. For both GenEval and T2I-CompBench, the reference set consists of the 5,000 images in COCO val2017. We compute CLIP-FID between the CLIP ViT-B/32 image features of the generated images and those of this reference set. The generated sets contain 553 and 1,800 images, respectively. These details are provided in the “Quality-metric evaluation protocol” paragraph of Appendix A.

Question 2. MILR configuration and the reported GenEval score

Thanks for your comment. We used MILR with Janus-Pro-7B and SelfReward. The reported score of 0.95 was obtained with OracleReward. OracleReward uses feedback from the benchmark evaluator as the reward during optimization. This directly optimizes the test metric and assumes feedback that is typically unavailable in real-world use. We therefore chose SelfReward, in which the model evaluates its own generated images. MILR's own reward ablation reports 0.79 on GenEval with SelfReward, close to our reproduced score of 0.77 (supplementary Table 4 in Appendix A). On T2I-CompBench, MILR reports 0.4501, and the mean of our reproduced six category scores is 0.4483 (supplementary Table 3 in Appendix A). We will explicitly label this configuration as MILR with SelfReward.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your thoughtful and constructive feedback. We have posted detailed responses to your comments, along with additional analyses and experimental results. We would appreciate it if you could take a moment to review our responses and let us know whether they resolve your concerns. If any questions remain, we would welcome the opportunity to clarify them during the discussion period.

Thank you again for your time and consideration.

Regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.