acceptodds
Under review as a conference paper at ICLR 2027

Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation

Abstract

Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and editing, adapting their actions based on interaction history.However, they typically fix the continuation state (i.e., the image from which refinement proceeds) to the latest image.Since iterative editing is not necessarily monotonic, the agent may continue refining from the latest image even when it has already degraded. In this work, we introduce GenRevisit, an agentic framework that extends action selection to joint continuation-state and action selection. The agent can not only use interaction history to adjust editing instructions but also select an image candidate to edit, jointly deciding where to continue from and how to refine. We train GenRevisit with SFT followed by agentic reinforcement learning. To address the coarse credit assignment of final outcome reward in GRPO, which assigns the same trajectory-level advantage to all actions, we further propose Normalized Action Credit Assignment (NACA) to redistribute positive advantage according to each action's contribution to reducing the remaining objective. On GenEval2, GenRevisit achieves GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, surpassing the strongest baselines using the same generators by 12.7 and 10.9 points respectively. GenRevisit achieves state-of-the-art performance on GenEval, GenEval++, and GenEval2 among baselines. Code is available at https://anonymous.4open.science/r/GenRevisit-7274/.

Primary area:
generative models
Keywords:
Agentic Image Generation, Compositional Text-to-Image Generation, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XZcROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces GenRevisit, an image-generation agent that chooses a previously generated image and an instruction for editing it. Training combines selective teacher-guided SFT with NACA, which allocates positive trajectory advantages based on normalized progress at each step. Experiments cover three compositional benchmarks, transfer across image generators, training ablations, and comparisons with controlled access to source images. GenRevisit achieves a GM score of 84.3 on GenEval2.

Critical strengths
  • The method addresses a concrete failure of iterative editing: later edits can destroy previously satisfied constraints. Historical source selection provides a practical recovery mechanism.
  • Strong reported benchmark results are complemented by matched-prefix comparisons, providing evidence beyond selected qualitative examples.
Critical weaknesses
  • The final trained policy is not directly compared under Joint, Best-only, and Latest-only access with otherwise identical settings. The archived student study always selected the current best during historical edits, leaving the central claim of beneficial non-best source selection insufficiently established.
  • NACA is compared mainly against standard GRPO, without simpler action-credit baselines or multi-seed results. Its specific normalization benefit and robustness are therefore unclear.
Questions for authors
  • With the final checkpoint fixed, how does performance change under Joint, Best-only, and Latest-only source access? How often do beneficial edits use a source that is neither latest nor best?
  • Are NACA’s gains consistent across training seeds, and how does it compare with simply removing positive credit from regressive actions?
  • Does “two-seed selection” in Figure F.7 apply only to qualitative examples or also to the reported GenEval++ scores?
3 comments
AuthorsOpenReview ↗

Response to Reviewer XZcR

Thank you for the constructive assessment and for identifying the need for more direct validation of the final policy and NACA. Your comments have helped us organize additional controlled source-access evaluations, clarify the credit-assignment analysis, and make the evaluation protocol more explicit. We respond to each question below.


Q1: Joint Source Selection and Non-Best Continuations

To directly address this concern, we evaluate the same final RL checkpoint under three source-access settings: Latest-only, Best-only, and Joint. All settings start from the same initial images and follow the same evaluation protocol, with only the set of eligible editing sources differing.

Source accessGenEval2 GM
Latest-only80.9
Best-only82.1
Joint (Ours)84.3

Our objective is not to maximize non-best selections, but to avoid restricting continuation to a fixed strategy. While best-only may work well in many cases, the highest-scoring image is not necessarily the most promising starting point for every subsequent edit. Joint selection allows the policy to choose the current best, the latest image, or another historical candidate, depending on the remaining errors and the editing instruction suited to each source. Through this process, the policy implicitly learns to assess the continuation potential of different candidates and determine where and how to continue.

To further characterize this behavior, we examine how frequently the Joint policy selects sources that are neither the latest nor the current GM-best. Of the 800 trajectories, 445 contain at least one editing action, with 1,267 editing actions in total. Source-selection frequencies use these editing actions as the denominator; initial generation and regeneration calls are excluded. Such selections account for 6.24% (79/1,267) of all editing actions, and 63.29% (50/79) of these edits produce a new candidate surpassing the historical GM-best. This frequency is not an optimization target; the value of Joint selection lies in learning when different continuation states are useful, rather than enforcing a fixed preference for any particular source.

Q2: NACA Robustness and Regressive-Action Masking

We agree that removing positive credit from regressive actions is a meaningful alternative to NACA. Our motivation is to distinguish meaningful progress during iterative refinement: binary masking identifies whether an action improves its selected source, while NACA further differentiates the relative progress made by improving actions. The qualitative cases below illustrate these two aspects.

Case A: Binary Masking vs. NACA (a bear and four glass horses)

Editing actionObserved effectΔU\Delta UMaskingNACA
Edit 1Counting error remains0.00081.0000.002
Edit 2Counting error remains0.00381.0000.007
Edit 3Fourth horse added; count resolved1.22811.0002.396

Although all three edits yield positive utility gains, only the final edit resolves the remaining counting constraint. Masking assigns equal positive credit, whereas NACA places substantially greater weight on the effective repair.

Case B: Absolute Improvement vs. NACA (a metal raccoon to the right of two suitcases)

Editing actionObserved effectΔU\Delta UAbsoluteNACA
Edit 1Material and count repaired1.16890.6380.724
Edit 2Identity restored; prior gains preserved1.08070.5901.394

Despite similar absolute utility gains, the second edit closes nearly all of its remaining utility gap (99.97%, compared with 51.95% for the first edit), restoring the raccoon's identity while preserving the previously established material and count requirements. NACA assigns greater credit based on normalized progress, whereas absolute-gain allocation assigns similar weights to both actions.

The reported coefficients are obtained by offline reweighting of fixed trajectories, using proxy action-token lengths and a trajectory advantage of +1+1. They illustrate the allocation mechanism rather than results from separately trained policies.

To assess robustness and compare against simpler alternatives, we evaluate standard GRPO, regressive-action filtering, absolute-improvement allocation, and NACA. NACA achieves 84.30 and 84.14 GM across two independent training runs, demonstrating consistent results across different training seeds. Regressive-action filtering achieves 83.1 GM, and absolute-improvement allocation achieves 82.9 GM. NACA improves GenEval2 GM by 4.0 points over standard GRPO, 1.2 points over regressive-action filtering, and 1.4 points over absolute-improvement allocation.

Credit assignmentGenEval2 GMNACA gain (GM)
Standard GRPO (two-seed average)80.3+4.0
Regressive-action filtering83.1+1.2
Absolute-improvement allocation82.9+1.4
NACA (two-seed average)84.3—

Q3: Clarification on Two-Seed Selection in Figure F.7

Thank you for pointing out this ambiguity, and we apologize for the unclear wording in Figure F.7. The two-seed selection applies exclusively to the qualitative examples in Figure F.7. The GenEval++ overall score of 89.3 reported in Table 2(b) was obtained by averaging results from two seeds across all 280 prompts, following the same evaluation protocol for both seeds. We are clarifying this distinction in the revised figure caption.


We are revising the manuscript and preparing the supplementary material for upload to incorporate these results and clarifications.


We sincerely appreciate your time and valuable suggestions. Your thoughtful feedback has helped us further strengthen the experimental validation and improve the clarity of our work. Thank you again for your careful review and constructive comments.

AuthorsOpenReview ↗

Additional Baselines and Component Ablations for NACA

Thank you again for raising the question of whether NACA's improvements arise from its specific credit-assignment design. In our previous response, we compared NACA against standard GRPO, regressive-action filtering, and absolute-improvement allocation. To further strengthen this analysis, we have conducted additional controlled comparisons, including credit-conserving filtering, an EditThinker-inspired Differential GRPO baseline, and an ablation of credit capping.


The complete results are summarized below:

Training Method / ConfigurationGenEval2 GM
Standard GRPO80.3
Regressive-action Filtering83.1
Absolute-improvement Allocation82.9
Credit-Conserving Filtering83.3
EditThinker-inspired Differential GRPO82.6
NACA without Capping84.1
NACA (Ours)84.3

These comparisons examine three complementary aspects of NACA's design: remaining-gap normalization, token-weighted credit redistribution, and credit capping.

(1) Remaining-gap progress normalization. To examine the role of remaining-gap normalization, we compare NACA against absolute-improvement allocation, which distributes credit according to absolute utility gains rather than remaining-gap-normalized local progress, while retaining the same token-weighted credit redistribution and credit capping. NACA achieves 84.3 GM, compared with 82.9 GM for absolute-improvement allocation (+1.4). This supports the benefit of accounting for the remaining improvement gap when measuring the relative progress of individual editing actions.

(2) Credit allocation under a conserved budget. Regressive-action filtering removes positive credit from non-improving actions, thereby reducing the total token-weighted positive credit mass. We introduce credit-conserving filtering to control for this change by restoring the original credit mass while assigning equal per-token advantages to retained actions. NACA achieves 84.3 GM, compared with 83.3 GM for credit-conserving filtering (+1.0). This demonstrates that conserving the total credit budget alone is insufficient; credit must also be differentiated according to the progress made by individual actions.

(3) Credit capping. We further conduct an ablation by removing credit capping while keeping the other NACA components unchanged. The resulting model achieves 84.1 GM, compared with 84.3 GM for the full method (+0.2). This small difference is consistent with the rare occurrence of extreme credit allocations observed in our training trajectories, suggesting that credit capping primarily serves as a safeguard against excessive credit concentration, rather than being the main source of NACA's performance gains.

Comparison with differential-reward training. We additionally implement and train an EditThinker-inspired Differential GRPO baseline adapted to GenRevisit. The baseline combines differential rewards with step-wise group-relative optimization over candidate joint actions. This provides a direct empirical comparison between local differential-reward optimization through step-wise GRPO and NACA's trajectory-level optimization with source-relative action credit assignment. The results show that the adapted Differential GRPO provides a competitive approach to multi-round policy optimization. Nevertheless, NACA achieves a further 1.70 GM improvement, highlighting the benefit of its trajectory-level optimization and source-relative action credit assignment.


Taken together, these experiments support the complementary roles of NACA's three design components. NACA combines remaining-gap normalization, progress-aware credit allocation under a conserved token-weighted budget, and credit capping while retaining the trajectory-level terminal objective.Together, these mechanisms provide an effective credit-assignment approach for multi-round refinement.


We sincerely thank you again for your thoughtful review and constructive suggestions. We hope these additional experiments and analyses further clarify the motivation and empirical effectiveness of NACA's design, and provide stronger empirical support for its role in multi-round refinement. We greatly appreciate your time and consideration.

Reviewer_UYeMOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2(+1)
Presentation
3(+1)
Soundness
3(+2)
Read review
Paper summary

The paper proposes GenRevisit, an agentic framework for compositional image generation that jointly selects a historical image candidate and an editing instruction. By allowing the agent to revisit earlier candidates, the framework aims to recover from edits that violate previously satisfied constraints. The planner operates with frozen generation and editing models and is trained through selective supervised fine-tuning followed by reinforcement learning. The proposed Normalized Action Credit Assignment (NACA) redistributes positive trajectory advantages according to the local progress of individual image actions. Experiments on GenEval, GenEval++, and GenEval2 report improved compositional alignment, supported by source-selection analyses and cross-generator evaluation.

Critical strengths
  1. Revisiting historical candidates is a sensible approach to handling edits that break previously satisfied constraints.

  2. The controlled teacher experiments and cross-generator evaluation provide useful evidence for the framework.

Critical weaknesses
  1. Many works have explored learning-based visual refinement, and this paper’s novelty beyond existing approaches is insufficiently established. Representative works EditThinker [1] and Thinking-while-Generating [2] are neither discussed nor compared. In particular, NACA’s use of local score improvements as a learning signal closely resembles EditThinker’s differential edit reward (improvements between consecutive editing outputs as a training reward.). The paper should clarify the difference, supported by appropriate empirical comparisons.

  2. Besides, positioning/comparison against training generation agents is limited. Representative works in this filed Gen-Searcher [3], GenEvolve [4] are all omitted despite their relevance to learned tool use and generation-agent training. Learning-based visual refinement or SFT/RL-trained generation agents have already been explored in prior work [1–4]. Discussion and comparisons on overlapping tasks are needed to establish the contribution relative to these approaches.

  3. Local progress does not necessarily reflect useful credit. NACA can reward source-relative improvements that never improve the archive, while assigning zero positive credit to temporarily degrading actions that enable later success. Although acknowledged in the limitations, this tension with non-monotonic refinement is not empirically investigated.

  4. The necessity of NACA’s specific design is unclear. Table 3 compares the complete method against standard GRPO but does not isolate progress filtering, remaining-gap normalization, or weight capping. Simpler alternatives, such as uniformly rewarding improving actions or weighting by absolute improvement, would clarify what drives the gain.

  5. The final agent’s advantage over best-only editing remains insufficiently isolated. In Appendix E.3, all 365 historical edits select the current best candidate, demonstrating historical access rather than selection beyond best-only. The teacher and SFT comparisons are informative, but a matched Joint-versus-Best-only evaluation of the final RL policy is still needed.

  6. External comparisons are not sufficiently controlled. Baseline scores come from their respective published configurations, with different planners and no common compute protocol. Sharing the generator alone does not control feedback access or editing budgets, making the reported margins difficult to attribute specifically to the proposed method.

  7. Evaluation only using GenEval, GenEval++, and GenEval2, leaving generalization beyond single benchmark family unclear. Broader evaluation on widely-used benchmarks such as WISE, together with appropriately adapted editing tests on ImgEdit-Bench/RISE, are necessary.

In summary, the discussion of and comparison with representative prior work are incomplete, leaving the paper’s position and performance within the existing literature unclear. Together with insufficient evidence isolating the central contributions, these issues raise substantial concerns about novelty and comparative advantage. Overall, I recommend reject this paper.

References

[1] EditThinker: Unlocking Iterative Reasoning for Any Image Editor

[2] Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation

[3] Gen-Searcher: Reinforcing Agentic Search for Image Generation

[4] GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

Questions for authors

See weaknesses

8 comments
AuthorsOpenReview ↗

Response to Reviewer UYeM

Thank you for the detailed review and for drawing our attention to several closely related directions. We acknowledge that the original manuscript did not sufficiently discuss important work on iterative visual refinement and learning-based image-generation agents, and we apologize for these omissions. Your comments have also helped us sharpen the research question addressed by GenRevisit.


Q1: Relation to EditThinker and Thinking-while-Generating

Thank you for highlighting EditThinker [1] and Thinking-while-Generating (TwiG) [2]. We acknowledge their relevant contributions and apologize for overlooking these works in the original Related Work. We will incorporate both into the revised version.

Although these works are broadly related to visual generation, they address fundamentally different problems and operate at different stages of the generation process:

  • Generation paradigm: TwiG interleaves textual reasoning with autoregressive visual-token generation. In contrast, GenRevisit operates across complete image-generation and editing attempts using diffusion-based generators, where the agent evaluates candidates and decides how to proceed.
  • Agentic refinement: EditThinker uses a critique–refine–repeat process to improve editing instructions. GenRevisit investigates an additional decision: using generation history as a set of selectable continuation states. It therefore jointly learns where to continue and how to edit.
  • RL optimization: We acknowledge the connection between EditThinker’s differential edit reward and NACA’s use of local progress. EditThinker incorporates edit-level score differences into its reward, whereas NACA retains the trajectory-level terminal reward and redistributes positive advantage among intermediate image actions according to source-relative, remaining-gap-normalized progress. NACA thus changes action-level credit allocation rather than introducing an additional edit-level reward term.

Q2: Relation to Gen-Searcher and GenEvolve

Thank you for bringing Gen-Searcher [3] and GenEvolve [4] to our attention. We will expand our discussion of related work to better position GenRevisit with respect to these studies.

Gen-Searcher emphasizes search-augmented generation through external knowledge and visual-reference acquisition, while GenEvolve studies tool orchestration and visual-experience distillation. GenRevisit focuses on non-monotonic, multi-round refinement for compositional generation. Our central question is whether historical generation results can serve not only as feedback for improving subsequent actions, but also as alternative starting points for further refinement. We therefore jointly learn where to continue and how to edit, rather than following a predetermined source-selection rule.


Q3: Source-Relative Progress and Long-Term Credit Assignment

Thank you for raising this important question about the relationship between local progress and global optimization.

Our goal is to learn a better retry policy, rather than reinforce only actions that immediately surpass the historical best. No fixed continuation strategy is necessarily optimal in every context. A lower-scoring candidate may preserve different useful content or offer an opportunity to resolve an error that remains unsolved in the current best. Improving such a candidate is still a capability we want the policy to learn, even if the resulting image does not immediately surpass the global best.

The historical best measures the strongest outcome achieved so far, whereas source-relative utility progress measures how effectively an intervention improves its selected input. These serve complementary roles: the terminal reward evaluates the final submitted image at the episode level, while NACA assigns action-level credit according to normalized local progress. This allows the policy to learn effective interventions across different continuation states without requiring every intermediate action to immediately improve the global best. NACA is designed as a local progress-based credit assignment mechanism, rather than a guarantee of globally optimal search or an estimator of every action's long-term contribution.


References

[1] EditThinker: Unlocking Iterative Reasoning for Any Image Editor

[2] Thinking-while-Generating: Interleaving Textual Reasoning throughout Visual Generation

[3] Gen-Searcher: Reinforcing Agentic Search for Image Generation

[4] GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

Q4: NACA Design Choices and Component Ablations

Our design distinguishes meaningful progress during iterative refinement: initial generation can yield large absolute gains, whereas subsequent edits target residual errors. Absolute-gain allocation may underweight these later refinements.

(1) Credit allocation analysis. We replayed Absolute Improvement and NACA on identical recorded trajectories, changing only the progress measure. Across five RL training rounds , we analyze 766 positive-advantage trajectories using recorded advantages and sampled action-token masks, with identical token-weighted normalization.

Credit shareAbsolute ImprovementNACA
Initial generation82.51%60.46%
Subsequent image actions17.49%39.54%

NACA increases the mean credit share for subsequent retries by 2.26×. The highest-credit action changes in 376/766 trajectories, including 372 shifts to a subsequent retry. All 372 retries increase the number of satisfied constraints, and 294 (79.0%) achieve full constraint satisfaction for the first time. These trajectory-level replays illustrate how normalized progress reallocates credit toward subsequent actions that resolve remaining constraints, motivating the controlled training ablations below.

(2) Controlled training ablations. We conducted two matched-training ablations:

  • Regressive-action filtering: removes positive credit from non-improving image actions while retaining the original credit for other actions.
  • Absolute-improvement allocation: replaces normalized progress with absolute positive utility gains while preserving token-weighted normalization and capping.
Credit assignmentGenEval2 GM
Standard GRPO80.3
Regressive-action filtering83.1
Absolute-improvement allocation82.9
NACA84.3

NACA improves GM by 4.0, 1.2, and 1.4 points over GRPO, filtering, and absolute-improvement allocation, respectively. The comparison with absolute-improvement allocation further suggests that accounting for the remaining utility gap is beneficial, as absolute gains alone may favor large improvements during initial generation over meaningful refinements in subsequent rounds.


Q5: Final-Policy Comparison under Different Source-Access Settings

We compare the same final RL checkpoint under Latest-only, Best-only, and Joint access. All settings start from the same initial image and follow the same evaluation protocol, differing only in eligible editing sources.

Source accessGenEval2 GM
Latest-only80.9
Best-only82.1
Joint (Ours)84.3

Across 800 GenEval2 evaluation trajectories, 445 contain at least one editing action, yielding 1,267 editing actions in total. Excluding initial generation and regeneration, sources that are neither latest nor current GM-best account for 6.24% (79/1,267) of edits. 50/79 (63.29%) produce a candidate surpassing the historical GM-best.

We recognize that our description of the 365 historical edits in Appendix E.3 could be clearer. Of the 481 editing actions in the Full arm, 365 select best historical sources. 39 edits (8.11% of all edits) select sources that are neither latest nor best. We will incorporate these additional statistics into the revised Appendix E.3.


Q6: External Baseline Comparisons and Configuration Differences

Our external comparisons examine whether a smaller planner trained for iterative refinement can achieve strong compositional-generation performance relative to systems using more powerful planners. GenRouter and GEMS use Kimi-K2.5 with Qwen-Image-2512, whereas GenRevisit trains Qwen3-VL-8B for joint source and action selection. GenRevisit obtains 84.3 GM, exceeding the reported GenRouter and GEMS results (71.6 and 70.4 GM) by 12.7 and 13.9 points.

Agentic generation systems differ in their planners, feedback mechanisms, interaction protocols, and inference budgets. We therefore use their reported results to contextualize benchmark performance, rather than to claim a controlled comparison of computational efficiency. We examine the contributions of individual design choices separately through controlled source-access comparisons and credit-assignment ablations.

Q7: Generalization Beyond the GenEval Benchmark Family

We additionally evaluate GenRevisit on T2I-CompBench++ [5] to assess generalization beyond the GenEval family within compositional generation.

MethodColorNumeracyTextureShapeSpatialAvg-5
Flow-GRPO0.83790.67520.72360.61300.49590.6691
Qwen-Image0.83920.77050.75680.59410.45720.6836
OctoT2I0.84100.73200.72590.62700.48030.6812
Nano Banana0.87280.79230.78550.64200.51390.7213
EvoGen-Harness [6]0.86710.83440.79130.71200.63250.7675
GenRevisit0.89650.89480.82640.70510.51330.7672

Our current study focuses on compositional generation with fine-grained, verifiable feedback. Extending the framework to knowledge-intensive generation (e.g., WISE) or general image-editing benchmarks (e.g., ImgEdit-Bench and RISE) would require adapting the feedback and interaction protocols for multi-round source selection and refinement. We view these as valuable future directions.


We sincerely appreciate your thoughtful suggestions, which helped clarify GenRevisit’s positioning and strengthen our evaluation.


References

[5] T2I-CompBench++: An Enhanced and Comprehensive Benchmark for Compositional Text-to-image Generation

[6] EvoGen-Harness: Learning Where and How to Evolve Image-Generation Harnesses

Reviewer_mWp5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the non-monotonic problem in agentic text-to-image generation, where existing iterative edit agents are forced to keep editing only the latest image that has already degraded. This paper proposes GenRevisit, where the agent jointly selects which historical image to backtrack and what edit instruction to apply. To train this policy, the authors use masked SFT on teacher traces, followed by GRPO with a new Normalized Action Credit Assignment (NACA) to avoid rewarding bad intermediate steps in successful trajectories. Evaluated on GenEval2, the method achieves great improvement over baselines and shows good zero-shot transfer on Z-Image-Turbo, alongside strong ablations.

Critical strengths

1. Clear motivation. The insight that multi-turn editing is non-monotonic and needs explicit backtracking is practical. The GM gain on GenEval2 and the zero-shot transfer to Z-Image-Turbo show that the learned policy captures a generalizable logic rather than over-fitting a specific generator.

2. Tackles a real credit-assignment issue in RL. The authors show that standard GRPO degrades the policy by blindly assigning positive advantages to harmful intermediate actions. NACA's progress-based credit redistribution turns this into a clean gain, giving the paper algorithmic value beyond simple prompt engineering.

Critical weaknesses

1. Lack of comparison to explicit search baselines. Dynamically choosing from candidate states is like tree search. The paper compares against simple latest-only and best-only heuristics, but it remains unclear whether the learned joint policy outperforms an untuned agent running with a small verifier-guided beam search.

2. Heavy dependence on automated verifiers. The entire loop and reward rely on an atomic VQA verifier. The paper lacks validation with an independent evaluation model or human study, leaving it ambiguous whether the large performance leap comes from better composition or from gaming the specific verifier.

Questions for authors
  1. How does GenRevisit compare to an explicit search baseline, such as standard iterative editing paired with a simple Beam Search (e.g., width 2) guided by GM scores?
  2. What is the practical runtime overhead (average total visual/text tokens and latency per prompt) of feeding multiple historical images into the planner at each step?
  3. Have you run any cross-evaluator check (e.g., a human preference test or an independent MLLM judge) on the final images to confirm that the +12.7 gain isn't overfitted to the training verifier?
1 comment
AuthorsOpenReview ↗

Response to Reviewer mWp5

We appreciate the reviewer’s thoughtful reading of our work and the concrete suggestions for strengthening its empirical evaluation. In response, we have prepared additional comparisons with explicit search, analyzed the token overhead of historical-image conditioning, and evaluated the final images with independent MLLM evaluators. We respond to each question below.


Q1: Comparison with an Explicit Search Baseline

Thank you for suggesting an explicit search baseline. We implemented a GM-guided Beam Search baseline with width 2 to examine whether maintaining and expanding multiple candidates can provide similar benefits to the learned joint source–action policy.

Both methods use the same maximum budget of five image calls per prompt. On GenEval2, Beam Search with the RL-trained iterative editing planner obtains 82.5 GM, compared with 84.3 GM for GenRevisit, a difference of 1.8 GM. The average number of image calls is 3.89 for Beam Search and 2.62 for GenRevisit.

MethodGenEval2 GMAvg. image calls
Beam Search (width 2)82.53.89
GenRevisit84.32.62

The two approaches differ in how they allocate a limited generation budget. Beam Search maintains multiple candidates through fixed-width, GM-guided expansion and pruning, creating a breadth–depth trade-off. GenRevisit instead learns to select a continuation state jointly with its editing action, allowing it to refine a promising trajectory while retaining the option to revisit historical candidates. These results suggest that learned joint state–action selection makes effective use of the available image-call budget compared with this fixed-width search strategy.


Q2: Computational Overhead of Historical-Image Conditioning

Thank you for raising the practical cost of conditioning the planner on interaction history. We measured the additional image calls and token usage under the maximum five-image-call budget.

MetricGenRevisit
Average image calls per prompt2.62
Additional calls beyond initial generation1.62
Additional input tokens6,784
Additional output tokens600
Historical visual tokens (excluding latest image)251.52
Historical visual-token ratio2.74%

Although the planner can access multiple historical images, the visual tokens contributed by images other than the latest one account for only 2.74% of the total input tokens across the workflow. Using the final RL checkpoint, the initial-generation outputs achieve 59.27 GM with one image call. The full iterative workflow achieves 84.33 GM with an average of 2.62 image calls, yielding a 25.06-point improvement for 1.62 additional image calls on average.


Q3: Cross-Evaluator Assessment

Thank you for highlighting the possibility that the improvement could be specific to the verifier used during interaction. To examine this, we re-evaluated the same 800 final images produced by GenRevisit using two independent, held-out open-weight MLLMs, Qwen3.5-35B-A3B and Qwen3.8-27B.

EvaluatorGenEval2 GM
Qwen3-VL-8B84.33
Qwen3.5-35B-A3B85.84
Qwen3.8-27B82.86

Both held-out evaluators assign similarly high scores to the unchanged final images. The scores differ from the original evaluator by -1.47 to +1.51 GM, providing evidence that the observed performance is not specific to the original Qwen3-VL-8B evaluator.

Our original evaluator is Qwen3-VL-8B, which provides feedback during interaction and is used for the primary evaluation. For cross-evaluator model selection, we were guided by the human-agreement evaluation protocol and findings reported in T2I-CoReBench [1], selecting Qwen3.5-35B-A3B for its high agreement with human judgments. To also account for ongoing model updates, we included Qwen3.8-27B as an additional evaluator. Both models evaluate the same final images.

References

[1] Easier Painting than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal

We sincerely thank all three reviewers for their thoughtful reviews and constructive suggestions. We are encouraged that they recognize the value of GenRevisit from complementary perspectives. Reviewer mWp5 recognizes the clear motivation and highlights that the learned policy captures a “generalizable logic rather than over-fitting a specific generator\textcolor{green}{\text{generalizable logic rather than over-fitting a specific generator}}” and that NACA tackles “a real credit-assignment issue in RL\textcolor{green}{\text{a real credit-assignment issue in RL}}”. Reviewer UYeM recognizes that “Revisiting historical candidates is a sensible approach\textcolor{green}{\text{Revisiting historical candidates is a sensible approach}}” and appreciated the supporting evidence from controlled teacher experiments and cross-generator evaluation. Reviewer XZcR acknowledges that “Historical source selection provides a practical recovery mechanism\textcolor{green}{\text{Historical source selection provides a practical recovery mechanism}}” and appreciates the “Strong reported benchmark results\textcolor{green}{\text{Strong reported benchmark results}}” supported by controlled comparisons. We appreciate these assessments and summarize below the experiments and manuscript revisions that address the reviewers’ concerns.

Reviewer mWp5

  1. Comparison with Explicit Search.\textcolor{blue}{\textbf{Comparison with Explicit Search.}} We introduce a GM-guided Beam Search baseline and show that learned joint source–action selection achieves a higher score with fewer image calls.
  2. Computational Overhead.\textcolor{blue}{\textbf{Computational Overhead.}} We quantify the additional inference costs and show that historical-image conditioning introduces limited token overhead.
  3. Independent Evaluation.\textcolor{blue}{\textbf{Independent Evaluation.}} We re-score the same final images with two independent MLLMs and observe similarly high scores across evaluators.

Reviewer UYeM

  1. Comparison with Closely Related Methods.\textcolor{blue}{\textbf{Comparison with Closely Related Methods.}} We clarified the methodological distinctions between GenRevisit and TwiG, Gen-Searcher, and GenEvolve, and adapted EditThinker's differential-reward and step-wise GRPO approaches into a directly trained baseline for comparison.
  2. Long-Term Credit Assignment Examination.\textcolor{blue}{\textbf{Long-Term Credit Assignment Examination.}} We analyzed the subsequent reuse of zero-credit regressive outputs, clarifying that their resulting images can still be selected as useful continuation states.
  3. NACA Ablations.\textcolor{blue}{\textbf{NACA Ablations.}} We added comparisons with alternative credit-assignment strategies and an ablation of NACA's credit-capping component, further validating the effectiveness of progress-based credit allocation.
  4. Final-Policy Comparison under Different Source-Access Settings.\textcolor{blue}{\textbf{Final-Policy Comparison under Different Source-Access Settings.}} We evaluated the same final RL policy under Joint, Best-only, and Latest-only settings, demonstrating the advantage of learned joint source selection over fixed selection rules.
  5. Controlled External Baseline Comparison.\textcolor{blue}{\textbf{Controlled External Baseline Comparison.}} We reproduced GenRouter under matched generator, feedback, and image-call budget settings. Comparisons within the fixed GenRevisit harness support the contribution of policy learning.
  6. Generalization Beyond the GenEval Benchmark Family.\textcolor{blue}{\textbf{Generalization Beyond the GenEval Benchmark Family.}} We added T2I-CompBench++ evaluation, demonstrating GenRevisit's generalization to an independent compositional image-generation benchmark.
  7. Manuscript Revisions and Reference Formatting.\textcolor{blue}{\textbf{Manuscript Revisions and Reference Formatting.}} We incorporated the additional experiments and analyses into the revised manuscript and standardized the arXiv references, while retaining all previous revisions.

Reviewer XZcR

  1. Joint Source-Selection Effectiveness.\textcolor{blue}{\textbf{Joint Source-Selection Effectiveness.}} We compare the same final RL policy under Joint, Best-only, and Latest-only access, demonstrating the benefit of adaptive historical source selection.
  2. NACA Component Ablations.\textcolor{blue}{\textbf{NACA Component Ablations.}} We evaluate NACA against multiple credit-assignment alternatives and isolate the contributions of its key design components.
  3. Multi-Seed Robustness and Statistical Significance.\textcolor{blue}{\textbf{Multi-Seed Robustness and Statistical Significance.}} We report consistent results across independent NACA training runs and paired statistical evidence for historical-source access in a separate matched-prefix experiment.
  4. Two-Seed Evaluation Clarification.\textcolor{blue}{\textbf{Two-Seed Evaluation Clarification.}} We clarify the distinction between qualitative example selection and the reported GenEval++ evaluation protocol.

We sincerely thank the reviewers for their constructive suggestions, which have helped us further improve the paper's completeness. All these additional experiments and clarifications are incorporated into the revised manuscript. \textcolor{green}{\textbf{All these additional experiments and clarifications are incorporated into the revised manuscript. }}We believe we have sufficiently addressed the main concerns and respectfully ask the reviewers to reassess our work. We thank the reviewers again for their time and effort.

Best,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.