acceptodds
Under review as a conference paper at ICLR 2027

EvoEdit: A Self-Evolving Agent Framework for Complex, Multi-Step Dependent Image Editing

Abstract

With the rapid advancement of generative AI, image editing models have become increasingly powerful and gained significant attention. Despite this progress, existing models and agent systems often remain fragile and struggle to execute complex, especially chain-dependent, image editing tasks. While this vulnerability is partly due to poor planning and a reliance on step-by-step verification, a more fundamental limitation is their inability to learn from accumulated interaction history, causing them to persistently repeat past mistakes. To address these critical bottlenecks, we present EvoEdit, a novel intelligent agent framework designed to solve complex chain-dependent image editing problems, built on the ComfyUI platform. EvoEdit introduces three core innovations: Multi-Path Exploration (MPE), which concurrently explores diverse execution trajectories to identify a high-quality plan; a Recursive Verification mechanism, which backtracks through the execution history to pinpoint the earliest fault responsible for chain failures; and a Dual-Dimensional Self-Evolving mechanism, which dynamically synthesizes macroscopic planning rules and refines microscopic step executions by learning from accumulated experience. Together, these components improve the stability, physical plausibility, and visual harmony of complex interdependent editing workflows. Experimental results demonstrate that EvoEdit significantly outperforms Nano Banana, a strong closed-source image editing model, achieving a 20.68% improvement in average score and a 31.55% increase specifically on three-chain dependent tasks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Image editing, AI agent framework, Self-evolving

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_BLaJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoEdit, a ComfyUI-based agentic image editing framework that combines multi-path exploration (MPE), recursive verification, and dual-dimensional self-evolving to improve complex dependent editing tasks. It leverages GPT-5.4 for planning and reasoning, GPT-4o for verification, and historical experience retrieval to refine future plans and workflow selection. Experiments show promising results on chain-dependent editing tasks and the ImgEdit Hybrid benchmark.

Critical strengths
  1. Practical problem setting. The method targets complex dependent editing tasks and reports promising results on Two-chain and Three-chain settings.

  2. Potentially useful framework. It attempts to improve future planning and workflow selection by accumulating global planning and local execution experiences without updating model parameters.

Critical weaknesses
  1. Insufficient evaluation of verifier reliability. The verifier (GPT-4o) plays a crucial role in assessing editing success and identifying erroneous steps. However, the paper lacks a direct evaluation of its accuracy against human annotations. Such analysis would help establish the reliability of the verification mechanism and its impact on subsequent refinement.
Questions for authors
  1. How accurate is the GPT-4o verifier in assessing editing success and identifying erroneous steps compared with human annotations? Could the authors also investigate whether alternative verification models lead to better performance?
3 comments
AuthorsOpenReview ↗

Further Discussion and Follow-up on Our Response

Dear Reviewer,

We hope this message finds you well. As we are now more than halfway through the discussion period, we wanted to gently follow up and ensure that we have addressed all your concerns satisfactorily. If there are any additional points or feedback you would like us to consider, please let us know while there is still ample time for discussion. Your insights are invaluable to us, and we are eager to address any remaining questions to further improve our work.

Thank you again for your time and effort in reviewing our paper.

Thank you for your response. One remaining concern is the reliance on GPT-4o as the core runtime verifier. As this model may become unavailable in the future, this dependency could affect the long-term usability and reproducibility of your work. I encourage the authors to discuss this limitation.

Thank you for this follow-up. We fully agree this is a valid long-term concern.

We want to emphasize that GPT-4o is not architecturally indispensable to EvoEdit. As demonstrated in our verifier ablation, Gemini-2.5-flash and qwen3-vl-flash serve as drop-in replacements with only modest performance variation, and the core recursive verification logic relies on structured step-level binary judgments that any capable multimodal model can perform. If GPT-4o were deprecated, practitioners could substitute it with any then-available VLM of comparable capability without modifying the framework architecture. All prompts and workflow configurations are already publicly available in our anonymous repository, which further facilitates reproduction with alternative verifiers.

We therefore view this as a manageable engineering dependency rather than a fundamental architectural constraint, and we appreciate your raising it as a point worth highlighting.

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

The paper proposes EvoEdit for multi-step image-editing requests in which later edits depend on earlier ones. The framework runs three different global plans in parallel, verifies the final result, and recursively traces failures back to an earlier editing step. It also stores planning-level and workflow-level experience for use on future tasks.

EvoEdit performs well on ComplexBench-Edit, especially on two- and three-step dependency chains. The paper also reports results on the Hybrid subset of ImgEdit and comparisons with modified versions of ComfyMind and GenArtist.

Critical strengths

The paper studies a real weakness of existing editing systems. Treating a dependent instruction chain as a collection of independent edits often leads to irreversible errors. Exploring several global decompositions and locating the earliest failed step are sensible design choices.

The ablation is useful: performance drops considerably when evolution, multi-path exploration, or recursive verification is removed. I also appreciate the appendix’s attempt to report latency and monetary cost rather than presenting quality improvements without any resource context.

Critical weaknesses

The evaluation relies heavily on a VLM-based 0–5 score and a multiplicative chain metric. There is little human validation of the main ranking, and the paper provides no confidence intervals, task-level significance tests, or results over multiple generation seeds. Because multiplying step scores can magnify relatively small differences, the headline percentage improvements may be sensitive to the evaluator and aggregation rule.

The comparison is also not compute-matched. EvoEdit runs three paths and may repeatedly refine individual steps, whereas Nano Banana appears to receive a much smaller inference budget. The appendix reports that EvoEdit costs roughly 40% more, but it does not compare against giving the baseline additional samples.

Finally, the experience protocol is unclear. I could not determine whether the experience library is frozen before evaluation or updated using earlier benchmark tasks. If the latter, results may depend on task order and include online adaptation unavailable to the baselines.

Questions for authors
  1. Is the experience library frozen before evaluation? If not, how sensitive are the results to task order?

  2. How does EvoEdit compare with baselines under equal cost, latency, and number of generated candidates?

  3. Does blinded human evaluation reproduce the reported ranking?

  4. Are ComfyMind and GenArtist given the same path and refinement budgets?

3 comments
AuthorsOpenReview ↗

Follow-up on Author Response and Further Discussion

Dear Reviewer,

We hope this message finds you well. As we are now more than halfway through the discussion period, we wanted to gently follow up and ensure that we have addressed all your concerns satisfactorily. If there are any additional points or feedback you would like us to consider, please let us know while there is still ample time for discussion. Your insights are invaluable to us, and we are eager to address any remaining questions to further improve our work.

Thank you again for your time and effort in reviewing our paper.

AuthorsOpenReview ↗

Final Follow-up on Our Rebuttal — Less Than 1 Day Remaining (Reviewer JujK)

Dear Reviewer,

We hope this message finds you well. As the discussion period is now coming to a close with less than one day remaining, we are reaching out one final time to kindly follow up on our rebuttal.

We have done our best to provide thorough responses to all of your comments and concerns, and we sincerely hope that our clarifications have been helpful. If there are any remaining questions or points that you feel we have not addressed adequately, we would very much welcome the chance to discuss them before the deadline.

We truly appreciate the time and effort you have dedicated to reviewing our work, and we hope to have the opportunity to hear your thoughts before the discussion window closes.

Thank you very much for your consideration.

Best regards, The Authors

Thanks for the response from the authors, All my concerns are addressed. I raise my score to be positive,

Reviewer_ZLa5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EvoEdit for complex image editing with step dependencies. The method combines multi-path exploration, recursive verification that traces back from final results to historical states, and two-dimensional evolution that reuses historical experience: at the macro level it adjusts step merging and execution order, and at the micro level it improves workflow choices for specific steps. The framework is evaluated on ComplexBench-Edit and the Hybrid subset of ImgEdit, and compared with image editing models and agent systems adapted to the same underlying tools.

Critical strengths

Proposes a task-targeted pipeline design for dependent-chain editing. Multi-path planning allows merging or reordering operations, and recursive verification uses subsequent operations to judge whether an intermediate state truly needs repair. This design targets the problem that local verification may hinder the final goal in complex editing; execution cases in the body and appendix help understand its practical role.

Experiments include tool-adapted agent comparisons and valuable ablations. The authors adapt ComfyMind and GenArtist to the same underlying models or tools as EvoEdit, reducing interference from directly using older tools. The full method achieves good results on the tested tasks across two editing benchmarks; a control that keeps multi-path and recursive verification but removes evolution also shows a clear drop. The appendix further provides parameter selection, latency, and cost analysis, making the real price of performance gains easier to judge.

Critical weaknesses

The boundary between building the experience library and using it for testing is unclear. The paper states that experience is retrieved and reused in later tasks, but does not sufficiently explain which tasks the experience library for the main results comes from, whether it was built and fixed before evaluation, or whether it keeps updating with evaluation tasks. Task order, repeated visits, and the data range used for parameter selection are also unspecified. For work centered on cross-task experience accumulation, this affects judgments of generalization gains and reproducibility; the current information does not warrant concluding that data leakage occurred.

The separate contributions of the two evolution dimensions are not sufficiently distinguished. Existing ablations already support the gain of the overall evolution module, so it should not be called lacking an evolution control. However, macro planning experience and micro execution experience are enabled or removed together, making it hard to tell their individual contributions and whether joint use brings extra gains. Identical multi-path counts and repair limits do not mean identical actual model-call costs; the relationship between the gains of the two evolution types and extra computation still needs explanation.

Key scoring definitions and independent validity validation remain unclear. The paper provides per-instruction scoring prompts, but does not clearly explain how the body's chain-product scores, table aggregate scores, and the average scores used for experience extraction correspond, nor sufficiently explain the relationship between the final result scoring model and the runtime verification and candidate-comparison scoring models. The appendix's small-scale human comparison mainly sets the experience-extraction threshold and cannot directly substitute for human alignment validation of the final benchmark scores. This limits interpretation of numerical gains and self-feedback reliability.

Questions for authors

How is the experience library in the main results initialized, which tasks build it, and when is it updated? Is evaluation an independent test under a fixed experience library, or an online evaluation that continuously accumulates experience in task order? Please explain the boundary between repeated visits, parameter selection, and final testing; if online evaluation is used, please provide existing records or limited checks of task order and its impact.

Can the gains from macro planning experience and micro execution experience be stated separately? Under comparable underlying tools, multi-path, and repair budgets, can a limited-scale control enabling each separately, together with actual model-call and cost statistics, support the complementarity of the two mechanisms?

How exactly are the chain product, the scores in result tables, and the experience-extraction average computed and converted? Is final evaluation independent of runtime feedback and candidate comparison? Can human ranking or an independent evaluator be checked on a small number of existing results to confirm the main advantages still hold?

I'm willing to raise my rating if all problems are solved

4 comments
AuthorsOpenReview ↗

Follow-up on Author Response and Further Discussion

Dear Reviewer,

We hope this message finds you well. As we are now more than halfway through the discussion period, we wanted to gently follow up and ensure that we have addressed all your concerns satisfactorily. If there are any additional points or feedback you would like us to consider, please let us know while there is still ample time for discussion. Your insights are invaluable to us, and we are eager to address any remaining questions to further improve our work.

Thank you again for your time and effort in reviewing our paper.

AuthorsOpenReview ↗

Final Follow-up on Our Rebuttal — Less Than 1 Day Remaining (ReviewerZLa5)

Dear Reviewer,

We hope this message finds you well. As the discussion period is now coming to a close with less than one day remaining, we are reaching out one final time to kindly follow up on our rebuttal.

We have done our best to provide thorough responses to all of your comments and concerns, and we sincerely hope that our clarifications have been helpful. If there are any remaining questions or points that you feel we have not addressed adequately, we would very much welcome the chance to discuss them before the deadline.

We truly appreciate the time and effort you have dedicated to reviewing our work, and we hope to have the opportunity to hear your thoughts before the discussion window closes.

Thank you very much for your consideration.

Best regards, The Authors

Thank you for the detailed responses and additional experiments. The macro-only and micro-only ablations, together with the actual model-call and cost statistics, substantially address my concern about the individual contributions of the two evolution mechanisms. The blind human evaluation also provides useful evidence that the main performance advantage holds beyond the default VLM evaluator. I appreciate these substantive additions.

The clarification of the empty-initialized, single-pass online protocol is helpful. Two remaining details would help me reconcile this description with the experiments in the paper:

  1. Hyperparameter selection and data splits. The response states that the three main hyperparameters were selected using 15 pilot tasks independent of the final evaluation split. However, Figures 8–9 and Table 3 report the same default score of 47.87 as the main Three-chain evaluation, and Table 3 also repeats the no-evolution score of 36.52. Please clarify which data were used for these figures and table, and whether they were produced after the hyperparameters had already been fixed. Please also identify the source and split of the 50 tasks in Figure 10 used to justify the experience-extraction threshold. A brief mapping from each analysis to its data split would be sufficient.

  2. Experience-library resets. Please specify whether the library is reset between benchmark subsets, between ComplexBench-Edit and ImgEdit, and before each shuffled-order run. If any evaluations share accumulated experience, please state their execution order and the scope of that shared history.

Regarding evaluator independence, the response establishes procedural separation between generation and final scoring. The verifier substitution experiment supports robustness to the runtime verifier choice, while the human study more directly addresses concerns about shared evaluator preferences. I suggest moderating the claims of complete independence and minimal preference bias to reflect this distinction.

Overall, the rebuttal substantially alleviates my original concerns. Clarifying the remaining protocol details and incorporating them into the paper would help me finalize my assessment.

AuthorsOpenReview ↗

Revision and response

Thank you for your thorough engagement with our rebuttal. We address the three remaining points below, and summarize the corresponding revisions made to the manuscript.

Hyperparameter selection and data splits. The 15-task pilot set used for initial hyperparameter selection is a small held-out subset drawn from the Three-chain portion of ComplexBench-Edit and is strictly disjoint from the final evaluation split. Figures 8–9 and Table 3 were produced as post-hoc sensitivity analyses on the full Three-chain evaluation split after the hyperparameters had already been fixed; the score of 47.87 appears there because they use the same final evaluation data, not because those figures were used to select the hyperparameters. Figure 10 uses 50 tasks drawn from the same Three-chain evaluation split with all hyperparameters already fixed, solely to visualize the score-difference distribution and justify the threshold value of 2. We have added explicit data-split clarification paragraphs to Appendices A.1, A.2, and A.3 to make this mapping clear to readers.

Experience-library resets. The library is reset to empty at the start of each independent evaluation run. To ensure a fair comparison across subsets, it is also reset between the Parallel, Two-chain, and Three-chain subsets of ComplexBench-Edit, so each subset is evaluated independently without inheriting experience from another. The ImgEdit evaluation likewise begins from an empty library, and each shuffled-order run starts from an empty library with no shared history. We have added an explicit "Experience Library reset policy" paragraph to Appendix C.1 documenting all three reset rules, and updated Appendices A.6 and A.7 to cross-reference it.

Evaluator independence. We agree with your suggestion. We have moderated the language in Appendix A.7 to note that the verifier substitution experiment addresses robustness to runtime verifier choice, while the human evaluation in Appendix A.8 provides the more direct evidence against shared-preference bias at the final scoring stage. We have also updated the Limitations section (Appendix D) accordingly, reporting the risk of long-term reproducibility risk of using the GPT-4o dependency in the reported benchmark scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.