acceptodds
Withdrawn from ICLR 2027

EditThinker: Enhancing Image Editors via Iterative Reasoning with a Plug-and-Play VLM

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image editors can produce realistic images while still failing to satisfy the requested changes. Improving their instruction following requires identifying these failures and translating visual feedback into instructions that the editor can execute. We introduce EditThinker, an 8B vision-language model (VLM) that enhances existing image editors through iterative reasoning while keeping their weights fixed. EditThinker jointly critiques semantic consistency and visual quality, reasons about editing failures, and refines the instruction in a structured response. It interacts with the editor through a Critique-Refine-Repeat loop, providing a plug-and-play interface for feedback-driven editing. Learning effective refinements requires more than imitating plausible expert suggestions: the instructions must work with the capabilities of the underlying editor. We therefore combine supervised fine-tuning on expert demonstrations with reinforcement learning from executed editing outcomes, using format, critic, and edit rewards. To support this training, we construct THINKEDIT-167k, a dataset of 167k step-wise samples with unified supervision for critique, reasoning, and instruction refinement. Experiments on four benchmarks show improved overall instruction-following scores across the tested editors. Ablations examine the joint model design, data allocation, and training across VLM backbones. Our round-budget study further shows a favorable quality-editor-call trade-off relative to best-of-N selection. We will release our data construction framework, datasets, and models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Instruction-based Image Editing, Iterative Reasoning, Instruction Refinement, Multimodal LLM

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

This paper proposes EditThinker, a plug-and-play image-editing framework built around an 8B vision-language model (VLM). Given the source image, original editing instruction, and current edit, the Thinker generates an analysis of editing failures, semantic-consistency and perceptual-quality scores, and a refined instruction. The underlying image editor remains frozen and applies each revised instruction to the original source image. The authors construct the ThinkEdit-167k dataset and train the Thinker using supervised fine-tuning (SFT) on expert demonstrations followed by GRPO with rewards based on actual editing outcomes. Experiments across four image-editing benchmarks and multiple editors report improvements in aggregate performance, along with ablations on reward configurations, model design, data partitioning, and iteration budgets.

Critical strengths
  1. A practically relevant problem and a reusable framework. The paper addresses cases in which edited images appear plausible but fail to satisfy the requested changes. Refining instructions through visual feedback without modifying the underlying editor is a potentially useful and broadly applicable approach.

  2. Broad evaluation with targeted ablations. The authors evaluate the framework across multiple editors and four benchmarks, and examine SFT/RL, the joint Critic-Refiner design, output ordering, and iteration budgets. The reported results support the overall effectiveness of the framework on several editing tasks.

Critical weaknesses
  1. The distinction from recent related work is not sufficiently clear. MIRA (Zeng et al., CVPR 2026 Findings) also uses a plug-and-play VLM, visual feedback, and SFT+GRPO for iterative image editing. The authors should clarify the specific benefits of the joint Critic-Refiner design and full-instruction rewriting, ideally through comparisons using the same editor and editor-call budget.

  2. The critique and refinement capabilities are not directly evaluated. As the Introduction states, “A VLM may correctly describe an editing error yet suggest a revision that the same editor cannot realize” (Section 1, p. 2). However, the experiments mainly report final editing quality. It would be useful to measure the prediction error of the Thinker's score output, as well as the improvement rate and magnitude after executing its answer instruction.

  3. The distinct role of the Edit Reward remains unclear. If candidates in a GRPO group share the same previous editing result, the quality-difference reward may be equivalent to an absolute-quality reward after standard group-wise normalization. The authors should clarify the implementation, compare the two reward formulations, and report single-step improvement rates before and after RL.

  4. Quantitative analysis of per-round editing quality is missing. Table 3(b) compares iteration budgets and Best-of-N selection, but does not show whether quality improves consistently across rounds. Given the degradation examples in Appendix G, the authors should report improvement/degradation rates and the score gap between the best intermediate result and the final result.

  5. End-to-end inference costs are not quantified. The authors acknowledge that editor-call counts exclude VLM computation. Reporting the end-to-end latency of the Thinker and Editor, the time spent in each component, and the hardware configuration would make the efficiency analysis more complete.

Questions for authors

None

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

EditThinker trains an 8B VLM through SFT and GRPO to critique editing results and refine instructions for fixed image editors (§§2–3). Its main contribution is an editing-specific training and feedback system.

Critical strengths

Fixed editors remove generator adaptation as a confound. Table 3a supports the usefulness of Thinker training and the quality–iteration trade-off. Table 3b compares against best-of-N selection; Table 4 confirms improvements under a non-GPT evaluator. Larger reasoning-editing gains support practical significance: RISE Qwen 8.9→17.8, LongCat 11.9→22.2 (Table 2).

Critical weaknesses
  1. Incremental novelty — partial evidence. Critique–refine–repeat, expert distillation, and GRPO are established ingredients, as acknowledged in Appendix A.2. The experiments support their editing-specific integration, but do not establish a distinct reasoning paradigm.
  2. Incomplete reasoning attribution — missing control. Table 3b provides sufficient evidence against the tested best-of-N baseline. Table 3c supports the evaluated reasoning-first ordering, but does not establish that explicit reasoning content is necessary. A rewriter trained on the same examples, with matched training and inference budgets but without explicit reasoning, remains needed.
  3. Efficiency and gain size — partial evidence. Fewer editor calls are demonstrated, but end-to-end efficiency is unestablished: §4.3 explicitly excludes VLM computation from these counts. Strong-editor ImgEdit gains—Qwen 4.36→4.41, LongCat 4.53→4.57—are small, and their reliability requires uncertainty estimates or repeated runs (Table 1). The larger Table 2 gains prevent dismissing the overall contribution as negligible. 43885_EditThinker_Enhancing_Im
Questions for authors
  1. What substantive methodological or empirical insight distinguishes EditThinker from existing refinement systems beyond task-specific data construction and SFT/GRPO adaptation?
  2. Does EditThinker outperform a direct rewriter without reasoning traces when backbone, inputs, training data, rewards, and training/inference budgets are matched?
  3. Do the advantages persist when total latency or compute includes VLM inference, and are the small strong-editor gains reliable under repeated runs and paired confidence intervals?
Reviewer_1abhOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an image editing framework that takes a user instruction of how to editing image, iteratively refine it, and finally leads to a better prompt that used to produce faithful and high-quality image editing result. The editing framework has two components: EditThinker, the key model being trained in this framework, capable of evaluating the current generated results and propose edited prompts for the editor, and a editing model, which is fixed in the framework. To train the EditThinker, the paper propose to first do SFT on high-quality dataset, then perform RL which leverages GPT to evaluate whether EditThinker’s predicted evaluation scores agree with GPT's scores and whether the generated image improves compared with last iteration. The paper compares the editing results with a series of baselines and leverages ablation studies to show the competitiveness of the editing framework and the trained model.

Critical strengths
  • The experiment plan is very sound. To show the benefits of the EditThinker, the experiment needs to show that (A) the entire editing framework, as a blackbox, outperforms strong instruction-based image editing model, and (B) within the editing framework, changing the trained think model to the other zero-shot open-source/close-sourced model will degrade the performance. The paper tries to conduct all these experiments with a solid experiment plan.
  • Trying to obtain a good prompt writer can usually add benefits on top of the current image editing model. This work lies in a different research line from most of the work that studying how to improve the image quality/editing faithfullness to instruction, etc.
  • A relatively large-scale training dataset, THINKEDIT-167k.
Critical weaknesses
  1. Training objective is not proper. The current training relies on GPT evaluation on semantic consistency and image quality, and when measuring semantic consistency, the reward function relies on the original instruction of the user. However, the original instruction itself can be incomplete. Figure 1 last row is a good example: the original instruction is just "Animate the cat in the image". It does not mention how the edited image should behave in terms of background, cat's texture, cat's actions, etc. With that, it is GPT-4.1 that determine whether the edit image aligns with the sentence "Animate the cat in the image" or not, and GPT-4.1 has its own bias on this. In other words, maximizing the reward means we have a thinker that mostly aligns with GPT-4.1 in terms of interpreting the instruction when the instruction does not clearly say some features, while different users may have different opinions on this (to me, I think the background should roughly not changed, and the cat's actions should be kept, which means the final generated image is a bad example to me personally).
  2. Model is outdated. The paper still uses GPT-4.1 to develop the model, but now we have GPT-6. Why using this model? I think we should roughly agree that GPT-4.1 definitely has drawbacks compared with today's model.
  3. Training highly relies on close-source models. During GRPO process, it needs to query GPT-4.1 for each reward generation. This is a huge cost. This also means the method can be difficult to evolve - for other works, you can evolve by simply adopting a stronger baseline models and better training datasets/methods, but for this work, it means queries the latest GPT model again during training, which suffers from the large cost again. In other words, this training method seems not persistent and cannot contribute to further method's training.
  4. Outdated baseline. When compare with the baseline thinker model, the model is comparing with GPT-4.1, Geminni-2.5-Pro, GPT-4o, etc., which are also out-dated. I actually doubt that with today's GPT-6/Qwen3.6+ the thinker is quite direct and no need to be trained. Again, why compared with this out-dated model?
  5. Lack of human evaluation on the generated results.
  6. Performance. The trained model is not consistently better than the direct editing method like UniWorld-V2, nor better than using stronger thinker like GPT-4.1. This seems mean the model cannot get state-of-the-art result in either dimension of comparison.
  7. Inference time need. With iterative editing, how long in terms of time should the user wait to get the final results? Compared with directly using the image editing model, would the time needed be significantly longer?
Questions for authors

The critical weaknesses I mentioned above are all my major concerns. In addition, I am actually worried about the current relevance of this work. Many baselines are outdated and the external reward model is outdated. Besides, the methodology behind this work, that to get a good instruction-based edited image, we need to have a systematic way to have better prompt, is also established. Therefore it seems this work could be a good work one year ago, but I am hesitant to conclude that this work offers sufficient value to today’s audience.

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

Related papers

Discussion (0)

Sign in to comment.