acceptodds
Withdrawn from ICLR 2027

R2E-Prompt: Reasoning-to-Executable Prompt Generation for Implicit-Instruction Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present R2E-Prompt, a training-free framework that formulates implicit-instruction image editing as a structured reasoning task in language space. Existing editors translate explicit commands faithfully but struggle with abstract conditions whose visual consequences are not directly named, leaving the entailed changes unrealized in the output. R2E-Prompt addresses this by moving reasoning entirely upstream of the editor: a VLM and an LLM produce, in a single pass, an explicit element-level edit specification consumable by any off-the-shelf instruction-based editor. The framework comprises three stages: Target-State Reasoning first grounds the source image in a source caption and then derives, via chain-of-thought reasoning, an explicit target caption that makes the implicit visual consequences of the instruction explicit; Edit-Aware Scene Graph Transformation parses the source and target captions into text-derived scene graphs and casts source-to-target editing as a structured prediction problem over scene-graph elements; and Executable Prompt Generation linearizes the labeled graph into a natural-language editing command. Without architectural modification or fine-tuning, R2E-Prompt achieves state-of-the-art performance among open-source baselines on RISEBench.

Primary area:
generative models
Keywords:
Reasoning-based Image Editing, Instruction-based Image Editing, Training-free, Vision-Language Models, Large Language Models

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_6adWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims at the implicit‑instruction image editing problem and presents R2E‑Prompt, a training‑free, single‑pass, editor‑agnostic framework. The pipeline consists of three sequential language‑space reasoning stages:

  1. Target‑State Reasoning to infer the explicit target visual caption from source image and implicit user instruction;
  2. Edit‑Aware Scene Graph Transformation to build element‑level edit specifications via labelled scene graph nodes;
  3. Executable Prompt Generation to linearize graph representations into natural‑language editing prompts for off‑the‑shelf image editors. Evaluations are conducted on RISEBench, RE‑Edit and ImgEdit‑Bench. Results show consistent improvements across multiple open‑source editing backbones, narrowing the performance gap against several leading closed‑source models. The authors also provide ablation studies, editor transferability analysis, human user study and extensive supplementary system prompts and pipeline traces.
Critical strengths

S1: This work locates a pain point in practice: existing instruction‑guided image editors fail heavily on implicit instructions that require inferring visual consequences rather than literal word matching. Shifting all reasoning upstream into language space without fine‑tuning editing models is a clean, appealing technical paradigm. S2: Extensive experiments demonstrate consistent gains across diverse off-the-shelf image editors (Qwen‑Image‑Edit, multiple FLUX‑2 variants and some other recent models). Both automatic benchmarks and human user study verify the effectiveness of this prompt‑level plug‑and‑play intervention. S3: The three-stage pipeline makes reasoning interpretable and trackable. The authors release full system prompts and concrete pipeline traces, improving reproducibility.

Critical weaknesses

W1: Even with R2E‑Prompt, the logical reasoning category accuracy on RISEBench remains very low (11.8%). The paper attributes it to joint limits of upstream LLM and downstream image editor, but lacks targeted diagnostic analysis or preliminary mitigation attempts. W2: Although single‑pass and faster than iterative baselines like EditThinker, the VLM+LLM reasoning steps bring non‑negligible additional computational cost per sample. This work does not explore lightweight distillation or optimization directions. W3: Errors from upstream LLM and VLM will propagate to the final editing prompt. The paper does not systematically characterize failure cases or try some self‑checking or self‑correction mechanisms to catch or avoid bad intermediate outputs.

Questions for authors

Q1: If a very strong reasoning VLM model is directly employed to directly transfer source image, instruction prompt to the final prompt(in other words, in place of all of the three stages), how about the performance compared with the proposed R2E-prompt?

Q2: In your ablation comparing against naive prompt‑engineering baselines, scene‑graph transformation brings clear gains for spatial reasoning. Could you elaborate on which specific scene‑graph properties contribute most to spatial performance improvement?

Reviewer_2xTLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes R2E-Prompt, a framework that uses LLMs to translate implicit image-editing instructions into explicit, element-level editing specifications. The pipeline consists of target-state reasoning, scene-graph transformation with edit-operation labels, and final editing-instruction generation. The method improves performance across multiple editors and reasoning-based editing benchmarks without task-specific training.

Critical strengths
  • The paper is well-structured and logically organized, making it easy to follow and understand. The overall presentation is clear and coherent.
  • The paper provides extensive visualization results, which offer valuable qualitative insights and help illustrate the effectiveness of the proposed method.
Critical weaknesses
  • Paper provides insufficient prompt-level comparisons. The authors should provide the complete prompts used by the baselines and ablation variants, alongside comparisons of the final editor-facing prompts generated for the same inputs.
  • The contribution of explicit edit labels is not sufficiently isolated. Controlled comparisons with and without these labels are needed to attribute the gains to the proposed constraints. Paired qualitative examples where the variant without scene graphs fails while the full method succeeds would also help explain the mechanism.
  • The expressiveness of the operation taxonomy remains unclear. Can composite operations that are not explicitly included, such as swapping two objects, be reliably decomposed and executed using the predefined operations?
  • Scalability to complex scenes is not verified. As the number of instances and relations increases, does the method suffer from growing graph complexity, instance confusion, or information truncation?
  • The source caption may omit task-critical visual information. Since source captioning is not conditioned on the editing instruction, it may overlook relevant instances or attributes. For example, in the lunar-phase case in Fig. 11, the source caption does not specify the current lunar phase, yet the target caption predicts a specific subsequent change. The visual basis for this inference should be clarified.
Questions for authors

Please see weaknesses.

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

This paper addresses implicit-instruction image editing, where the desired visual changes must be inferred from a condition or event. R2E-Prompt is a training-free pipeline that generates source and target captions, constructs an edit-aware scene graph, and converts it into a natural-language instruction specifying changes and preservation requirements. A frozen downstream editor then executes this instruction once. On RISEBench, the method improves Qwen-Image-Edit’s overall success rate from 19.4% to 31.4%, with the largest gains in temporal and causal reasoning. Ablations, prompt-engineering baselines, transfer experiments across seven editors, RE-Edit evaluation, and a user study provide supporting evidence.

Critical strengths
  1. This paper aims to mitigate the problem that executing an implicit instruction requires inferring its visual consequences, rather than simply following its literal wording. Separating target-state inference from edit specification provides a reasonable mechanism for addressing this difficulty. The resulting natural-language interface allows the framework to improve existing editors without architectural changes or additional training.

  2. On RISEBench, R2E-Prompt improves Qwen-Image-Edit’s overall accuracy from 19.4% to 31.4%, with particularly substantial gains in temporal and causal reasoning. Appendix C reports consistent overall improvements across seven editor backbones, and Appendix D provides additional validation on RE-Edit. Together, these results support the practical value of upstream reasoning across the evaluated implicit-editing settings.

Critical weaknesses
  1. UniEdit-I [1] already explores the use of source and target captions together with scene-graph updates. The paper should discuss the differences between the two methods in greater detail and include an experimental comparison with UniEdit-I.

  2. The method relies on additional VLMs and LLMs for understanding and reasoning, rather than integrating these capabilities into the editing model itself. The limitations of this decoupled design require further discussion, particularly how errors from these external components affect the final edits.

  3. How does R2E-PROMPT perform on reasoning-aware editing tasks involving implicit instructions, such as those in Reason50K [2]?

[1] Bai C, Chen J, Bai X, et al. Uniedit-i: Training-free image editing for unified vlm via iterative understanding, editing and verifying[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026

[2] He, Q., Chen, X., Wang, C., Pan, Y., et al. Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning[C]// Proceedings of the 43rd International Conference on Machine Learning, 2026

Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.