PaperlayerView exact revision on arXiv

Source-linked AI summary

EditThinker: Unlocking Iterative Reasoning for Any Image Editor

Hongyu Li, Manyuan Zhang, Dian Zheng, Ziyu Guo, Yimeng Jia, Kaituo Feng, Hao Yu, Yexin Liu, Yan Feng, Peng Pei, Xunliang Cai, Linjiang Huang, Hongsheng Li, Si Liu

arXiv:2512.05965v1cs.CV

TL;DR

Single-turn image editing leaves limited opportunity to correct intermediate errors, constraining instruction following. EditThinker adds iterative critique and instruction refinement to image editing, and experiments report improved instruction following across four benchmarks.

  • Problem

    Single-turn editing gives models limited opportunity to self-correct intermediate errors, leaving them reactive rather than reflective.

    Source: §1 · p. 2

  • Method

    EditThinker uses an MLLM Thinker to iteratively critique editing results and refine instructions for an existing Editor, with reinforcement learning aligning its reasoning with editing outcomes.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    Experiments across four benchmarks demonstrate that the framework significantly improves instruction-following capability across image editing models.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The reported results support iterative reasoning as a way to improve instruction following across image editing models.

    Source: §6 · p. 9

Abstract

from arXiv · show

Instruction-based image editing has emerged as a prominent research area, which, benefiting from image generation foundation models, have achieved high aesthetic quality, making instruction-following capability the primary challenge. Existing approaches improve instruction adherence via supervised or reinforcement learning, yet single-turn success rates remain limited due to inherent stochasticity and a lack of deliberation. In this work, we propose a deliberative editing framework to 'think' while they edit, which simulates the human cognitive loop by iteratively executing a Think-while-Edit cycle: Critiquing results and Refining instructions , followed by Repeating the generation until satisfactory. Specifically, we train a single MLLM, EditThinker, to act as the reasoning engine of this framework, which jointly produce the critique score, reasoning process, and refined instructions. We employ reinforcement learning to align the EditThinker's thinking with its editing, thereby generating more targeted instruction improvements. Extensive experiments on four benchmarks demonstrate that our approach significantly improves the instruction-following capability of any image editing model by a large margin. We will release our data construction framework, datasets, and models to benefit the community.

1. Introduction

EditThinker reframes instruction-based image editing as an iterative reasoning process: a Thinker critiques results, refines instructions, and repeats generation. The work also introduces a trained reasoning MLLM and the THINKEDIT-140k dataset, reporting consistent improvements across four benchmarks and evaluated editing models.

Introduction

  • Single-turn editors must understand instructions, plan visually, and generate content in one step, leaving no opportunity to correct intermediate errors such as missing attributes.The authors characterize current models as reactive executors rather than reflective thinkers.Source: §1 · p. 2
  • The Think-while-Edit loop critiques editing results, refines instructions around identified deficiencies, and resubmits them for regeneration.Using GPT-4.1 as an expert Thinker improved Qwen-Image-Edit [28], Flux-Kontext, and Omnigen2 [30] without fine-tuning those editors.Source: §1 · p. 2
  • EditThinker is a single MLLM that jointly produces critique scores, refined instructions, and reasoning, using SFT for output format and RL to align planning with editors’ capabilities and failure modes.The framework is intended to implement Think-while-Edit for any image editor.Source: §1 · pp. 2–3
  • THINKEDIT-140k is a large-scale multi-round dataset of high-fidelity source images, diverse editing requests, and detailed reasoning traces for instruction refinement and reasoning-based training.It is constructed through an automated multi-round instruction-refinement framework.Source: §1 · pp. 2–3
  • EditThinker yields consistent performance improvements across four benchmarks, diverse editing scenarios, and evaluated edit models.The paper also reports ablations of the thinking paradigm, reasoning-turn count, training strategy, and expert-thinker choice.Source: §1 · pp. 2–3

2. Related Work

Image editing has progressed from controlled and instruction-tuned methods to stronger general-purpose models, but single-turn instruction following remains limited. Existing reward models improve evaluation and alignment, yet post-hoc feedback can miss intermediate reasoning needed for complex edits.

2.1. Image Editing

  • Single-turn editing remains limited by stochasticity and lack of deliberation, motivating this work’s multi-round instruction-refinement paradigm for improving any editing model.Prior work spans inversion-based and spatial-control techniques, instruction tuning, and robust instruction-tuned or general-purpose multimodal models, each advancing editing capabilities while leaving instruction-following challenges.Source: §2.1 · p. 3

2.2. Reward Models for Image Editing

  • Specialized reward models advanced editing evaluation: EditReward [33] trained on human preferences, and EditScore [20] surpassed general-purpose VLM judges to enable online RL.Earlier discrete-score [9] and dense logit-based [31] signals often missed the fine-grained nuances needed for precise visual modifications.Source: §2.2 · p. 3
  • Post-hoc reward feedback cannot correct intermediate logic in complex edits [44], prompting MLLMs to serve as internal planners [19] [38] rather than external judges.This work shifts from maximizing static post-hoc rewards to using structured MLLM reasoning to guide editing during execution.Source: §2.2 · p. 3

3. Think-while-Edit

Think-while-Edit turns single-pass image editing into a human-inspired cycle of critique, reflection, and revision. EditThinker evaluates edits and refines instructions, while reinforcement learning grounds its reasoning in the outcomes produced by real editors.

3.1. Overall Framework

  • The framework replaces single-turn editing with a Critique-Refine-Repeat cycle in which a Thinker scores and reasons about each result, refines the instruction, and an Editor executes it until the goal is achieved.The Editor can be any existing image-editing model.Source: §3.1 · p. 3, §3 · p. 3

3.2. Design of the EditThinker

  • EditThinker combines evaluation and planning in one forward pass, generating its refined instruction only after reasoning from a critique of the previous edit.At each iteration it receives the source image, current edit, original instruction, and previous instruction, and outputs reasoning, scores, and a refined prompt; the two scores range from 0 to 10.Source: §3.2 · pp. 3–4

3.3. Training of EditThinker

  • Training uses a two-stage strategy: supervised fine-tuning teaches the structured output and basic reasoning, then reinforcement learning optimizes refinement using editing feedback.The data construction process is described in Section 4.Source: §3.3 · p. 4

3.3.1. Supervised Fine-Tuning (Cold Start)

  • Using GPT-4.1 expert demonstrations, supervised fine-tuning teaches the model the structured I/O format, expert-style reasoning, and principles for critiquing and refining instructions.Source: §3.3.1 · p. 4

3.3.2. Reinforcement Learning Tuning (RLT)

  • Because SFT reasoning may not reflect real editors’ limitations, GRPO uses editing feedback to train both more accurate critique and instruction refinement grounded in execution outcomes.The critic reward penalizes score prediction error against GPT-4.1 evaluation, while the differential edit reward is positive only when the refined instruction produces an improved image.Source: §3.3.2 · pp. 4–5

4. THINKEDIT Dataset

THINKEDIT-140k is built through a four-stage pipeline that generates and filters iterative editing trajectories, extracts balanced step-wise samples, and splits data for SFT and RL.

4. THINKEDIT Dataset

  • The four-stage construction pipeline comprises Trajectory Generation, Trajectory Filter, Step-wise Filter, and Data Partition, producing THINKEDIT-140k.The pipeline is designed to simulate the multi-turn Think while Edit cycle.Source: §4 · p. 5

4.1. Trajectory Generation

  • GPT-4.1 iteratively evaluates the source and current edit, generates refined instructions and reasoning, and continues editing until a stop token or maximum iteration limit N.The expert does not output scores because it struggles to maintain high performance in both task refinement and output scoring simultaneously.Source: §4.1 · pp. 5–6

4.2. Trajectory Filter

  • An edit scorer assigns step scores; trajectories are retained only if a later step scores at least as high as the initial step, then truncated at their highest-scoring step.Trajectories failing max(St>1) ≥ S1 are discarded, and steps after the best-scoring step are removed.Source: §4.2 · p. 6

4.3. Step-wise Filter

  • Each retained trajectory step becomes a training sample pairing source and prior-edit inputs with expert reasoning and refined-instruction outputs; samples are balanced by task type and score level.The step score is retained as metadata for subsequent filtering.Source: §4.3 · p. 6

4.4. SFT and RL Data Split

  • 10k high-variance trajectories yielded 27k RL step-wise samples, while low-variance or consistently high-quality examples formed 140k SFT step-wise samples.RL trajectories were selected for high intra-trajectory score variance (Var(St) > θ); SFT examples were selected for stable refinement behavior.Source: §4.4 · p. 6

5. Experiments

EditThinker’s Think-while-Edit framework improves the reported general and reasoning-based editing results across the tested base models. The experiments also examine its setup, iterative reasoning, training stages, and expert-model choices.

5.1. Experimental Setup

  • The evaluation covers ImgEdit-Bench and GEdit-Bench-EN for general editing, and RISE-Bench and KRIS-Bench for reasoning-based editing.EditThinker is built on Qwen3-VL-8B-Instruct [1], trained with one SFT epoch and one RL epoch, and tested in the Think-while-Edit paradigm with OmniGen2, Flux Kontext [dev], and Qwen-Image-Edit.Source: §5.1 · p. 6, §5.2 · p. 6

5.2. Main Results

  • 3.44 to 3.98 on ImgEdit-Bench Overall for FLUX.1-Kontext [Dev] with EditThinker, with gains also reported for OmniGen2 and Qwen-Image-Edit.On ImgEdit-Bench Overall, OmniGen2 rises from 3.4 to 3.5 and Qwen-Image-Edit from 4.36 to 4.37. On GEdit-Bench-EN, FLUX.1-Kontext [Dev] rises from 6.18 to 7.05, OmniGen2 from 6.19 to 6.28, and Qwen-Image-Edit from 7.49 to 7.73.Source: §5.2 · pp. 6–7, Table 1 · p. 7
  • 5.8 to 14.4 on RISE-Bench for FLUX.1-Kontext [Dev] with EditThinker, with reported gains for OmniGen2 and Qwen-Image-Edit as well.OmniGen2 rises from 3.1 to 3.4, and Qwen-Image-Edit from 8.9 to 17.8. RISE-Bench assesses complex spatial, causal, and temporal reasoning.Source: Table 2 · p. 7, §5.2 · p. 7
  • 4.13 versus 3.98 on ImgEdit-Bench for FLUX.1-Kontext [Dev] with the stronger GPT-4.1 expert versus EditThinker-8B, respectively.The paper reports that the same pattern holds across other models and benchmarks, linking greater performance enhancement to a more capable expert.Source: §5.2 · p. 7, Table 1 · p. 7

5.3. Ablation Study

  • 7.19 to 7.06 on GEdit-Bench-EN when Think-while-Edit is initialized with a Think-before-Edit step, while Think-while-Edit alone outperforms Think-before-Edit.The authors hypothesize that the initial step biases first-round reasoning, causing incomplete information transfer and poorer downstream performance.Source: §5.3 · pp. 7–8, Table 3 · p. 8
  • 6.18 to 7.30 G O on GEdit-Bench-EN as the maximum allowed turns increase from one to eight.The reported scores are 6.95 at two turns, 7.13 at four turns, and 7.16 at six turns; the authors attribute the scaling to iterative error correction.Source: Table 4 · p. 8, §5.3 · p. 8
  • 6.18 to 6.93 G O on GEdit-Bench-EN with SFT alone, followed by 7.02 with RL; on ImgEdit-Bench Overall, SFT raises 3.44 to 3.57 and RL raises 3.57 to 3.95.The authors describe SFT as providing foundational refinement capabilities and RL as optimizing the expert’s judgment and decision policy.Source: §5.3 · p. 8, Table 5 · p. 8
  • 6.00 to 7.19 G O on GEdit-Bench-EN when the baseline FLUX model uses GPT-4.1 as its expert instead of EditThinker-8B.The result supports the paper’s claim that the framework can use different experts and that performance benefits from stronger expert reasoning and critique.Source: Table 6 · p. 8, §5.3 · p. 8

6. Conclusion

EditThinker enables image editing models to think while they edit through an iterative cycle of critique, instruction refinement, and repeated generation. A single MLLM produces critiques, reasoning, and refined instructions, with reinforcement learning aligning its reasoning to editing outcomes; experiments show gains across four benchmarks.

6. Conclusion

  • EditThinker simulates human deliberation through repeated critique, instruction refinement, and generation until outcomes are satisfactory.The framework addresses limited instruction-following capability attributed to stochasticity and lack of deliberation in existing single-turn approaches.Source: §6 · pp. 8–9
  • Experiments on four benchmarks show that EditThinker significantly enhances the instruction-following capability of any image editing model.The framework, datasets, and models will be released to benefit the research community.Source: §6 · p. 9
  • A single MLLM jointly produces critique scores, reasoning, and refined instructions, while reinforcement learning aligns its reasoning with editing outcomes.This alignment enables more targeted instruction improvements.Source: §6 · p. 9

Supplementary Material · A. Kris-Bench Result

On Kris-Bench, the method improves performance on reasoning-driven edits, with gains in Overall Score across three image editing models.

A. Kris-Bench Result

  • Overall Score rises from 61.81 to 69.53 for FLUX.1 Kontext [Dev], from 50.52 to 53.09 for OmniGen2, and from 64.43 to 71.91 for Qwen.These Kris-Bench results measure reasoning-driven editing capability.Source: A. Kris-Bench Result · p. 11

B. More Ablation Analysis … C.2. Details of EditThinker

EditThinker improves through repeated editing turns, while its expert and inference prompts train a shared critique-and-revision process. At inference, it uses the current editing state to produce a score, reasoning trace, and refined instruction.

B. More Ablation Analysis

  • EditThinker-8B’s score rose from 6.18 at baseline to 7.03 at Turn 8, with the largest jump—from 6.18 to 6.90—at Turn 2.The initial prompt typically performs worst, so its first refinement brings the most direct improvement; later turns reflect further refinement.Source: B. More Ablation Analysis · p. 11

C.1. Details of EditThinker Expert

  • GPT-4 supervises EditThinker with reasoning traces and refined instructions, generating a trace, revised instruction, and success flag from the editing state at each iteration.The expert process is capped at N = 5 iterations.Source: §C.1 · p. 11
  • The expert prompt directs a Critique–Revise cycle: assess instruction fulfillment, reason about failure causes, and improve the instruction without introducing inconsistencies.The full expert prompt template is provided in Figure 6.Source: §C.1 · p. 11

C.2. Details of EditThinker

  • EditThinker uses one prompt format for training and inference, aligning supervised behavior with the capabilities needed to evaluate results, reason about issues, and refine instructions.Source: §C.2 · p. 11
  • At each iteration, EditThinker receives the source image, user instruction, previous edited image, and the instruction that produced it, with a maximum of N = 5 iterations.Source: §C.2 · p. 11
  • EditThinker outputs an instruction-following score, reasoning trace, and refined instruction; its continuous Semantic and Quality Scores support nuanced supervision and threshold-based refinement control.The full prompt template is provided in Figure 7.Source: §C.2 · p. 11

D. Details of ThinkEdit-140K Dataset

ThinkEdit-140K combines editing instances from three sources and converts them into filtered trajectories and high-quality SFT and RL training data.

D. Details of ThinkEdit-140K Dataset

  • ThinkEdit-140K draws 40K, 40K, and 60K editing instances from OpenGPT-4o-Image, ShareGPT-4o-Image, and Pico-Banana-400K, totaling 140K raw samples.The samples were selected to distribute editing categories as evenly as possible.Source: D. Details of ThinkEdit-140K Dataset · p. 11
  • GPT-4.1 served as the EditThinker-Expert, with OmniGen2, FLUX.1 Kontext [Dev], and Qwen-Image-Edit as editors; trajectory filtering retained 70K valid trajectories.The samples were divided into three splits.Source: D. Details of ThinkEdit-140K Dataset · p. 11
  • Of the retained trajectories, 10K were selected for RL, while the remaining 60K underwent step-wise refinement and filtering to produce 140K SFT samples.Additionally, a subset of 27K filtered trajectories was used for RL training.Source: D. Details of ThinkEdit-140K Dataset · p. 11

E. Visualization

The visualizations show EditThinker producing edits with multiple editors and iteratively critiquing results to refine instructions. They also present the prompts that define the Thinker’s and Expert’s inputs and outputs.

E. Visualization

  • EditThinker produces high-quality edits with FLUX.1 Kontext [Dev], OmniGen2, and Qwen-Image-Edit.Figure 4 shows qualitative outputs paired with all three editors.Source: E. Visualization · p. 12, Figure 4 · p. 13
  • With FLUX.1 Kontext [Dev], EditThinker critiques intermediate results and progressively refines instructions over multiple rounds.Figure 5 visualizes the reasoning traces and intermediate edits in this iterative process.Source: E. Visualization · p. 12, Figure 5 · p. 14
  • The Expert prompt takes the source image, previous edit, source instruction, and prior refined instruction, then produces a stop flag, reasoning, and refined instruction.Figure 7’s unified Thinker prompt uses those inputs to produce semantic and quality scores, reasoning, and a refined instruction.Source: Figure 6 · p. 15, Figure 7 · p. 16

Related papers

Loading 2512.05965v1…