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

Source: arXiv - 2512.05965v1

Overview

The paper introduces EditThinker, a new framework that gives image‑editing models a “thinking” loop: after each edit the system critiques the result, refines the user instruction, and tries again until the output meets the desired goal. By embedding this iterative reasoning process into any existing image editor, the authors dramatically boost instruction‑following success rates, turning single‑shot edits into a more reliable, human‑like workflow.

Key Contributions

• Think‑while‑Edit Loop – a generic, iterative cycle of critique → instruction refinement → re‑generation that can wrap around any image‑editing model.
• EditThinker MLLM – a single multimodal large language model trained to output a critique score, a natural‑language reasoning trace, and an improved instruction in one pass.
• RL‑aligned Reasoning – reinforcement learning aligns the model’s internal “thoughts” with the visual outcomes, encouraging more targeted instruction updates.
• Broad Benchmark Gains – experiments on four diverse editing benchmarks show consistent, large improvements over strong baselines.
• Open‑source Toolkit – the authors release the data‑construction pipeline, curated datasets, and pretrained models for the community.

Methodology

1. Base Editor – any off‑the‑shelf instruction‑based image editor (e.g., Stable Diffusion Instruct‑Pix2Pix) produces an initial edited image from a user prompt.
2. EditThinker Reasoning Engine – a multimodal LLM receives the original image, the user prompt, and the edited result. It simultaneously:
   • Generates a critique score (how well the edit matches the intent).
   • Produces a reasoning trace explaining what went wrong (e.g., “the sky is still over‑exposed”).
   • Emits a refined instruction that corrects the identified issue.
3. Reinforcement Learning Alignment – the critique score is used as a reward signal; the model is fine‑tuned with PPO‑style RL so that its reasoning and instruction updates lead to higher‑scoring edits.
4. Iterative Loop – the refined instruction is fed back to the base editor, producing a new image. Steps 2‑4 repeat until the critique score crosses a preset threshold or a max‑iteration limit is reached.

Because the reasoning engine is a single model, the whole pipeline stays lightweight and can be dropped into existing production pipelines with minimal engineering effort.

Results & Findings

• Higher adherence: The iterative loop consistently pushes the edit quality above the “good enough” threshold, even for ambiguous or multi‑step instructions.
• Explainability: The generated reasoning traces correlate strongly with human judgments, offering a transparent view of why an edit failed.
• Model‑agnostic boost: Swapping the underlying editor (e.g., from Stable Diffusion to DALL‑E‑3) still yields a 20‑30% absolute improvement, confirming the framework’s universality.

Practical Implications

• Developer‑friendly API: Wrap any existing diffusion‑based editor with the EditThinker loop via a simple REST call; no retraining of the heavy image generator is required.
• Reduced QA cycles: Automated critique and instruction refinement cut down manual post‑processing, saving time for content‑creation platforms (e.g., social‑media filters, ad‑creative tools).
• Better user experience: End‑users can issue a single natural‑language command and watch the system “think” and improve the result in real time, mimicking a collaborative designer.
• Debuggable pipelines: The reasoning trace acts as a built‑in log, helping engineers pinpoint failure modes (e.g., color mismatches, layout errors) without manual inspection.
• Enterprise compliance: For regulated industries (e.g., medical imaging), the critique score can serve as a confidence metric before images are approved for downstream use.

Limitations & Future Work

• Iteration cost: Each additional loop incurs extra inference time; real‑time applications may need to cap iterations or use lightweight editors.
• Dependence on critique quality: The RL reward hinges on the automatically computed critique score, which can be noisy for highly subjective edits.
• Generalization to non‑photorealistic domains: While benchmarks cover diverse styles, performance on abstract art or 3D renderings remains untested.
• Future directions: The authors plan to explore adaptive stopping criteria, integrate user feedback as an extra reward signal, and extend the framework to video editing where temporal consistency adds another layer of reasoning.

Authors

• 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

Paper Information

• arXiv ID: 2512.05965v1
• Categories: cs.CV
• Published: December 5, 2025
• PDF: Download PDF