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
2025-12-08
Summary
This paper introduces a new way to make image editing programs better at following your instructions, focusing on making edits that truly match what you want.
What's the problem?
Current image editing programs, while good at creating visually appealing images, often struggle to accurately follow complex instructions. They might get close, but often miss details or make unintended changes because they essentially 'guess' at what to do, and only get one shot at it. This is because image generation is a bit random, and programs don't really 'think' about how to improve their edits step-by-step.
What's the solution?
The researchers created a system called 'EditThinker' that allows the program to 'think' while it edits. It works by going through a cycle: first, it critiques its own work, then it refines the instructions based on that critique, and finally, it tries editing the image again. This 'Think-while-Edit' process repeats until the result is satisfactory. EditThinker is a single program that handles all parts of this cycle – judging the edits, figuring out why they aren't quite right, and rewriting the instructions to be more precise. They used a technique called reinforcement learning to train EditThinker to improve its reasoning and editing skills together.
Why it matters?
This research is important because it significantly improves the ability of image editing programs to understand and follow instructions. This means you'll have more control over the final image and get results that are closer to your vision, and the tools they created will be shared with other researchers to help advance the field.
Abstract
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.