Source-linked AI summary
ReasonEdit: Towards Reasoning-Enhanced Image Editing Models
Fukun Yin, Shiyu Liu, Yucheng Han, Zhibo Wang, Peng Xing, Rui Wang, Wei Cheng, Yingming Wang, Aojie Li, Zixin Yin, Pengtao Chen, Xiangyu Zhang, Daxin Jiang, Xianfang Zeng, Gang Yu
TL;DR
ReasonEdit addresses limited visual reasoning in instruction-based image editing by unlocking MLLM reasoning through Thinking and Reflection. Its interleaved reasoning-and-editing framework interprets abstract instructions and iteratively corrects results, achieving state-of-the-art performance among open-source methods across multiple benchmarks.
Problem
Frozen MLLM encoders limit visual reasoning for complex or abstract editing instructions and restrict benefits from test-time scaling.
Method
ReasonEdit combines Thinking for abstract-to-actionable instruction conversion with Reflection for iterative self-correction in an MLLM–DiT editing framework.
Results
ReasonEdit achieves state-of-the-art performance among open-source methods while remaining highly competitive with several closed-source models.
Takeaways & Limitations
Structured instruction understanding and iterative self-correction support image editing with high fidelity and consistency across simple and complex tasks.
Abstract
from arXiv · showhide
Recent advances in image editing models have shown remarkable progress. A common architectural design couples a multimodal large language model (MLLM) encoder with a diffusion decoder, as seen in systems such as Step1X-Edit and Qwen-Image-Edit, where the MLLM encodes both the reference image and the instruction but remains frozen during training. In this work, we demonstrate that unlocking the reasoning capabilities of MLLM can further push the boundaries of editing models. Specifically, we explore two reasoning mechanisms, thinking and reflection, which enhance instruction understanding and editing accuracy. Based on that, our proposed framework enables image editing in a thinking-editing-reflection loop: the thinking mechanism leverages the world knowledge of MLLM to interpret abstract instructions, while the reflection reviews editing results, automatically corrects unintended manipulations, and identifies the stopping round. Extensive experiments demonstrate that our reasoning approach achieves significant performance gains, with improvements of ImgEdit (+4.3%), GEdit (+4.7%), and Kris (+8.2%) when initializing our DiT from the Step1X-Edit (ReasonEdit-S), and also outperforms previous open-source methods on both GEdit and Kris when integrated with Qwen-Image-Edit (ReasonEdit-Q).
1. Introduction
ReasonEdit addresses limited reasoning in instruction-based image editing by combining Thinking for abstract instruction interpretation with Reflection for iterative result correction.
- 1. Introduction: Current multimodal editing models often freeze their MLLM encoders, limiting visual reasoning for complex or abstract instructions.This limitation also restricts their ability to benefit fully from test-time scaling.
- 1. Introduction: ReasonEdit introduces Thinking and Reflection to improve instruction understanding and editing accuracy.Thinking converts ambiguous instructions into actionable directives, while Reflection supports iterative self-correction and termination.
- 1. Introduction: Thinking decomposes abstract instructions into clear commands, while Reflection iteratively refines intermediate images toward more accurate results.The two mechanisms respectively target instruction interpretation and output auditing or correction.
- 1. Introduction: ReasonEdit trains an MLLM Reasoner and DiT Generator through separate initial tasks followed by joint training.The staged strategy begins with image editing and thinking tasks before unified optimization.
2. Related Work
Related work progresses from mask-based and instruction-driven editing toward multimodal reasoning, but ReasonEdit specifically targets thinking and reflection for image editing.
- 2. Related Work: Earlier editing methods rely on masks, while later systems incorporate multiple visual conditions or natural-language instructions for more flexible control.These developments span localized mask-based editing, multimodal conditioning, and purely instruction-driven systems.
- 2. Related Work: Instruction-based editing lowers the usage threshold but often struggles to align semantic understanding with fine-grained manipulation.Existing systems also continue to face fidelity and quality challenges.
- 2. Related Work: Frozen MLLM encoders limit performance on complex or abstract instructions, motivating joint optimization with the diffusion decoder.ReasonEdit uses this optimization to unlock MLLM reasoning for image editing.
- 2. Related Work: Visual-generation research has explored thinking before generation, reflection-based refinement, and multimodal chain-of-thought reasoning.Examples include BAGEL, OmniGen2, Reflect-DiT, Image-CoT, MINT, and IRG.
- 2. Related Work: ReasonEdit focuses on thinking and reflection mechanisms for editing, distinguishing it from broader reasoning-enhanced visual generation approaches.The paper states that its base models, training-data composition, and training paradigm also differ from concurrent Uni-CoT.
3. Method
ReasonEdit builds reasoning-enhanced editing data and integrates an MLLM Reasoner with a DiT Generator through multi-stage training and iterative reflection.
- 3. Method: ReasonEdit combines Thinking Pairs and Reflection Triples with an MLLM Reasoner and DiT Generator to support cascaded image-editing reasoning.Thinking Pairs map abstract requests to actionable directives, while Reflection Triples model intermediate and refined image states.
- 3. Method: Thinking Pairs translate ambiguous, colloquial, or informal requests into precise, standardized, executable editing directives.The dataset uses categorization, annotation, and review to create abstract-to-concrete instruction pairs.
- 3. Method: 150k high-quality Thinking Pairs are selected from 500k image-instruction pairs, including 62k simplified complex instructions and 88k abstracted simple instructions.The final collection also includes 50k simple, unedited instructions.
- 3. Method: The proposed reflection design addresses hallucinations in dual-image evaluation while retaining before-and-after information through an intermediate generated image and refined target.This structure supports subsequent adjustments when further edits are needed.
- 3. Method: Reflection evaluates generated outputs through a multi-round single-image process that produces success, refinement, or failure decisions and determines termination.The pipeline generates a target description, computes consistency, and can return revised editing instructions or a final score.
- 3. Method: Training proceeds from reasoning learning, to frozen-MLLM DiT adaptation, to unified fine-tuning of both components.The progressive strategy separates reasoning acquisition, generator adaptation, and joint optimization.
4. Experiments
Experiments evaluate ReasonEdit across foundational editing and abstract reasoning benchmarks, showing gains from thinking and reflection. The full approach improves fidelity on complex edits and benefits from multi-round reflection, though later rounds add little.
- Evaluation Setup: The experiments use GEdit-Bench, ImgEdit-Bench, and KRIS-Bench with task-specific automated or model-based metrics.The benchmarks span foundational editing, instruction adherence, perceptual quality, detail preservation, and abstract reasoning.
- Quantitative Results: +4.3% and +4.7% on ImgEdit and GEdit for ReasonEdit-S, while ReasonEdit-Q gains +2.8% and +3.4% over its base model.These benchmarks assess foundational image editing capabilities.
- Qualitative Results: Qualitative comparisons show more precise target edits and better preservation of backgrounds, facial features, and hairstyles.The results address consistency and fidelity failures in image editing.
- Quantitative Results: +8.2% on KRIS-Bench for ReasonEdit-S and +6.1% for ReasonEdit-Q, with the approach outperforming open-source models and several closed-source methods.KRIS-Bench evaluates abstract reasoning and difficult instructions, making it a central test of the proposed reasoning capabilities.
- Reasoning Mechanisms: Thinking and reflection simplify abstract instructions, detect subtle errors, and support self-correction during multi-step editing.The two mechanisms reduce misinterpretation and prevent mistakes from compounding across editing steps.
- Ablation Studies: Two reflection rounds raise the KRIS-Bench score from 58.64 to 60.93, while three or four rounds add only +0.06 and +0.14 at higher computational cost.The reflection-round study uses ReasonEdit-S and compares against a thinking-only baseline.
5. Conclusion
ReasonEdit introduces thinking and reflection as explicit reasoning capabilities for image editing. Experiments support its effectiveness across simple and complex edits, with strong open-source performance and competitiveness with closed-source models.
- Conclusion: ReasonEdit trains thinking and reflection on Thinking Pairs and Reflection Triples to support instruction understanding and self-correction.The framework is presented as a structured reasoning pipeline for image editing.
- Conclusion: The model achieves state-of-the-art performance among open-source methods while remaining highly competitive with several closed-source models.The conclusion presents structured instruction understanding and self-correction as central to high-fidelity editing.
- Conclusion: The framework is designed to handle both simple and complex editing tasks with high fidelity and consistency.This conclusion is stated as the supported scope of the approach.
Contributors and Acknowledgments
The paper lists its core contributors and identifies corresponding authors.
- Contributors: The paper credits Fukun Yin, Shiyu Liu, Yucheng Han, and additional collaborators as core contributors.The passage lists the complete contributor group.
- Contributors: Gang Yu is included among the listed core contributors.The contributor list contains multiple research and engineering authors.
- Acknowledgments: The passage separately labels corresponding authors after the contributor list.Their names are not included in the supplied passage.
A.1. Illustration of Reflection Pipelines
The appendix compares alternative reflection pipelines and motivates the proposed multi-round design. It reports hallucinations in a unified dual-image input scheme for current MLLMs.
- Reflection Pipelines: The ablation compares dual-image, pure single-image, and proposed multi-round reflection pipelines.The comparison is reported in Table 4 and illustrated in Fig. 6.
- Dual-Image Reflection: The dual-image pipeline jointly provides the reference image, instruction, and result image to the MLLM for concurrent thinking and reflection.This design directly compares the initial input with the generated output.
- Dual-Image Reflection: Current MLLMs frequently hallucinate during image editing under the unified dual-image input scheme.This finding motivates evaluating alternative reflection designs.
B. More Results
Additional qualitative results cover diverse editing instructions, multi-round reasoning and reflection, and comparisons with state-of-the-art methods.
- Qualitative examples span diverse editing instructions, multi-round reasoning-and-reflection processes, and comparisons against state-of-the-art methods.
C. Failure Cases
The failure-case analysis examines how reasoning and reflection diagnose and correct image-editing errors across abstract instructions, visual transformations, and multi-round edits. It also identifies failures caused by incomplete planning and limited generation capability.
- The qualitative figures compare reflection pipelines and illustrate thinking, reflection, and progressive multi-round correction across editing examples.The figure descriptions distinguish dual-image and sequential single-image reflection and present qualitative reasoning examples.
- Reflection identifies semantic mismatches, such as replacing a Greek mythological wisdom bird with an eagle, then proposes an owl or phoenix.The later reflection distinguishes a real animal from the requested mythical bird and recommends a corrective edit.
- Reflection corrects structural and stylistic errors by revising an Egyptian-looking stepped pyramid into a distinctly Maya structure with platforms and carvings.The second round reports successful transformation while preserving the desert setting without noticeable artifacts.
- Multi-round reflection improves environmental blending and object-count compliance by changing a chameleon’s colors and removing an extra durian.The examples show reflection detecting the wrong background change and the presence of two durians instead of one.
- Other examples show reflection verifying successful edits, including cigarette removal and adding water to dry ice with visible vapor and no unintended artifacts.These cases illustrate reflection’s role in judging whether an edit satisfies the instruction and can stop successfully.
- Some failures persist when planning is incomplete or generation cannot selectively retain content, including an inconsistent hand gesture and removal of all apples.The failure analysis attributes these cases respectively to incomplete planning and insufficient generation capability.