Research map / Vision-language and multimodal / Multimodal LLMs
ReasonEdit: Towards Reasoning-Enhanced Image Editing Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ReasonEdit, an image editing model that adds MLLM-based thinking (abstract-to-concrete instruction rewriting) and multi-round single-image reflection to a Step1X-Edit or Qwen-Image-Edit backbone, trained in three stages.
Problem
Instruction-based image editing with frozen MLLM encoders struggles with abstract or complex instructions and lacks self-correction, limiting accuracy and test-time scaling.
Approach
Unfreeze the MLLM encoder and train it with thinking pairs (abstract-to-concrete instructions) and reflection triples (multi-round single-image self-assessment), so the model runs a thinking-editing-reflection loop that stops at a chosen round; reflection is decomposed into target description, result assessment, and refinement conclusion to avoid dual-image hallucination.
Key results
- ReasonEdit-S (thinking+reflection) reaches 60.93 on KRIS-Bench overall vs 56.33 base (+8.2%), GEdit overall 7.58 (+4.7%), ImgEdit 4.40 (+4.3%).
- ReasonEdit-Q reaches 61.57 KRIS-Bench overall (+6.1%) and GEdit overall 7.77, the best among open-source models on GEdit and KRIS.
- Multi-round single-image reflection (58.29) outperforms dual-image (53.79) and single-image (53.04) reflection pipelines on KRIS-Bench ablation with the base generator.
Surprising findings
- Naive re-roll sampling peaks at 59.24, only marginally above thinking-only, while structured reflection gives larger gains.
- Frozen base Qwen reasoning adds only a marginal 0.82-point gain over the pre-trained generator, while fine-tuned Qwen reasoning yields much larger gains.
Limitations
- Gains on GEdit and ImgEdit are smaller since these simpler benchmarks may not exercise thinking and reflection fully.
- Dual-image reflection and naive pipelines suffer from MLLM hallucination; the multi-round pipeline still depends on VLM-generated annotations and manual screening.
Experimental setup
Datasets: GEdit-Bench ImgEdit-Bench KRIS-Bench
Benchmarks: GEdit-Bench ImgEdit-Bench KRIS-Bench
Metrics: Semantic Consistency Perceptual Quality Overall score KRIS overall ImgEdit overall
Baselines: Step1X-Edit v1.1 Qwen-Image-Edit-2509 BAGEL BAGEL-thinking OmniGen2 UniWorld-Qwen-Image-Edit Flux-Kontext-dev GPT4o Gemini 2.5 flash Seedream 4.0
Builds on: Step1X-Edit Qwen-Image-Edit BAGEL OmniGen2 VIEScore LoRA
Closest papers by meaning
- Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning (ICML 2026 · similarity 0.94)
- SmartEdit: Exploring Complex Instruction-based Image Editing with Multimodal Large Language Models (CVPR 2024 · similarity 0.92)
- KRIS-Bench: Benchmarking Next-Level Intelligent Image Editing Models (NeurIPS 2025 · similarity 0.91)
- Envisioning Beyond the Pixels: Benchmarking Reasoning-Informed Visual Editing (NeurIPS 2025 · similarity 0.90)
- InsightEdit: Towards Better Instruction Following for Image Editing (CVPR 2025 · similarity 0.90)
- Guiding Instruction-based Image Editing via Multimodal Large Language Models (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.