ReasonUni improves image editing understanding and generation in AI models
ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Matching images to text and creating images from text are two connected tasks for AI models that use pictures and words together. Existing methods mainly focus on recognizing details but don’t help much with understanding what changes should be made or why. The researchers developed ReasonUni, which helps models think through editing instructions step-by-step and then produce better edited images based on that reasoning. They also invented a new learning method that teaches the AI to improve both its understanding and image creation at the same time. Tests show this approach consistently improves results on multiple challenges.
What this means in practice
- •For computer vision engineers: Build AI systems that better understand editing commands and generate accurate edited images for image manipulation tools.
- •For digital content creators: Use improved AI models to create and refine images from text descriptions with greater precision in visual editing tasks.
Authors
Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan, Li Yuan
Abstract
Image-to-text (I2T) understanding and text-to-image (T2I) generation are dual tasks in unified multimodal models (UMMs). Reconstructive post-training couples these tasks, but existing reconstruction-based methods remain perception-centric and provide limited supervision for reasoning about structure, intent, and transformation. We propose ReasonUni, a reasoning-augmented reconstructive post-training framework that extends this coupling beyond perception. Given an original image and an editing instruction, the UMM first reasons about what and how to edit and describes the target state, then generates the corresponding target image conditioned on this understanding output. To jointly optimize both modules, we introduce TreeGRPO, a tree-structured reinforcement learning method that samples multiple reasoning paths and candidate images per path, then aggregates visual rewards to update both modules under a unified objective. Experiments across multiple benchmarks demonstrate consistent gains in both understanding and generation, showing that reasoning-augmented reconstruction yields genuine mutual gains for UMMs.