GenRevisit improves text to image generation by revisiting earlier results
Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Creating detailed images from text instructions can be hard, especially when making complex scenes step-by-step. Existing methods usually keep improving only the latest image even if it got worse. This paper introduces GenRevisit, a system that can choose not just how to edit but also which earlier image to go back to when refining. This helps it make better pictures overall. The authors also developed a way to better teach the system which editing steps matter most.
What this means in practice
- •For ai developers: Build more reliable image creation tools by allowing systems to revisit previous image versions during stepwise refinement.
- •For game designers: Generate complex visual game assets by iteratively improving images with selective backtracking to earlier design states.
Authors
Zixu Wang, Xiaoran Zhao, Yahui Liu, Feng-ao Wang, Junhao Xiao, You Shu, Bing Liu, Yiting Liu, Songlei Jian, Zhiliang Tian, Dongsheng Li
Abstract
Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and editing, adapting their actions based on interaction history. However, they typically fix the continuation state (i.e., the image from which refinement proceeds) to the latest image. Since iterative editing is not necessarily monotonic, the agent may continue refining from the latest image even when it has already degraded. In this work, we introduce GenRevisit, an agentic framework that extends action selection to joint continuation-state and action selection. The agent can not only use interaction history to adjust editing instructions but also select an image candidate to edit, jointly deciding where to continue from and how to refine. We train GenRevisit with SFT followed by agentic reinforcement learning. To address the coarse credit assignment of final outcome reward in GRPO, which assigns the same trajectory-level advantage to all actions, we further propose Normalized Action Credit Assignment (NACA) to redistribute positive advantage according to each action's contribution to reducing the remaining objective. On GenEval2, GenRevisit achieves GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, surpassing the strongest baselines using the same generators by 12.7 and 10.9 points respectively. GenRevisit achieves state-of-the-art performance on GenEval, GenEval++, and GenEval2 among baselines.