CanvasAgent: Enabling Complex Image Creation and Editing via Visual Tool Orchestration

This paper introduces CanvasCraft, a multimodal tool-use dataset with 140K fully annotated executable trajectories and 10K reinforcement-learning task specifications, together with CanvasAgent, an agent for complex image creation and editing. The workflow covers synthesis, object localization, segmentation, selected-region editing, compositing, text reading, and enhancement. CanvasAgent is first trained with supervised fine-tuning on executable reasoning-action trajectories, then optimized with GRPO using hybrid outcome- and process-level rewards. During multi-turn rollouts, it inspects intermediate images, tracks visual assets, and adapts tool choices as the visual state changes. The abstract reports evaluations of both final image quality and trajectory behavior.