arxiv:2607.05465

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

Published on Jul 6

· Submitted by

wh on Jul 8

Authors:

Abstract

A large-scale multimodal tool-use dataset and agent are presented for complex image creation workflows that orchestrate multiple visual tools through multi-turn interactions and hybrid reward optimization.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

Complex image creation and editing often require more than a single generation or editing model. A user request may involve synthesizing images, localizing objects, segmenting regions, editing selected content, compositing intermediate assets, reading text, and enhancing the final result. Such tasks shift multimodal agents from perception-augmented reasoning to manipulation-centered visual creation, where tools must actively transform visual states rather than merely inspect them. However, existing multimodal tool-use agents are mostly optimized for perception, search, or domain-specific editing, and lack large-scale supervision for executable image-creation trajectories. In this paper, we introduce CanvasCraft, a large-scale multimodal tool-use dataset for complex image creation and editing, and CanvasAgent, a tool-augmented multimodal agent that learns to orchestrate heterogeneous visual tools through multi-turn interaction. CanvasCraft contains 140K fully annotated executable trajectories and 10K RL task specifications. CanvasAgent is first trained with SFT to learn executable reasoning-action trajectories, and is then optimized with GRPO using a hybrid reward that combines outcome- and process-level signals. During rollout, CanvasAgent inspects intermediate results, tracks visual assets, and adapts tool decisions to the evolving visual state. Experiments evaluate both final image quality and trajectory behavior, demonstrating the effectiveness of CanvasAgent and the proposed dataset for complex multi-tool image creation workflows.

View arXiv page View PDF Add to collection

Community

cswhjiang

Paper submitter 2 days ago

This paper presents CanvasAgent, a tool-augmented multimodal agent for complex image creation and editing. Unlike single-step image generation/editing methods, CanvasAgent decomposes user requests into executable multi-tool trajectories involving generation, grounding, segmentation, OCR, compositing, transformation, and super-resolution. A key contribution is CanvasCraft, a large-scale dataset with 140K supervised executable trajectories and 10K RL task specifications, enabling both SFT-based tool-use bootstrapping and GRPO-based policy optimization. The hybrid reward design is also interesting, as it jointly evaluates final image quality and process-level tool-use validity. Experimental results show strong gains over general MLLMs and image-only baselines, especially on alignment and trajectory quality. Overall, this is a timely and useful step toward more controllable, interpretable, and stateful visual creation agents.

O96a

2 days ago

CanvasAgent hits on something I've been running into building agentic systems: the gap between "perception-augmented reasoning" and "manipulation-centered creation" is real. Most multimodal agents today are great at looking at an image and telling you what's in it, but ask them to composite, mask, inpaint, and re-render in a single pipeline and they fall apart because each tool boundary is a context switch with no shared state.

The orchestration approach here — routing through a visual state that tools actively transform rather than just inspect — is the right abstraction. I've seen similar patterns in agentic coding tools where a shared file tree replaces ad-hoc tool outputs. The same principle applies to images: if your segmentation model returns a mask and your inpainting model can't read that mask's format, you've already lost.

What I'd want to see benchmarked is the failure rate on multi-step edits. Not just whether the final image looks good, but how often a tool call in the middle of a 5-step pipeline silently corrupts state. That's the kind of reliability metric that determines whether this ships as a demo or as a product. The paper's framing is solid — the hard part is making the orchestration robust enough that the agent doesn't need a human babysitter for every composite operation.