COMFYCLAW Agent Evolves Skills for Image Generation Workflows

This research introduces COMFYCLAW, an innovative agentic system designed to enhance image generation workflows, particularly within environments like ComfyUI. The system treats workflow creation as a process of typed graph editing, providing tools for each stage and automatically correcting invalid modifications. A key feature is its integration of a region-level vision-language model (VLM) that analyzes visual failures and translates them into actionable suggestions for repair. COMFYCLAW's core innovation lies in its ability to evolve a skill library. It distills insights from previous workflow executions, including successful trajectories, encountered errors, and verifier feedback, into reusable agent skills. This continuous learning mechanism significantly improves the agent's reliability and performance. Evaluations across various benchmarks, agent models, and image backbones demonstrate that COMFYCLAW consistently achieves superior image generation scores compared to baselines without skill evolution. Human annotations further confirm a preference for the system's output, highlighting the effectiveness of its skill evolution approach in complex visual workflow construction.