GenClaw explores code-driven agentic image generation: instead of only rewriting prompts, an image generation agent uses code as a controllable visual canvas before calling image generation models for final rendering.
The core idea is simple: think, sketch with code, then render.
🎨 Code as a Visual Brush. The agent creates by writing executable visual sketches—SVG, HTML/CSS, Python, lightweight 3D code—turning object count, spatial layout, and text rendering into executable, verifiable, debuggable programs. Image synthesis shifts from implicit diffusion sampling to an explicit, reasoning-friendly process.
✋ Draw as a Human Artist. We mirror the human creative loop—conceptualize → sketch → coloring → refine—and make every stage transparent: ideation, reference retrieval, drafting, and incremental rendering are all surfaced as inspectable, editable, revertible artifacts. Generation becomes an iterative collaboration rather than one-shot black-box inference.
🔌 Agent Harness for Image Generation. We plug an LLM agent's proven planning, tool-use, and reflection abilities directly into image synthesis, exploring an agent harness for image generation—so that creating images becomes a first-class capability inside the agent's toolbox, not an isolated standalone model.
The technical report is available now. Code and demos are being prepared and will be released later.
- Paper: https://huggingface.co/papers/2605.30248
- arXiv: https://arxiv.org/abs/2605.30248
- Repository: https://github.com/yejy53/GenClaw
If you find this project interesting, please consider giving it a star and voting for the paper on Hugging Face.
If you find GenClaw useful, please consider citing our technical report:
@article{ye2026genclaw,
title={GenClaw: Code-Driven Agentic Image Generation},
author={Ye, Junyan and others},
journal={arXiv preprint arXiv:2605.30248},
year={2026}
}