yejy53/genclaw

GenClaw is a research project exploring a different way to generate images with AI. Standard AI image generation works by describing what you want in words, and a model like Stable Diffusion or DALL-E produces the image in one shot. The result can be hard to control, especially for things like exact text placement, precise object counts, or specific spatial layouts. GenClaw's approach inserts a coding step between the description and the final image. An AI agent first writes executable code, such as SVG graphics, HTML with CSS, or lightweight 3D code, to sketch out the visual composition. This code is a verifiable, editable representation of the scene: you can inspect it, run it, fix mistakes, and iterate before committing to an expensive image generation call. Only after the code sketch is settled does the system call an image generation model to produce the final rendered result. The README describes this as mirroring how a human artist works: conceptualize, sketch, add color, refine. Each stage produces a tangible artifact you can review and modify rather than a single black-box output. The project frames this as plugging image generation into an AI agent's existing toolbox for planning, tool use, and reflection, making image creation a first-class capability of an agent rather than an isolated model. The repository accompanies a technical report published on arXiv. At the time of writing, the README notes that code and demos are being prepared for release and are not yet available. The current contents appear to be the paper description and showcase images rather than runnable code. If you are looking for something to install and use today, this repository is not ready for that yet.