GenClaw proposes a code-driven agentic pipeline that replaces black-box prompt rewriting with three stages: conceptual search and reasoning, executable code sketching (SVG/HTML/ThreeJS), and final texture/photorealism supplementation by an image model. The paper argues this decoupling gives agents direct control over composition and layout, overcoming language ambiguity. It reports improved compositional control, text rendering, and support for physical simulation and editing, though specific numbers are not given in this excerpt.
Code-driven agents generate images via sketching and texturing stages
GenClaw uses LLM-generated SVG, HTML, or Three.js as an intermediate canvas, then refines with image synthesis for photorealism.
Independent
Junyan Ye · Jun He · Zilong Huang · Dongzhi Jiang · Xuan Yang · Rui Chen · +1 more
Research Digest··2 min read
In one line
GenClaw stages image generation into conceptualize, sketch (via code), and color, enabling direct canvas manipulation and interpretability.
Before you read this
Two ideas this paper builds oncode as actionA paradigm where actions are expressed as executable code, such as method calls or program synthesis, rather than direct manipulation.26 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.216 papers →
Read it as
js), and final texturing via an image model.
The authors developed GenClaw, a multi-stage agentic pipeline for image generation.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.