yejy53/GenClaw

What it solves

GenClaw addresses the limitations of one-shot image generation, where users often struggle with precise spatial layouts, specific object counts, and accurate text rendering. Instead of relying solely on prompt engineering, it introduces a way to use code as a controllable visual canvas to ensure the final image matches the user's intent.

How it works

The system operates as an agentic loop that follows a "think, sketch with code, then render" workflow. A planner LLM uses a set of tools to create a visual draft before calling an image generation model for the final result.

1. Perception: A sub-LLM writes "painter notes" to guide the planning.
2. Planning: The planner creates a todo_write plan and selects a tool pipeline.
3. Drafting: The agent uses tools like code_scene_draft (SVG) or code_text_draft (HTML/CSS) to create an explicit, editable layout.
4. Rendering: The draft is passed to image-to-image (i2i) or text-to-image (t2i) models for final synthesis.

Who it’s for

It is designed for creators and developers who need high-precision control over image composition, such as those designing posters with specific text or complex scenes with exact object placements.

Highlights

• Code as a Visual Brush: Uses SVG, HTML/CSS, and Python to turn spatial layout and text rendering into verifiable and debuggable programs.
• Human-like Creative Loop: Mirrors the process of conceptualizing, sketching, coloring, and refining, making every stage transparent and editable.
• Agentic Harness: Integrates LLM planning, tool-use, and reflection directly into the image synthesis process.
• Knowledge-Grounded: Includes search tools (via Tavily) and multimodal reasoning to ground image generation in real-world knowledge.