<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> LL3M generates 3D assets in Blender via interpretable Python code, a novel approach distinct from data-driven methods. This multi-agent LLM system offers artists a modular, editable, co-creative workflow, enhancing integration and providing human-readable scene representations.

How It Works

LL3M uses specialized LLM agents to generate 3D assets through code. It reformulates shape generation as a code-writing task, coordinating agents to plan, retrieve (using BlenderRAG from API docs), write Python scripts leveraging Blender's constructs, and debug them. This code-centric method provides modularity, editability, and a co-creative loop for iterative refinement via code or parameters.

Quick Start & Requirements

• Installation: Clone repo, set up Conda env (python=3.12), install requirements (pip install -r requirements.txt).
• Blender: Requires v4.4.x; install ./blender/addon.py via preferences.
• Authentication: Login (python main.py --login) and accept terms (python main.py --accept-terms); 5 requests/day limit.
• Hardware: Min OS (Win 10/11, macOS 10.15+, Linux Ubuntu 18.04+), CPU (i5-8400/Ryzen 5 2600), 8GB RAM, 2GB VRAM GPU. Rec: Win 11, macOS 12+, Linux Ubuntu 20.04+, i7-10700K/Ryzen 7 3700X, 16GB RAM, RTX 3060/RX 6600 XT (8GB+ VRAM).
• Links: Blender 4.4 download.

Highlighted Details

• Code-Centric Generation: Produces human-readable, editable Python scripts for Blender, ensuring high modularity and interpretability.
• Multi-Agent System: Employs specialized LLM agents for planning, retrieval, code generation, and debugging, improving asset complexity and correctness.
• BlenderRAG: Integrates API documentation knowledge for advanced modeling.
• Co-Creative Workflow: Supports iterative refinement via user instructions and agent self-critique.
• Session Refinement: Enables resuming and modifying previous generation sessions.

Maintenance & Community

No explicit community links (Discord, Slack, forums) or roadmap are provided. The project appears academic, with authors listed and a BibTeX citation.

Licensing & Compatibility

The README does not specify a software license, requiring further investigation for adoption, especially for commercial use.

Limitations & Caveats

• LLM Hallucination: May produce inconsistent or flawed results; detailed prompts or re-runs are advised.
• Daily Request Limit: User accounts are limited to 5 requests per day.
• No Revert: Changes during user-guided refinement are not reversible in the demo.
• Performance: Complex rendering can freeze Blender/system; adjust rendering parameters (num_images, resolution_scale) for lower-end hardware.
• Reproducibility: Random seeds affect output; exact paper results may not be reproducible.