LL3M
lastmod 2025-08-19created 2025-08-191 edits
3D modeler based on large language models
About
- LL3M uses multiple large language models to automatically write Python code that creates and edits 3D assets in Blender
- Directly builds creative and precise shapes from the user's text instructions, and implements complex geometric operations in code
- Unlike existing 3D model generation tools, it offers unconstrained asset creation and fine-grained interaction
- The generated Blender code is clear and its parameters are highly transparent, so users or agents can easily modify it or refine it iteratively
- Demonstrates broad 3D asset processing capabilities, including consistent stylization, material editing, and hierarchy construction
LL3M overview
- LL3M is an innovative framework in which multiple large language model (LLM) agents write Python code to create and edit 3D assets in Blender
- When the user gives instructions in text, LL3M automates creative shape generation and precise geometric operations, and uses high-level code as the 3D representation, enabling iterative refinement and collaboration
- The code is clearly documented, exposing its many parameters and structure transparently, which makes further editing and continuous user feedback easy
Pipeline overview
- The pipeline consists of three main stages (initial creation, automatic refinement, and refinement based on user feedback)
- The initial creation stage builds the basic shape, while LL3M automatically detects and improves logically inappropriate structures or simplistic geometric elements
- The second stage applies more refined automatic corrections, also reflecting complex shapes and relationships
- The final stage accepts the user's additional edit requests, enabling interactive and iterative 3D asset creation
See also
- LLM
- 3D
- Deep Learning