I fine-tuned LLaMA 3.1 to generate 3D furniture models—essentially teaching an LLM to output mesh structures instead of just text.
The main technical challenge was handling furniture geometry complexity. I extended the context window to 20k tokens to capture the detailed mesh representations needed for furniture (sofas, chairs, tables, cabinets).
Dataset was curated from open-source 3D model repositories, filtered for furniture categories. Training was done on verda.com's GPU infrastructure.
Demo site: llm3d.space (currently in testing mode due to GPU costs)
The approach shows LLMs can bridge natural language understanding with 3D content generation. Potential applications in e-commerce visualization, interior design tools, and AR/VR.
Happy to answer technical questions about the fine-tuning process or mesh format adaptation!
I fine-tuned LLaMA 3.1 to generate 3D furniture models—essentially teaching an LLM to output mesh structures instead of just text. The main technical challenge was handling furniture geometry complexity. I extended the context window to 20k tokens to capture the detailed mesh representations needed for furniture (sofas, chairs, tables, cabinets). Dataset was curated from open-source 3D model repositories, filtered for furniture categories. Training was done on verda.com's GPU infrastructure. Demo site: llm3d.space (currently in testing mode due to GPU costs) The approach shows LLMs can bridge natural language understanding with 3D content generation. Potential applications in e-commerce visualization, interior design tools, and AR/VR. Happy to answer technical questions about the fine-tuning process or mesh format adaptation!