LLaMA-Mesh

LLaMA-Mesh is a groundbreaking text-to-3D generation model. Based on LLaMA architecture, it generates 3D mesh models (OBJ format) directly from text descriptions - no separate renderer needed. At just 8B parameters with GGUF quantization, it runs on consumer GPUs.

How it works

LLaMA-Mesh fine-tuned LLaMA to output 3D mesh vertices and faces as text tokens, enabling the model to generate complete 3D meshes directly in a single forward pass.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
 
tokenizer = AutoTokenizer.from_pretrained("Zengyi/LLaMA-Mesh")
model = AutoModelForCausalLM.from_pretrained("Zengyi/LLaMA-Mesh")
 
prompt = "A 3D model of a racing car"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=2000)
 
mesh_obj = tokenizer.decode(outputs[0])
# Save as .obj file
with open("output.obj", "w") as f:
    f.write(mesh_obj)

When to use

• Game development - quick 3D asset prototyping
• 3D printing - generate printable models
• Architecture - concept modeling
• E-commerce - product visualization
• Education - 3D geometry exploration

Limitations

• Mesh quality improves with prompt engineering
• Complex scenes may need refinement
• Output resolution limited by context window