DamianBoborzi/TRELLIS-text-xlarge_MeshFleet_ft

TRELLIS Text-to-3D XLarge Finetuned Model

This is a finetuned version of the TRELLIS Text-to-3D XLarge generation model. It was finetuned using the MeshFleet dataset.

Model Description

This finetuned model contains all the necessary components for the TRELLIS Text-to-3D pipeline, making it a fully standalone model that doesn't require external dependencies.

Finetuned Components

• Sparse Structure Text Flow Model (ss_flow_txt_dit_XL_16l8_fp16)
• SLaT Text Flow Model (slat_flow_txt_dit_XL_64l8p2_fp16)

Usage

from trellis.pipelines import TrellisTextTo3DPipeline

# Load the pipeline
pipeline = TrellisTextTo3DPipeline.from_pretrained("path/to/this/model")
pipeline.cuda()

# Generate 3D assets from text
text_prompt = "A vintage wooden chair with ornate carvings"
outputs = pipeline.run(text_prompt, seed=42)

# Access different 3D representations
gaussian_splatting = outputs['gaussian'][0]  # Gaussian splatting representation
radiance_field = outputs['radiance_field'][0]  # Radiance field representation
mesh = outputs['mesh'][0]  # Mesh representation

# Save outputs
gaussian_splatting.save("output_gs.ply")
radiance_field.save("output_rf.obj")
mesh.save("output_mesh.obj")

File Format

All model weights are stored in the same SafeTensors format (.safetensors) as the original weights. Configuration files (.json) contain model hyperparameters and architecture details.

Requirements

• CUDA-capable GPU with >= 24GB VRAM recommended
• TRELLIS library (see the official TRELLIS repository for installation instructions: https://github.com/microsoft/TRELLIS)

License

This model follows the same license as the original TRELLIS model.