HIVE-3D

This repository contains the two trained components used by HIVE-3D on top of microsoft/TRELLIS-image-large:

• denoiser: an IP-Adapter sparse-structure flow model.
• latent_projector: the latent conditioning projector used by the denoiser.

Both components are from training step 200000 and are distributed in the safetensors format. The HIVE-3D source code is required to construct the model classes and run the complete image-to-3D pipeline.

Files

denoiser/
├── config.json
└── model.safetensors
latent_projector/
├── config.json
└── model.safetensors

Loading

Download this repository through huggingface_hub, then pass the component files to the HIVE-3D pipeline:

from pathlib import Path

from huggingface_hub import snapshot_download
from trellis.pipelines.test_pipeline import TestImageToSlatPipeline

checkpoint_root = Path(snapshot_download("mocun123/HIVE-3D"))

pipeline = TestImageToSlatPipeline.from_pretrained(
    "microsoft/TRELLIS-image-large",
    checkpoint_root / "denoiser/model.safetensors",
    checkpoint_root / "latent_projector/model.safetensors",
    checkpoint_root / "denoiser/config.json",
    checkpoint_root / "latent_projector/config.json",
)
pipeline.cuda()

HIVE-3D follows the TRELLIS runtime requirements. A CUDA-capable NVIDIA GPU is required for the full pipeline; at least 16 GB of GPU memory is recommended.

Checksums

4c679dd55812709b2fa39547f714d63aeabe62f5b63fcf0191fee182ce4aa77a  denoiser/model.safetensors
a8ce9615599020a7bf5c26538c7764431440da51439d32aeb5dc0815ca97369f  latent_projector/model.safetensors

Base Model

HIVE-3D builds on the MIT-licensed microsoft/TRELLIS-image-large model. This repository does not redistribute the TRELLIS base-model weights; they are downloaded from the official repository at runtime.

License Notice

A license for the HIVE-3D weights has not yet been declared in this model card. Users should review the HIVE-3D project license, the TRELLIS license, and the licenses of any optional third-party components before redistribution or commercial use.