render_package — 3D Render + Latent Encoding Pipeline

Self-contained portable archive for rendering 3D assets with Blender and encoding them into latent representations (DINOv2, UniLat, TRELLIS SLAT, TRELLIS SS).

Download & Deploy

# Download
huggingface-cli download Dennis0626/render_package render_package_portable.tar.gz --local-dir .

# Extract & setup
tar xzf render_package_portable.tar.gz
cd render_package
bash setup.sh                   # extract bundled conda env
source envs/env/bin/activate    # activate
vi config/default.yaml          # set your data paths

Requirements

• Linux x86_64
• NVIDIA GPU with CUDA 12.x drivers (>= 525)
• No conda/pip/internet needed after extraction

What's inside (~6.9 GiB compressed)

Component	Size
Python pipeline code	~50K
Blender 3.5.1 binary	~1.2G
Model weights (UniLat + TRELLIS + DINOv2)	~2.1G
Third-party model code	~16M
Pre-built conda environment (PyTorch, flash-attn, spconv, open3d, X11/GL)	~4.5G

Usage

# Multi-GPU encode
SPCONV_ALGO=native python encode_all.py --render_root /path/to/renders --num_gpus 4

# Render .tar.zst shards
python render_github.py --num_shards 10

# Full interleaved pipeline
SPCONV_ALGO=native python run_pipeline.py --render_gpus 0,1 --encode_gpus 2,3