RUST · GENERATIVE 3D AI

hunyuan3d-rs — Text-to-3D & Image-to-3D Mesh and PBR Texture Engine in Rust

hunyuan3d-rs is a native Rust implementation of Tencent's Hunyuan3D-2 pipeline for synthesizing high-resolution 3D meshes and physically based rendering (PBR) textures from text or single images. It combines Dual-Stream Diffusion Transformer shape generation, octree Marching Cubes surface extraction, and UV-unwrapped albedo/roughness/metallic baking.

$ cargo install hunyuan3d-rs
GitHub ↗

Architected for Speed & Reliability

Engineered from the ground up in memory-safe Rust with zero unnecessary dependencies.

📐

Marching Cubes Surface Extractor

Pure Rust implementation of the 256-configuration edge and triangle lookup table with iso-surface interpolation and bounding box normalization.

🌊

Rectified Flow Matching ODE

High-order numerical solvers including Euler, Midpoint, Heun, and 4th-order Runge-Kutta (RK4) with shifted schedule matching upstream Hunyuan3D-2.1.

🧠

Spatio-Temporal DiT & 3D RoPE

3D Rotary Position Embeddings across spatial axes, scaled dot-product attention, AdaLN-Zero modulation, and Mixture of Experts (MoE) routing.

🧹

Automated Mesh Healing & QEM

Prunes detached floating debris, applies iterative Laplacian smoothing, and executes Quadric Error Metric (QEM) decimation to target polygon budgets.

🎨

PBR Material Synthesis

Bakes physically-based rendering textures including Albedo, Roughness, and Metallic maps directly into standard GLB material channels.

🖥️

Parallel Software Rasterizer

Multi-threaded CPU software rasterizer powered by Rayon for multi-view validation and texture projection without GPU driver lock-in.

Get Started in Seconds

Install via Cargo and run immediately on macOS, Linux, or Windows.

Terminal
# Build and install Hunyuan3D-RS
cargo install --git https://github.com/bhubbard/hunyuan3d-rs

# Generate a high-resolution 3D asset from prompt
hunyuan3d generate --prompt "ancient ornate brass lantern" --solver rk4 -o lantern.glb

# Decimate and repair an existing mesh using QEM
hunyuan3d decimate --input raw.obj --target-faces 15000 -o optimized.glb

# Run the test suite
cargo test --release

Engine Architecture

Streamlined execution pipeline with deterministic memory isolation and zero runtime overhead.

Pipeline Flow
┌───────────────────────────────────────────────────────────┐
│               Input: Text Prompt / Image Latent           │
└─────────────────────────────┬─────────────────────────────┘
                              │
┌─────────────────────────────▼─────────────────────────────┐
│          Rectified Flow Matching ODE (Euler/Heun/RK4)     │
└─────────────────────────────┬─────────────────────────────┘
                              │ 3D Density Grid
┌─────────────────────────────▼─────────────────────────────┐
│          MCSurfaceExtractor (Marching Cubes 256)          │
└─────────────────────────────┬─────────────────────────────┘
                              │ Raw Mesh & Triangles
┌─────────────────────────────▼─────────────────────────────┐
│     Mesh Post-Processing (Laplacian + QEM Decimation)     │
└─────────────────────────────┬─────────────────────────────┘
                              │
┌─────────────────────────────▼─────────────────────────────┐
│      PBR Material Synthesis & Texture Baking -> GLB       │
└───────────────────────────────────────────────────────────┘