TRELLIS 3D: Generate 3D Models from a Single Image or Text Description 85

What is TRELLIS 3D?

TRELLIS 3D is an open-source 3D generation model developed jointly by Microsoft Research Asia, Tsinghua University, and the University of Science and Technology of China. Using SLAT technology, it can directly generate high-quality 3D models from a single 2D image or text description.

The TRELLIS 3D model has 4 billion parameters. Based on the new O-Voxel representation, it achieves 16x spatial compression and can produce textured 512³ resolution models in just 3 seconds.

Main Features of TRELLIS 3D

• Lightning-fast generation: 512³ takes only 3 seconds on an H100, 1024³ around 17 seconds, and 1536³ about 1 minute

• High resolution support: From 512³ to 1536³, with impressive compression ratios

• Handles any shape: Unlike traditional methods that only produce closed meshes, it can handle open surfaces like clothing and foliage, as well as objects with internal structures

• Complete material system: Full PBR pipeline support, including base color, roughness, metallic, and transparency

• Wide format support: Exports to GLB, OBJ, STL, GLTF, USDZ, PLY, 3D Gaussian, and other common formats

• Two input modes: Both image-to-3D and text-to-3D are supported

• Local editing: After generation, you can add, remove, or modify specific regions

• Simple processing pipeline: Converting textured mesh to O-Voxel takes under 10 seconds on CPU; the reverse conversion on CUDA takes just 100 milliseconds

Use Cases for TRELLIS 3D

• Game development: Quickly produce scene props and characters, no need for artists to model from scratch

• Film & animation: Convenient for creating VFX assets and preview models

• VR/AR: Generate realistic virtual objects to enhance immersion

• Industrial design: Quickly produce product prototypes to validate designs

• E-commerce: Create interactive 3D product displays

• 3D printing: Directly generate printable model files

• Education & research: Used as a tool for generating 3D data

Pros & Cons of TRELLIS 3D

What it does well

• Truly high quality: 4 billion parameters combined with a sparse 3D VAE deliver impressive detail and fidelity

• Unmatched speed: 3 seconds to generate a model — far faster than traditional modeling

• Few topology restrictions: Handles complex geometric structures, including open surfaces and internal structures

• Excellent material support: Complete PBR workflow, with transparency effects supported

• Fully open source: MIT license, free for commercial use

• Modest hardware requirements: Runs with 16GB VRAM; RTX 3090/4090 are both supported

Drawbacks

• Hardware-hungry: For top-tier speed you still need an H100-class GPU; consumer cards will be noticeably slower

• Limitations on fine details: Facial features and tiny textures (like ropes) aren't perfectly generated yet

• High training cost: Large-scale pretraining requires massive 3D datasets — training from scratch is very expensive

• Still essentially low-poly: Edge details look a bit rough when zoomed in — fundamentally a low-polygon model

TRELLIS 3D Tutorial

1. Go directly to the TRELLIS 3D official website

2. Upload a JPG/PNG image (1920×1080 or higher recommended)

3. Click generate, wait a few seconds, and the 3D model appears

4. Once satisfied with the preview, download as GLB or another format

FAQ

• TRELLIS 3D generation is too slow?

  Make sure CUDA acceleration is enabled; an RTX 4090 or above is recommended. Lowering the resolution speeds things up.

• Running out of VRAM?

  Setting the environment variable PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True helps save VRAM.

• Transparency effect not showing?

  GLB is opaque by default. You need to manually connect the texture Alpha to transparency in your 3D software.

• Details not sharp enough?

  Try 1024³ or 1536³ resolution; extremely complex structures still hit a technical ceiling.

• Which image formats are supported?

  JPG, PNG, and WEBP all work; images with alpha channels are treated as masks.

• Can I use it commercially?

  Yes, it's open source under MIT, but note that dependencies (such as nvdiffrast) may have their own separate licenses.

• Want to fine-tune it yourself?

  TRELLIS.2 provides complete training code. Use train.py to fine-tune or train from scratch.

Related Articles:

AI-Generated 3D Models