Microsoft TRELLIS.2: 4B Image-to-3D Generative Model Using O-Voxels

X postSupermanTool

Microsoft has released the repository for TRELLIS.2, a 4-billion parameter generative model designed for high-fidelity image-to-3D synthesis. It introduces a field-free sparse voxel structure called O-Voxel to generate complex 3D meshes with full PBR materials.

Key points

  1. 4B-parameter model utilizing vanilla DiTs and a Sparse 3D VAE with 16× spatial downsampling.
  2. Implements O-Voxel representation to support non-manifold geometry, open surfaces (e.g. foliage, clothing), and internal cavities.
  3. Models complete PBR attributes including base colour, metallic, roughness, and opacity.
  4. Inference on an NVIDIA H100 takes approximately 3 seconds at 512³, 17 seconds at 1024³, and 60 seconds at 1536³ resolution.
  5. System prerequisites require Linux, Python 3.8+, CUDA 12.4, and an NVIDIA GPU with at least 24 GB of VRAM.
  6. Independent user feedback highlights topology issues, such as thousands of fragmented connected components within a single primitive.

Why it matters

Single-image 3D generation has frequently been constrained by the limitations of continuous implicit fields. By adopting a discrete, field-free voxel approach, TRELLIS.2 improves speed and material accuracy, bridging an important gap for rapid 3D asset prototyping.

Analysis

While conversion between O-Voxels and textured meshes is fast and rendering-free, early hands-on tests reveal practical downstream limitations. Generated GLB files may present overly dense and fragmented geometry requiring significant post-processing or decomposition before being suitable for real-time production engines.

Original post

Superman

@thesupermanmx · 29 July 2026

Repo: t.co/eW7JpjXoED

156 likes2 replies292 bookmarks12.3k views

View on X
Selected replies (1)
  • @FabioAngela79 @thesupermanmx I tested the “production-ready” part on a real TRELLIS.2 GLB. Visually good, but one primitive had 287k faces and 3,684 connected components. A normal split turned it into confetti. So I benchmarked downstream decomposition on my 3090 Ti: t.co/9oA9NwXTlS

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Repository:microsoft/TRELLIS.2

    TRELLIS.2: Microsoft Releases 3D Generative Model with O-Voxel Structure

    TRELLIS.2 is an open-source 4B-parameter 3D generative model by Microsoft designed for high-fidelity image-to-3D generation. Using a field-free sparse voxel structure named O-Voxel and Diffusion Transformers, it handles arbitrary topologies, sharp geometry, and complete PBR materials.

    11.1kstarsPython

    AI & LLMs· Tool

  • Repository:lightningpixel/modly

    Modly: Local Desktop App for AI-Powered 3D Mesh Generation

    Modly is an open-source desktop application for Windows, Linux, and Apple Silicon macOS that creates 3D models from images or text prompts directly on local GPUs.

    7.4kstarsTypeScript

    AI & LLMs· Tool

  • Repository:img2threejs/img2threejs

    img2threejs: Procedural Code-Only Image-to-3D for Three.js

    img2threejs turns reference images into procedural, animation-ready Three.js models generated purely as code, avoiding traditional mesh extraction or photogrammetry.

    15.1kstarsPython

    AI & LLMs· Tool

  • X post:Mia

    Uncensored GGUF Quantisations Released for Qwen-Image 2.1

    A Hugging Face repository provides GGUF and safetensors quantisations of the Qwen-Image 2.1 base model for local execution in ComfyUI, featuring no integrated safety checkers or content filters.

    2.5kbookmarks138kviews

    AI & LLMs· Resource

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.