# TRELLIS.2-4B by Microsoft: Open-Weight Model
Source: https://savrn.com/models/trellis-2-4b
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## SAVRN's Notes on TRELLIS.2-4B

TRELLIS.2-4B generates 3D assets from a single image: the card describes a 4-billion-parameter flow-matching transformer over a field-free O-Voxel sparse-voxel structure, outputting geometry with full PBR materials, including transparency. Published weights are 9 safetensors files totalling 16.2 GB. The stats row states no parameter count, no context and no GPU table — plan around the 16.2 GB plus your runtime's own headroom, and assume a modern 24 GB-class card for comfortable inference.

MIT permits commercial use — a notably clean grant for a 3D generator, with notices kept and changes stated. The repo dates to 2025-12-01, last updated 2025-12-27, with the paper at arXiv:2512.14692 and code linked from the card. At 1.7M downloads a month it is among the most-pulled image-to-3D models we list. Test the export path into your DCC toolchain before building a pipeline around it; generation quality means little if meshes do not round-trip.

SAVRN Research

## Model Card

By Microsoft, published under mit, revision af44b45f2e35.

### TRELLIS.2: Native and Compact Structured Latents for 3D Generation

Model Name: TRELLIS.2-4B

Paper: [https://arxiv.org/abs/2512.14692](https://arxiv.org/abs/2512.14692)

Repository: [https://github.com/microsoft/TRELLIS.2](https://github.com/microsoft/TRELLIS.2)

Project Page: [https://microsoft.github.io/trellis.2](https://microsoft.github.io/TRELLIS.2)

### Introduction

TRELLIS.2 is a state-of-the-art large 3D generative model designed for high-fidelity image-to-3D generation. It leverages a novel "field-free" sparse voxel structure termed O-Voxel and a large-scale flow-matching transformer (4 Billion parameters).

Unlike previous methods that rely on iso-surface fields (e.g., SDF, Flexicubes) which struggle with open surfaces or non-manifold geometry, TRELLIS can reconstruct and generate arbitrary 3D assets with complex topologies, sharp features, and full Physical-Based Rendering (PBR) materials—including transparency/translucency.

### Model Details

- Developed by: Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, Jiaolong Yang
- Model Type: Flow-Matching Transformers with Sparse Voxel based 3D VAE
- Parameters: 4 Billion
- Input: Single Image
- Output: 3D Asset (Mesh with PBR Materials)
- Resolution: Varies from 512³ to 1536³ (Voxel Grid Resolution)

### Key Features

[Read the full model card (697 words)](https://savrn.com/models/trellis-2-4b/card)

## Identity and Version

Repository

microsoft/TRELLIS.2-4B

Publisher

Microsoft

Task

Image to 3d

Modality

Other

Library

trellis2

Parameters

Not stated by the source

Languages

en

Revision

af44b45f2e35a493886929c6d786e563ec68364d

First published

2025-12-01

Last updated

2025-12-27

## Files and Weights

22 files, 16.2 GB in total. The weights are 9 files totalling 16.2 GB in safetensors.

Weights9 files · 16.2 GB

Configuration11 files · 12.2 KB

Documentation1 file · 6.4 KB

Repository1 file · 1.5 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| ckpts/shape_dec_next_dc_f16c32_fp16.safetensors | Weights | 948.5 MB | e3b718d3e43e |
| ckpts/shape_enc_next_dc_f16c32_fp16.safetensors | Weights | 708.8 MB | f37c5ff5b983 |
| ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16.safetensors | Weights | 2.6 GB | 07cd0596f634 |
| ckpts/slat_flow_img2shape_dit_1_3B_512_bf16.safetensors | Weights | 2.6 GB | ec5e0917ef9b |
| ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16.safetensors | Weights | 2.6 GB | 580401269059 |
| ckpts/slat_flow_imgshape2tex_dit_1_3B_512_bf16.safetensors | Weights | 2.6 GB | 8371aa1c5d13 |
| ckpts/ss_flow_img_dit_1_3B_64_bf16.safetensors | Weights | 2.6 GB | ca01377c485b |
| ckpts/tex_dec_next_dc_f16c32_fp16.safetensors | Weights | 948.5 MB | 97ea69addea2 |
| ckpts/tex_enc_next_dc_f16c32_fp16.safetensors | Weights | 708.8 MB | dd109f75f84b |
| ckpts/shape_dec_next_dc_f16c32_fp16.json | Configuration | 678 B | — |
| ckpts/shape_enc_next_dc_f16c32_fp16.json | Configuration | 653 B | — |
| ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16.json | Configuration | 458 B | — |
| ckpts/slat_flow_img2shape_dit_1_3B_512_bf16.json | Configuration | 458 B | — |
| ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16.json | Configuration | 458 B | — |
| ckpts/slat_flow_imgshape2tex_dit_1_3B_512_bf16.json | Configuration | 458 B | — |
| ckpts/ss_flow_img_dit_1_3B_64_bf16.json | Configuration | 467 B | — |
| ckpts/tex_dec_next_dc_f16c32_fp16.json | Configuration | 705 B | — |
| ckpts/tex_enc_next_dc_f16c32_fp16.json | Configuration | 676 B | — |
| pipeline.json | Configuration | 4.2 KB | — |
| texturing_pipeline.json | Configuration | 3.0 KB | — |
| README.md | Documentation | 6.4 KB | — |
| .gitattributes | Repository | 1.5 KB | — |

## License and Download

License

mit

Access

Open weights, no gate

Download size

16.2 GB

[Download from Microsoft](https://huggingface.co/microsoft/TRELLIS.2-4B)

Released by Microsoft through its official repository on Hugging Face. [Read the license](https://opensource.org/license/mit).

## Built From

- Described by arXiv:2512.14692

## Memory Requirements

| Precision | Weights in memory |
| --- | --- |
| As published | 16.2 GB |

Weights only, from the published parameter count; the key-value cache and runtime add to this.

## Questions About TRELLIS.2-4B

### Can I use TRELLIS.2-4B commercially?

Yes. TRELLIS.2-4B is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

## Similar Models

Model · Image to 3d

### [TRELLIS-image-large](https://savrn.com/models/trellis-image-large)

[Microsoft](https://savrn.com/model-publishers/microsoft)

The image conditioned version of TRELLIS, a large 3D genetive model. It was introduced in the paper Structured 3D Latents for Scalable and Versatile 3D Generation.

Open weights mit trellis

[View model](https://savrn.com/models/trellis-image-large)

Model · Image to 3d

### [QtMeshEditor-models](https://savrn.com/models/qtmesheditor-models)

[Fernando Tonon](https://savrn.com/model-publishers/fernandotonon)

The models used by QtMeshEditor's AI-assisted authoring features. This repo is what the app downloads from at runtime (each model on first use, then it runs locally/offline). Licenses are per model — see the table and each dedicated repo. The dedicated repos carry the full model cards (I/O contracts, provenance, reproduction scripts) for anyone who wants the converted weights standalone. 1x-PBRifyNormalV3.onnx, 1x-PBRifyRoughnessV2.onnx, 1x-PBRifyHeight.onnx generate tangent-space normal / roughness / height maps from a single albedo texture. ONNX re-exports of the CC0 SPAN models from Kim2091/PBRifyRemix — all RealESRGANx2plus.onnx, RealESRGANx4plus.onnx — 2×/4× super-resolution. ONNX…

Open weights other onnx

[View model](https://savrn.com/models/qtmesheditor-models)

Model · Image to 3d

### [matchanu-1.5](https://savrn.com/models/matchanu-1-5)

[PGR](https://savrn.com/model-publishers/pygrassreal)

Matchanu 1.5 is PyGrassReal's generative 3D asset engine capable of synthesizing production-ready watertight 3D meshes (STL, OBJ, PLY) and real-time 3D Gaussian Splats from text descriptions or single 2D concept images. Designed for Industrial Prototyping (3D Printing), Game & AR/VR Assets, and Architectural Elements, Matchanu 1.5 generates clean topology, UV unwrappings, and volume-accurate meshes without the point cloud artifacts common to open-source prototypes. 1. Prompt & Image-to-3D Mesh: - Converts single-view product photos or text prompts into clean, manifold 3D meshes ready for Blender, Rhino, Maya, or Unreal Engine. 2. 3D Printing Ready (Watertight Topology): - Automatically…

Open weights other

[View model](https://savrn.com/models/matchanu-1-5)

Model · Image to 3d

### [ABot-Recon-Axera](https://savrn.com/models/abot-recon-axera)

[AXERA](https://savrn.com/model-publishers/axera-tech)

ABot-Recon(流式前馈三维重建:视频 → 相机位姿 + 世界坐标点云 + 置信度)编译到 Axera AX650N NPU 的模型文件。 模型与算法来自原作者:amap-cvlab/ABot-Recon(主页 ),本仓库只做 Axera NPU 的格式转换。 代码与使用说明:https://github.com/AXERA-TECH/ABot-Recon-Axera 三个 axmodel 由 pulsar2 编译,芯片 AX650N,支持 AXCL PCIe 卡和 AX650 片上两种运行方式。 打开 http://:8011。Python 调用: 片上运行需要板子 CMM 预留 ≥ 6 GB。 这些权重转换自 ABot-Recon 发布的权重,后者派生自 Pi3;再分发须保留对 Pi3 与 ABot-Recon 的署名与 CC BY-NC 4.0 条款。商业使用需另行获得书面授权。 - presentvalid 封顶 8(应为 11),长序列有轻微漂移。

Open weights cc-by-nc-4.0

[View model](https://savrn.com/models/abot-recon-axera)

Model · Image to 3d

### [MeshAxiom-Base](https://savrn.com/models/meshaxiom-base)

[Alaya Lab](https://savrn.com/model-publishers/alayalab)

Open weights meshaxiom

[View model](https://savrn.com/models/meshaxiom-base)

Model · Image to 3d

### [etri_recon](https://savrn.com/models/etri-recon)

[Jiwoo Chung](https://savrn.com/model-publishers/wldn0202)

Open weights other diffusers

[View model](https://savrn.com/models/etri-recon)

## Microsoft

[All models and datasets](https://savrn.com/model-publishers/microsoft)

## Versions

- [af44b45f2e35](https://savrn.com/models/trellis-2-4b/versions/af44b45f2e35) · current 2026-10-24

## Explore More

- [All image to 3d models](https://savrn.com/models/tasks/image-to-3d)
- [All models under mit](https://savrn.com/models/licenses/mit)
- [Model comparisons](https://savrn.com/models/comparisons)
- [The model directory](https://savrn.com/models)
- [Open model prices by host](https://savrn.com/ai-index/pricing/open-models)

## Source

- Repository metadata, read 2026-10-19.
- [Microsoft website](https://github.com/microsoft)
- [Hugging Face record](https://huggingface.co/microsoft/TRELLIS.2-4B)
- [How the hub is built](https://savrn.com/model-hub/methodology)