TRELLIS.2：微软图生 3D 续作，O-Voxel 打破拓扑限制，4B 参数，3 秒 512³ 分辨率

by Mycelium Protocol

---

GitHub：microsoft/TRELLIS.2
论文：arxiv.org/abs/2512.14692
HF 模型：microsoft/TRELLIS.2-4B
HF Demo：huggingface.co/spaces/microsoft/TRELLIS.2
许可证：MIT
Stars：10,745
前作：microsoft/TRELLIS（CVPR’25 Spotlight，13,477 stars）

---

一、TRELLIS.2 是什么

TRELLIS 是微软研究院 2024 年末发布的结构化 3D 潜空间生成模型，入选 CVPR’25 Spotlight。TRELLIS.2 是它的直接续作，核心改变是3D 表示方式——从「有场」的等值面结构，换成「无场」的 O-Voxel 稀疏体素。

这不是参数更大的同一个模型，而是底层表示范式的升级。

---

二、为什么要换表示

传统 3D 生成（包括原版 TRELLIS）使用 SDF（有符号距离场）或占用场作为表示，再用 Marching Cubes 等方法提取网格。这套流程有一个根本限制：

等值面只能提取封闭、可定向、流形的曲面。

现实物体很多都不是这样的：

• 衣物、叶片、薄壳（开放表面）：Marching Cubes 处理后漏洞百出
• 铁丝网、网格结构（非流形几何）：完全无法正确表示
• 箱子里的物品（内部封闭结构）：直接被丢弃

O-Voxel 绕过了这个问题：不生成场，直接生成稀疏体素网格，每个体素携带几何和材质属性，导出时直接转换，没有有损的等值面提取步骤。

---

三、O-Voxel 是什么

O-Voxel（Open Voxel）是 TRELLIS.2 团队开发的一套新的 3D 表示方案：

• 稀疏体素结构：只存非空体素，内存高效
• 无场：不依赖距离场或占用场，避免拓扑约束
• 携带完整 PBR 属性：每个体素存储 Base Color、Roughness、Metallic、Opacity
• 双向即时转换：
  • 网格 → O-Voxel：单 CPU < 10 秒，渲染无关、优化无关
  • O-Voxel → 网格：CUDA < 100ms

Sparse 3D VAE 对 O-Voxel 做 16× 空间下采样，将资产编码进紧凑的潜空间，再用 vanilla DiT（扩散变换器）进行生成。

---

四、性能

模型：TRELLIS.2-4B（40 亿参数）
测试硬件：NVIDIA H100

分辨率	总耗时	形状 + 材质分解
512³	~3 秒	2s + 1s
1024³	~17 秒	10s + 7s
1536³	~60 秒	35s + 25s

512³ 分辨率 3 秒是非常快的速度，足以支持交互式工作流。

---

五、快速上手

安装（需要 NVIDIA GPU ≥ 24GB，CUDA 12.4，Linux）：

git clone -b main https://github.com/microsoft/TRELLIS.2.git --recursive
cd TRELLIS.2
. ./setup.sh --new-env --basic --flash-attn --nvdiffrast --nvdiffrec --cumesh --o-voxel --flexgemm

图生 3D（最简示例）：

from PIL import Image
from trellis2.pipelines import Trellis2ImageTo3DPipeline

pipeline = Trellis2ImageTo3DPipeline.from_pretrained("microsoft/TRELLIS.2-4B")
pipeline.cuda()

image = Image.open("your_image.png")
mesh = pipeline.run(image)[0]
mesh.simplify(16777216)  # nvdiffrast 上限

# 导出 GLB（WebP 纹理，支持透明度）
glb = o_voxel.postprocess.to_glb(
    vertices=mesh.vertices, faces=mesh.faces,
    attr_volume=mesh.attrs, coords=mesh.coords,
    attr_layout=mesh.layout, voxel_size=mesh.voxel_size,
    aabb=[[-0.5,-0.5,-0.5],[0.5,0.5,0.5]],
    decimation_target=1000000,
    texture_size=4096, remesh=True,
)
glb.export("output.glb", extension_webp=True)

Web Demo（本地运行）：

python app.py

不想自己部署的话，可以直接用 Hugging Face Spaces 上的演示。

---

六、与原版 TRELLIS 的对比

	TRELLIS（v1）	TRELLIS.2
发布时间	2024-12	2025-11
会议	CVPR’25 Spotlight	Tech Report
3D 表示	SLAT（稀疏结构化潜空间 + 等值面）	O-Voxel（无场稀疏体素）
开放表面	❌	✅
非流形几何	❌	✅
内部结构	❌	✅
PBR 材质	基础	完整（Base/Roughness/Metallic/Opacity）
参数量	未公开（多规格）	4B
训练代码	未完整公开	✅ 完整开放

---

七、训练代码完整开放

TRELLIS.2 开放了完整的训练流水线，包括：

1. 数据预处理：原始 3D 资产 → O-Voxel 格式（data_toolkit/）
2. SC-VAE 训练（形状和材质各一个）
3. Flow 模型训练（稀疏结构流 + 形状流 + 材质流）
4. 高分辨率微调（512 → 1024）

训练数据使用 Objaverse-XL（Sketchfab 子集）。这意味着可以用私有 3D 资产数据集微调一个垂直领域的模型——比如游戏资产、工业零件、家具。

---

八、社区生态

项目	说明
visualbruno/ComfyUI-Trellis2	ComfyUI 封装，直接拖拽工作流
UNES97/trellis-3d-docker	Docker 化部署方案
microsoft/TRELLIS.2-4B（HF）	官方模型权重
HF Spaces demo	无需本地环境，在线试用

---

图生 3D 的主要障碍一直是「拓扑限制」——生成的网格不干净，开放表面和复杂结构出问题。O-Voxel 在底层绕过了这个问题，而不是打补丁。训练代码全开让垂直领域微调成为可能。值得关注的方向。

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

TRELLIS.2: Microsoft’s Image-to-3D Successor — O-Voxel Breaks Topology Limits, 4B Params, 3-Second 512³

by Mycelium Protocol

---

GitHub: microsoft/TRELLIS.2
Paper: arxiv.org/abs/2512.14692
HF Model: microsoft/TRELLIS.2-4B
HF Demo: huggingface.co/spaces/microsoft/TRELLIS.2
License: MIT
Stars: 10,745
Predecessor: microsoft/TRELLIS (CVPR’25 Spotlight, 13,477 stars)

---

What TRELLIS.2 Is

TRELLIS was Microsoft Research’s structured 3D latent generation model, released late 2024 and accepted as a CVPR’25 Spotlight. TRELLIS.2 is its direct successor, with one fundamental change: the 3D representation — from iso-surface fields to O-Voxel (Open Voxel), a “field-free” sparse voxel structure.

This isn’t a larger version of the same model. It’s an upgrade in the underlying representation paradigm.

---

Why Change the Representation

Traditional 3D generation — including the original TRELLIS — uses SDF or occupancy fields, then extracts meshes via Marching Cubes or similar. This has a hard topological limit:

Iso-surfaces can only extract closed, orientable, manifold surfaces.

Most real objects violate this:

• Clothing, leaves, thin shells (open surfaces): Marching Cubes produces artifacts and holes
• Wire frames, mesh structures (non-manifold geometry): can’t be correctly represented
• Objects inside boxes (internal enclosed structures): silently discarded

O-Voxel sidesteps this entirely: skip the field, generate sparse voxel grids directly, each voxel carrying geometry and material attributes, with no lossy iso-surface extraction step at export.

---

What O-Voxel Is

O-Voxel is a new 3D representation developed by the TRELLIS.2 team:

• Sparse voxel structure: only non-empty voxels stored, memory-efficient
• Field-free: no SDF or occupancy field, no topological constraint
• Full PBR per voxel: Base Color, Roughness, Metallic, Opacity
• Instant bidirectional conversion:
  • Mesh → O-Voxel: <10s on a single CPU, rendering-free and optimization-free
  • O-Voxel → Mesh: <100ms on CUDA

A Sparse 3D VAE applies 16× spatial downsampling to encode assets into a compact latent space, then vanilla DiTs (diffusion transformers) handle generation.

---

Performance

Model: TRELLIS.2-4B (4 billion parameters)
Hardware: NVIDIA H100

Resolution	Total time	Shape + Material
512³	~3s	2s + 1s
1024³	~17s	10s + 7s
1536³	~60s	35s + 25s

3 seconds at 512³ is fast enough for interactive workflows.

---

Quickstart

Install (NVIDIA GPU ≥ 24GB VRAM, CUDA 12.4, Linux):

git clone -b main https://github.com/microsoft/TRELLIS.2.git --recursive
cd TRELLIS.2
. ./setup.sh --new-env --basic --flash-attn --nvdiffrast --nvdiffrec --cumesh --o-voxel --flexgemm

Image to 3D (minimal example):

from PIL import Image
from trellis2.pipelines import Trellis2ImageTo3DPipeline

pipeline = Trellis2ImageTo3DPipeline.from_pretrained("microsoft/TRELLIS.2-4B")
pipeline.cuda()

image = Image.open("your_image.png")
mesh = pipeline.run(image)[0]
mesh.simplify(16777216)

glb = o_voxel.postprocess.to_glb(
    vertices=mesh.vertices, faces=mesh.faces,
    attr_volume=mesh.attrs, coords=mesh.coords,
    attr_layout=mesh.layout, voxel_size=mesh.voxel_size,
    aabb=[[-0.5,-0.5,-0.5],[0.5,0.5,0.5]],
    decimation_target=1000000,
    texture_size=4096, remesh=True,
)
glb.export("output.glb", extension_webp=True)

For a web UI: python app.py. Or try the Hugging Face Spaces demo without any local setup.

---

vs Original TRELLIS

	TRELLIS (v1)	TRELLIS.2
Released	2024-12	2025-11
Venue	CVPR’25 Spotlight	Tech Report
Representation	SLAT (structured latent + iso-surface)	O-Voxel (field-free sparse voxel)
Open surfaces	❌	✅
Non-manifold geometry	❌	✅
Internal structures	❌	✅
PBR materials	Basic	Full (Base/Roughness/Metallic/Opacity)
Parameters	Multiple sizes (undisclosed)	4B
Training code	Partially available	✅ Fully open

---

Full Training Code Released

TRELLIS.2 opens its complete training pipeline:

1. Data preprocessing: raw 3D assets → O-Voxel format (data_toolkit/)
2. SC-VAE training (separate shape and texture VAEs)
3. Flow model training (sparse structure flow + shape flow + texture flow)
4. High-resolution fine-tuning (512 → 1024)

Training data: Objaverse-XL (Sketchfab subset). This makes domain-specific fine-tuning possible — game assets, industrial parts, furniture — with private 3D asset collections.

---

Community

Project	Description
visualbruno/ComfyUI-Trellis2	ComfyUI wrapper for drag-and-drop workflows
UNES97/trellis-3d-docker	Dockerized deployment
microsoft/TRELLIS.2-4B (HF)	Official model weights
HF Spaces demo	Try online without local setup

---

The persistent blocker for image-to-3D has been topology — generated meshes aren’t clean, open surfaces break, complex structures get mangled. O-Voxel addresses this at the representation level, not as a post-processing fix. With full training code open, vertical fine-tuning is now on the table.

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.