title: "TRELLIS.2-4B: Complete Guide to Microsoft's Revolutionary Image-to-3D Generation Model (2025)" description: "Complete guide to TRELLIS.2-4B: Microsoft's 4B parameter model for image-to-3D generation. Features O-Voxel technology, 3-60 second generation, full PBR materials, arbitrary topology support, and MIT open-source license. Includes installation, usage examples, benchmarks, and training guide." publishedAt: "2025-12-18" updatedAt: "2025-12-18" author: "CurateClick Team" tags: ["TRELLIS.2-4B", "Microsoft Research", "3D Generation", "Image-to-3D", "O-Voxel", "3D AI", "PBR Materials", "3D Modeling", "Computer Vision", "3D Assets", "Open Source", "MIT License"] featured: true seo: title: "TRELLIS.2-4B: Complete Guide to Microsoft's Revolutionary Image-to-3D Generation Model (2025)" description: "Complete guide to TRELLIS.2-4B: Microsoft's 4B parameter model for image-to-3D generation. Features O-Voxel technology, 3-60 second generation, full PBR materials, arbitrary topology support, and MIT open-source license. Includes installation, usage examples, benchmarks, and training guide." keywords: ["TRELLIS.2-4B", "Microsoft Research", "3D Generation", "Image-to-3D", "O-Voxel", "3D AI", "PBR Materials", "3D Modeling", "Computer Vision", "3D Assets", "Open Source", "MIT License", "3D Mesh Generation", "3D Asset Creation", "Neural 3D", "3D Deep Learning", "3D Reconstruction", "3D Content Generation"] canonical: "/blog/2025-trellis-2-4b"

TRELLIS.2-4B: The Complete Guide to Microsoft's Revolutionary 3D Generation Model (2025)

🎯 Core Highlights (TL;DR)

• TRELLIS.2-4B is Microsoft's state-of-the-art 4 billion parameter model for high-fidelity image-to-3D generation
• Introduces O-Voxel (Omni-Voxel), a breakthrough "field-free" representation handling arbitrary topologies including open surfaces and non-manifold geometry
• Achieves ultra-fast generation: 3 seconds for 512³ resolution, 17 seconds for 1024³ on NVIDIA H100
• Supports full PBR materials (Base Color, Metallic, Roughness, Opacity) for photorealistic rendering
• Compresses 1024³ assets into only ~9.6K latent tokens with negligible quality loss
• Open-source under MIT License with complete training code and 500K dataset

Table of Contents

1. What is TRELLIS.2-4B?
2. Evolution from TRELLIS to TRELLIS.2
3. Technical Innovations
4. Key Features and Capabilities
5. Performance Benchmarks
6. Installation and Setup
7. How to Use TRELLIS.2-4B
8. Comparison with Other 3D Generation Models
9. Training Your Own Model
10. Limitations and Considerations
11. Frequently Asked Questions
12. Conclusion and Next Steps

What is TRELLIS.2-4B?

TRELLIS.2-4B is Microsoft Research's latest breakthrough in 3D generative AI, representing a significant leap forward in image-to-3D conversion technology. As a 4 billion parameter model, it transforms single 2D images into fully textured, high-resolution 3D assets with unprecedented quality and speed.

Core Capabilities

• Input: Single RGB image
• Output: Fully textured 3D mesh with PBR materials
• Resolution: Supports 512³ to 1536³ voxel grid resolution
• Speed: 3-60 seconds depending on resolution (NVIDIA H100)
• License: MIT License (open-source)

💡 Key Innovation Unlike traditional methods that rely on implicit fields (SDF, NeRF) or iso-surface representations (Flexicubes), TRELLIS.2 uses a novel "field-free" approach that natively handles complex geometries without lossy conversions.

Research Background

Developed by a collaborative team from Tsinghua University and Microsoft Research, TRELLIS.2 builds upon the original TRELLIS model (CVPR'25 Spotlight) with fundamental architectural improvements. The research paper is available at arXiv:2512.14692.

Evolution from TRELLIS to TRELLIS.2

TRELLIS (First Generation)

The original TRELLIS introduced the concept of Structured LATent (SLAT) representation, enabling:

• Multiple output formats (Radiance Fields, 3D Gaussians, Meshes)
• Models up to 2B parameters
• Training on 500K diverse 3D objects

TRELLIS.2 Breakthrough

TRELLIS.2 represents a paradigm shift with:

✅ Native topology handling - No conversion artifacts
✅ Compact latent space - 16× spatial compression
✅ Instant processing - Rendering-free, optimization-free
✅ Full PBR support - Including transparency/translucency
✅ Higher resolution - Up to 1536³ voxel grids

Technical Innovations

1. O-Voxel: Omni-Voxel Representation

O-Voxel is the cornerstone innovation of TRELLIS.2, representing a "field-free" sparse voxel structure that simultaneously encodes geometry and appearance.

Geometry Component (f_shape)

• Flexible Dual Grids: Handles arbitrary topologies
• Sharp Edge Preservation: Maintains geometric details
• Topology Freedom: Supports open surfaces, non-manifold geometry, internal structures

Appearance Component (f_mat)

• Base Color: RGB texture information
• Metallic: Material reflectivity
• Roughness: Surface smoothness
• Alpha: Transparency/translucency support

⚠️ Technical Advantage Traditional iso-surface methods (SDF, Flexicubes) struggle with:

• Open surfaces (e.g., cloth, hair)
• Non-manifold geometry (e.g., intersecting surfaces)
• Internal structures (e.g., hollow objects)

O-Voxel handles all these cases natively without conversion artifacts.

2. SC-VAE: Sparse Compression VAE

The Sparse Compression 3D VAE employs a Sparse Residual Autoencoding scheme to achieve unprecedented compression ratios.

Key Features:

• Negligible perceptual degradation
• Efficient large-scale generative modeling
• Direct voxel compression without intermediate representations

3. Flow-Matching Transformer Architecture

TRELLIS.2-4B utilizes vanilla DiT (Diffusion Transformer) architecture with:

• 4 billion parameters
• Flow-matching training objective
• Efficient attention mechanisms for sparse data
• Multi-resolution training strategy

4. Instant Bidirectional Conversion

One of TRELLIS.2's most practical innovations is the ability to convert between meshes and O-Voxels instantly:

This enables:

• Rendering-free processing: No need for multi-view rendering
• Optimization-free workflow: Direct conversion without iterative refinement
• Minimalist pipeline: Simplified data preparation and post-processing

Key Features and Capabilities

High Quality and Resolution

TRELLIS.2-4B generates assets with exceptional fidelity across multiple resolutions:

📊 Generation Quality Metrics

Resolution: 512³
- Generation Time: 3 seconds (2s shape + 1s material)
- Detail Level: High
- Use Case: Rapid prototyping, real-time applications

Resolution: 1024³
- Generation Time: 17 seconds (10s shape + 7s material)
- Detail Level: Very High
- Use Case: Production assets, game development

Resolution: 1536³
- Generation Time: 60 seconds (35s shape + 25s material)
- Detail Level: Ultra High
- Use Case: Film production, high-end visualization

Arbitrary Topology Handling

Unlike traditional methods constrained by iso-surface representations, TRELLIS.2 robustly handles:

✔ Open Surfaces

• Cloth, curtains, flags
• Hair and fur
• Thin structures

✔ Non-manifold Geometry

• Intersecting surfaces
• Self-intersections
• Complex architectural elements

✔ Internal Structures

• Hollow objects
• Multi-layer constructions
• Enclosed cavities

Rich Texture Modeling with PBR

Full Physically Based Rendering (PBR) support enables photorealistic relighting:

✅ Best Practice The PBR material system is compatible with standard game engines (Unity, Unreal Engine) and 3D software (Blender, Maya), enabling seamless integration into production pipelines.

Shape-Conditioned Texture Generation

TRELLIS.2 supports two generation modes:

1. Image-to-3D: Generate complete 3D asset from single image
2. Texture Generation: Generate textures for existing 3D meshes with reference image

This flexibility allows:

• Re-texturing existing assets
• Style transfer to 3D models
• Texture variation generation

Performance Benchmarks

Generation Speed (NVIDIA H100)

Latent Space Efficiency

TRELLIS.2 achieves state-of-the-art compression while maintaining quality:

Reconstruction Accuracy vs. Latent Compactness

TRELLIS.2: ████████████████████ (Highest)
Method A:  ████████████░░░░░░░░
Method B:  ██████████░░░░░░░░░░
Method C:  ████████░░░░░░░░░░░░

Compactness (tokens per 1024³):
TRELLIS.2: ~9.6K
Method A:  ~25K
Method B:  ~40K
Method C:  ~60K

Hardware Requirements

Installation and Setup

Prerequisites

Before installing TRELLIS.2, ensure your system meets these requirements:

• Operating System: Linux (tested on Ubuntu 20.04+)
• GPU: NVIDIA GPU with 24GB+ VRAM
• CUDA Toolkit: Version 12.4 or higher
• Python: Version 3.8 or higher
• Conda: For dependency management

⚠️ Windows Users While primarily tested on Linux, Windows setup is possible but not officially supported. Refer to community discussions for Windows-specific configurations.

Step-by-Step Installation

1. Clone the Repository

git clone --recurse-submodules https://github.com/microsoft/TRELLIS.2.git
cd TRELLIS.2

2. Create Conda Environment

# Create new environment
conda create -n trellis2 python=3.10
conda activate trellis2

3. Install Dependencies

The installation script provides modular dependency installation:

# Install all dependencies for inference
. ./setup.sh --new-env --basic --xformers --flash-attn \
  --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrast

Installation Flags Explained:

4. Environment Configuration

Set environment variables for optimal performance:

export OPENCV_IO_ENABLE_OPENEXR=1
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True"

# For GPUs without flash-attn support (e.g., V100)
# export ATTN_BACKEND=xformers

# SPCONV algorithm selection
export SPCONV_ALGO=native  # Use 'auto' for benchmarking (slower first run)

5. Download Pre-trained Models

Models are automatically downloaded from Hugging Face on first use, or download manually:

# Models will be cached in ~/.cache/huggingface/
# No manual download required for basic usage

Troubleshooting Installation

Issue: CUDA version mismatch

# Check CUDA version
nvcc --version

# Set correct CUDA path
export PATH=/usr/local/cuda-12.4/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-12.4/lib64:$LD_LIBRARY_PATH

Issue: Out of memory during compilation

# Limit parallel compilation jobs
export MAX_JOBS=4

How to Use TRELLIS.2-4B

Basic Image-to-3D Generation

Here's a minimal example to generate a 3D asset from an image:

import os
os.environ['OPENCV_IO_ENABLE_OPENEXR'] = '1'
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

import cv2
import imageio
from PIL import Image
import torch
from trellis2.pipelines import Trellis2ImageTo3DPipeline
from trellis2.utils import render_utils
from trellis2.renderers import EnvMap
import o_voxel

# 1. Setup Environment Map for PBR rendering
envmap = EnvMap(torch.tensor(
    cv2.cvtColor(cv2.imread('assets/hdri/forest.exr', cv2.IMREAD_UNCHANGED), 
                 cv2.COLOR_BGR2RGB),
    dtype=torch.float32, device='cuda'
))

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

# 3. Load Image & Run Generation
image = Image.open("assets/example_image/T.png")
mesh = pipeline.run(image)[0]

# 4. Simplify mesh (nvdiffrast has 16M triangle limit)
mesh.simplify(16777216)

# 5. Render Video Preview
video = render_utils.make_pbr_vis_frames(
    render_utils.render_video(mesh, envmap=envmap)
)
imageio.mimsave("output.mp4", video, fps=15)

# 6. Export to GLB format
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,
    remesh_band         = 1,
    remesh_project      = 0,
    verbose             = True
)
glb.export("output.glb", extension_webp=True)

Advanced Usage: Multi-Resolution Generation

Generate assets at different resolutions based on your needs:

# High-speed generation (512³)
mesh_fast = pipeline.run(
    image,
    resolution=512,
    seed=42
)[0]

# Balanced quality (1024³) - Default
mesh_balanced = pipeline.run(
    image,
    resolution=1024,
    seed=42
)[0]

# Maximum quality (1536³)
mesh_ultra = pipeline.run(
    image,
    resolution=1536,
    seed=42
)[0]

Shape-Conditioned Texture Generation

Generate textures for existing 3D meshes:

from trellis2.pipelines import Trellis2TextureGenerationPipeline

# Load texture generation pipeline
texture_pipeline = Trellis2TextureGenerationPipeline.from_pretrained(
    "microsoft/TRELLIS.2-4B"
)
texture_pipeline.cuda()

# Load existing mesh and reference image
input_mesh = o_voxel.io.load_mesh("input_model.obj")
reference_image = Image.open("texture_reference.png")

# Generate texture
textured_mesh = texture_pipeline.run(
    mesh=input_mesh,
    image=reference_image,
    seed=42
)[0]

Batch Processing

Process multiple images efficiently:

import glob
from pathlib import Path

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

# Process all images in directory
image_paths = glob.glob("input_images/*.png")

for img_path in image_paths:
    image = Image.open(img_path)
    mesh = pipeline.run(image)[0]
    
    # Save with same filename
    output_name = Path(img_path).stem
    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
    )
    glb.export(f"output/{output_name}.glb", extension_webp=True)

Output Formats

TRELLIS.2 supports multiple output formats:

# Export to different formats
mesh.save_ply("output.ply")  # Gaussian representation
mesh.save_obj("output.obj")  # Traditional mesh
glb.export("output.glb")     # Optimized for web/games

Comparison with Other 3D Generation Models

TRELLIS.2 vs. Original TRELLIS

TRELLIS.2 vs. Other State-of-the-Art Models

✅ Competitive Advantage TRELLIS.2's combination of speed, quality, and topology flexibility makes it the most versatile solution for production-ready 3D asset generation.

When to Use TRELLIS.2

Best Use Cases:

• Production asset creation for games and films
• Rapid prototyping and concept visualization
• E-commerce 3D product visualization
• AR/VR content creation
• Architectural visualization
• Digital twin creation

Consider Alternatives When:

• You need text-only input (use TRELLIS-text models)
• You require real-time generation on mobile devices
• You need extremely high polygon counts (>10M triangles)
• You're working with specific artistic styles (may need fine-tuning)

Training Your Own Model

TRELLIS.2 provides complete training code for researchers and developers who want to:

• Fine-tune on custom datasets
• Experiment with architecture modifications
• Train domain-specific models

Training Dataset: TRELLIS-500K

Microsoft provides TRELLIS-500K, a curated dataset containing 500,000 high-quality 3D assets from:

All assets are filtered based on aesthetic scores and quality metrics.

Training Pipeline Overview

The training process follows a multi-stage approach:

Stage 1: VAE Training
├── Sparse Structure VAE (ss_vae)
└── SLat VAE with Decoders (slat_vae)
    ├── Gaussian Decoder
    ├── Radiance Field Decoder
    └── Mesh Decoder

Stage 2: Flow Model Training
├── Sparse Structure Flow (ss_flow)
└── SLat Flow (slat_flow)

Training Configuration

Example configurations are provided in the configs/ directory:

VAE Training:

# Train Sparse Structure VAE
python train.py \
  --config configs/vae/ss_vae_conv3d_16l8_fp16.json \
  --output_dir outputs/ss_vae \
  --data_dir /path/to/TRELLIS-500K \
  --num_gpus 8

# Train SLat VAE with Gaussian Decoder
python train.py \
  --config configs/vae/slat_vae_enc_dec_gs_swin8_B_64l8_fp16.json \
  --output_dir outputs/slat_vae_gs \
  --data_dir /path/to/TRELLIS-500K \
  --num_gpus 8

Flow Model Training:

# Train Image-conditioned Flow Model
python train.py \
  --config configs/generation/slat_flow_img_dit_L_64l8p2_fp16.json \
  --output_dir outputs/slat_flow_img \
  --data_dir /path/to/TRELLIS-500K \
  --num_nodes 4 \
  --num_gpus 8 \
  --master_addr $MASTER_ADDR \
  --master_port $MASTER_PORT

Multi-Node Distributed Training

For large-scale training across multiple machines:

# Node 0 (Master)
python train.py \
  --config configs/generation/slat_flow_img_dit_L_64l8p2_fp16.json \
  --output_dir outputs/slat_flow_img_distributed \
  --data_dir /path/to/TRELLIS-500K \
  --num_nodes 4 \
  --node_rank 0 \
  --num_gpus 8 \
  --master_addr 192.168.1.100 \
  --master_port 29500

# Node 1
python train.py \
  --config configs/generation/slat_flow_img_dit_L_64l8p2_fp16.json \
  --output_dir outputs/slat_flow_img_distributed \
  --data_dir /path/to/TRELLIS-500K \
  --num_nodes 4 \
  --node_rank 1 \
  --num_gpus 8 \
  --master_addr 192.168.1.100 \
  --master_port 29500

# Repeat for nodes 2 and 3...

Fine-tuning on Custom Data

To fine-tune TRELLIS.2 on your own dataset:

1. Prepare Data: Convert your 3D assets to O-Voxel format
2. Configure Training: Modify config files for your dataset
3. Resume from Checkpoint: Load pre-trained weights

python train.py \
  --config configs/generation/slat_flow_img_dit_L_64l8p2_fp16.json \
  --output_dir outputs/custom_finetuned \
  --data_dir /path/to/custom/dataset \
  --load_dir microsoft/TRELLIS.2-4B \
  --num_gpus 8

Training Hardware Requirements

Limitations and Considerations

Known Limitations

1. Geometric Artifacts

Issue: Generated meshes may occasionally contain small holes or minor topological discontinuities.

Impact:

• Affects applications requiring watertight geometry (3D printing, simulation)
• More common in high-complexity models with intricate details

Mitigation:

# Use provided post-processing scripts
from trellis2.utils import mesh_repair

cleaned_mesh = mesh_repair.fill_holes(mesh, max_hole_size=100)
cleaned_mesh = mesh_repair.remove_degenerate_faces(cleaned_mesh)

2. Base Model Without Alignment

Issue: TRELLIS.2-4B is a pre-trained foundation model without human preference alignment (RLHF).

Impact:

• Output style reflects training data distribution
• May require multiple generations to achieve desired aesthetic
• Not optimized for specific artistic styles

Recommendations:

• Generate multiple variants with different seeds
• Use post-processing for style refinement
• Consider fine-tuning for specific aesthetic requirements

3. Input Image Quality Dependency

Issue: Output quality heavily depends on input image characteristics.

Best Practices:

• Use high-resolution images (512×512 minimum, 1024×1024 recommended)
• Ensure clear object visibility with minimal occlusion
• Prefer images with good lighting and contrast
• Avoid heavily compressed or noisy images

4. Memory Requirements

Issue: High-resolution generation requires substantial GPU memory.

Memory Optimization:

# Enable memory-efficient settings
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

# Use gradient checkpointing during inference
pipeline.enable_memory_efficient_attention()

Responsible AI Considerations

⚠️ Important Notice TRELLIS.2 is a research project. Responsible AI considerations were factored into all stages:

• Dataset Curation: Public datasets reviewed for harmful content and PII
• Potential Bias: Internet-sourced data may contain inherent biases
• Intended Use: Academic and research purposes only
• Commercial Use: Requires careful evaluation of generated content

Ethical Guidelines:

• Do not generate content that infringes intellectual property rights
• Avoid creating misleading or deceptive 3D representations
• Respect privacy and consent when generating assets based on real objects
• Consider cultural sensitivity in generated content

Performance Considerations

Factors Affecting Generation Quality:

1. Input Image Characteristics

   • Resolution and clarity
   • Lighting conditions
   • Object visibility and occlusion
   • Background complexity

2. Generation Parameters

   • Resolution setting (512³ vs 1024³ vs 1536³)
   • Random seed selection
   • Sampling steps (if configurable)

3. Hardware Configuration

   • GPU model and memory
   • CUDA version compatibility
   • Driver version

Frequently Asked Questions

Q: What is the difference between TRELLIS and TRELLIS.2?

A: TRELLIS.2 represents a fundamental architectural upgrade from the original TRELLIS. The key differences are:

• Representation: TRELLIS uses SLAT (Structured Latent), while TRELLIS.2 uses O-Voxel (Omni-Voxel), a "field-free" approach
• Topology: TRELLIS.2 natively handles arbitrary topologies including open surfaces and non-manifold geometry, which TRELLIS cannot
• Efficiency: TRELLIS.2 compresses 1024³ assets into ~9.6K tokens vs. ~25K in TRELLIS
• Materials: TRELLIS.2 supports full PBR including transparency, while TRELLIS has partial support
• Processing: TRELLIS.2 offers instant mesh conversion (<100ms with CUDA) vs. multi-stage rendering in TRELLIS

Q: Can I use TRELLIS.2 for commercial projects?

A: Yes, TRELLIS.2 is released under the MIT License, which permits commercial use. However:

• Verify that generated assets don't infringe on existing intellectual property
• The model is a base model without alignment, so output quality may vary
• Some submodules may have different licenses (check the LICENSE file)
• Consider the ethical implications of AI-generated content in your use case

Q: What GPU do I need to run TRELLIS.2?

A: Minimum requirements:

• GPU: NVIDIA GPU with 24GB VRAM (e.g., RTX 3090, A5000, A100)
• Resolution: 512³ requires 16GB, 1024³ requires 24GB, 1536³ requires 40GB+
• Tested on: NVIDIA A100 and H100 GPUs
• Not supported: AMD GPUs, Apple Silicon (MPS), CPU-only inference

For optimal performance, use NVIDIA H100 or A100 (80GB) GPUs.

Q: How do I improve generation quality?

A: Follow these best practices:

1. Input Image Quality:

   • Use high-resolution images (1024×1024 recommended)
   • Ensure good lighting and clear object visibility
   • Remove complex backgrounds if possible
   • Avoid heavily compressed or noisy images

2. Generation Settings:

   • Use higher resolution (1024³ or 1536³) for detailed assets
   • Try different random seeds (generate 3-5 variants)
   • Experiment with different input angles if available

3. Post-Processing:

   • Use mesh repair tools for geometric artifacts
   • Apply texture enhancement in 3D software
   • Optimize topology for your specific use case

Q: Can TRELLIS.2 generate 3D assets from text prompts?

A: TRELLIS.2-4B is specifically designed for image-to-3D generation. For text-to-3D, you have two options:

1. Two-stage approach (Recommended):

   • Use a text-to-image model (DALL-E, Midjourney, Stable Diffusion)
   • Feed generated image to TRELLIS.2-4B
   • This typically produces better results

2. Use TRELLIS text models:

   • TRELLIS-text-base (342M)
   • TRELLIS-text-large (1.1B)
   • TRELLIS-text-xlarge (2.0B)
   • Note: These are from TRELLIS v1, not TRELLIS.2

Q: How long does generation take?

A: Generation time depends on resolution and hardware:

On NVIDIA H100:

• 512³: ~3 seconds (2s shape + 1s material)
• 1024³: ~17 seconds (10s shape + 7s material)
• 1536³: ~60 seconds (35s shape + 25s material)

On NVIDIA A100 (40GB):

• 512³: ~5 seconds
• 1024³: ~30 seconds
• 1536³: ~120 seconds

Older GPUs (RTX 3090, A6000) will be proportionally slower.

Q: What output formats are supported?

A: TRELLIS.2 supports multiple industry-standard formats:

• GLB/GLTF: Optimized for web, game engines (Unity, Unreal), and AR/VR
• PLY: Point cloud format, useful for Gaussian splatting
• OBJ: Traditional mesh format for 3D modeling software
• Mesh with PBR: Full material properties (Base Color, Metallic, Roughness, Alpha)

All formats include full PBR material information where applicable.

Q: Can I train TRELLIS.2 on my own dataset?

A: Yes, the complete training code is provided. You can:

1. Fine-tune the pre-trained model on your custom dataset
2. Train from scratch if you have sufficient data (100K+ assets recommended)
3. Modify architecture for research purposes

Requirements:

• Convert your 3D assets to O-Voxel format using provided tools
• Minimum 8× NVIDIA A100 GPUs for fine-tuning
• 32× A100 GPUs for full training of 4B model
• Training time: 1-4 weeks depending on model size

Q: Does TRELLIS.2 work on Windows?

A: TRELLIS.2 is primarily developed and tested on Linux (Ubuntu 20.04+). Windows support is:

• Not officially supported by the development team
• Possible with community workarounds (see GitHub issues)
• Recommended approach: Use WSL2 (Windows Subsystem for Linux) with GPU passthrough

For production use, Linux is strongly recommended.

Q: How does TRELLIS.2 handle transparent or translucent objects?

A: TRELLIS.2 has native support for transparency through the Alpha channel in its PBR material system:

• Opacity/Alpha attribute is part of the O-Voxel representation
• Supports both binary transparency (glass) and gradient translucency (smoke, water)
• Exports correctly to GLB format with alpha channel preserved
• Compatible with standard rendering engines that support PBR

This is a significant advantage over methods that only support opaque surfaces.

Q: What is the TRELLIS-500K dataset?

A: TRELLIS-500K is the training dataset for TRELLIS.2, containing:

• 500,000 curated 3D assets from multiple sources
• Filtered based on aesthetic scores and quality metrics
• Includes diverse categories: objects, furniture, toys, architectural elements
• Publicly available for research purposes
• Comes with data preparation toolkits for processing custom assets

Sources: Objaverse(XL), ABO, 3D-FUTURE, HSSD, Toys4k

Conclusion and Next Steps

Summary

TRELLIS.2-4B represents a significant breakthrough in 3D generative AI, offering:

✅ Unmatched Versatility: Handles arbitrary topologies including open surfaces, non-manifold geometry, and internal structures
✅ Exceptional Efficiency: 3-60 second generation time with compact 9.6K token representation
✅ Production-Ready Quality: Full PBR materials with photorealistic rendering capabilities
✅ Open Research: MIT License with complete training code and 500K dataset
✅ Minimalist Pipeline: Instant, optimization-free mesh conversion

Getting Started Checklist

• Verify hardware requirements (24GB+ NVIDIA GPU)
• Install CUDA Toolkit 12.4+
• Clone repository and install dependencies
• Download or prepare test images
• Run basic image-to-3D generation example
• Experiment with different resolutions and settings
• Export to GLB format for use in your pipeline

Recommended Next Steps

For Researchers:

1. Explore the technical paper: arXiv:2512.14692
2. Download TRELLIS-500K dataset for analysis
3. Experiment with architecture modifications
4. Benchmark against your own methods

For Developers:

1. Integrate TRELLIS.2 into your 3D content pipeline
2. Build applications using the API
3. Optimize for your specific hardware configuration
4. Contribute to the open-source project

For Artists and Designers:

1. Test with various input images to understand capabilities
2. Develop workflows combining text-to-image and TRELLIS.2
3. Experiment with post-processing in 3D software
4. Share results and feedback with the community

Resources and Links

Community and Support

• GitHub Issues: Report bugs and request features
• Discussions: Share results and ask questions
• Research Collaboration: Contact the authors for academic partnerships
• Commercial Inquiries: Review MIT License terms and conditions

Final Thoughts

TRELLIS.2-4B pushes the boundaries of what's possible in 3D generative AI, combining cutting-edge research with practical usability. Whether you're building the next generation of 3D content tools, conducting academic research, or creating immersive experiences, TRELLIS.2 provides a powerful foundation for innovation in 3D generation.

The open-source nature of the project, combined with comprehensive documentation and pre-trained models, makes it accessible to a wide range of users—from researchers exploring new architectures to developers building production applications.

Start generating high-quality 3D assets today with TRELLIS.2-4B!

Last Updated: December 2025
Model Version: TRELLIS.2-4B
License: MIT License

TRELLIS.2-4B Complete Guide