How to Use Microsoft TRELLIS 2: Official Ways, Steps & Demo

Last updated: October 4, 2026

There are four official ways to use Microsoft TRELLIS 2: clone the official GitHub repository and run it yourself, open the Hugging Face Spaces demo in your browser, drop the official workflow into ComfyUI, or use an online platform that runs the same model without any setup. GitHub gives you full control but the repo states you need an NVIDIA GPU with at least 24GB of memory; the hosted demo is free but runs on ZeroGPU quotas; ComfyUI fits if you already live in node graphs. This guide walks through each route with its real costs, then the step-by-step workflows — image-to-3D and text-to-3D — that stay the same wherever you run it.

What Is TRELLIS 2?

TRELLIS.2 is Microsoft Research's open-source image-to-3D model. In the project's own words, it is "an open-source 4B-parameter image-to-3D model producing up to 1536³ PBR textured assets, built on native 3D VAEs with 16× spatial compression" (official project page, accessed October 2026). Feed it one photo, get back a textured 3D asset you can open in Blender, drop into Unity, or send to a printer. The successor to TRELLIS v1, it handles thin geometry, transparent materials, and complex topology that tripped up earlier models. If you want the architecture deep-dive — O-Voxel, structured latents, the CVPR paper — see our complete TRELLIS.2-4B guide.

Official Ways to Use TRELLIS 2 at a Glance

Before picking a lane, here's the whole map. These are the channels Microsoft itself provides or endorses:

The details behind each row — including one licensing nuance most posts get wrong — are below.

What You Need Before Starting

That 24GB figure is not a typo or a conservative estimate. The official README states: "Hardware: An NVIDIA GPU with at least 24GB of memory is necessary. The code has been verified on NVIDIA A100 and H100 GPUs" (microsoft/TRELLIS.2, accessed October 2026). Older guides quoting 16GB are describing earlier releases or trimmed workflows. For most users, an online platform is the fastest way to start. If you want to attempt local installation anyway, see our TRELLIS 2 installation guide — and our low-VRAM workarounds if your card is smaller.

Try TRELLIS 2 online — no installation required

The Official Ways to Use TRELLIS 2 (Channel Map)

GitHub: the official open-source repository

The repo at github.com/microsoft/TRELLIS.2 is the source of truth. You get the inference code, pretrained 4B checkpoints, and a minimal Python example built around Trellis2ImageTo3DPipeline. The README's usage section is honest about what a run looks like: you set up an environment map from an HDR file, load your image through OpenCV, and let the pipeline do the rest. Prerequisites, quoted directly: tested on Linux only, 24GB GPU minimum, CUDA Toolkit 12.4 recommended, Conda for environment management. A community tester who ran it in August 2026 reported the main checkpoint plus the background-removal model takes about 18GB of disk, so budget storage too (Sayed Amman Akhtar, Medium, August 2026).

Hugging Face Spaces: the official hosted demo

Microsoft runs a demo space on Hugging Face — 2.18k likes and running on ZeroGPU hardware as of October 2026. It works, and it's the only Microsoft-operated way to try the model without installing anything. The trade-off is ZeroGPU's quota system: you get a share of GPU time, and at busy hours your job waits in line. For a one-off experiment, fine. For iterating on a client asset, the waiting adds up.

ComfyUI: the official workflow for pipelines

If your work already lives in ComfyUI, you don't need to leave it. The official ComfyUI documentation ships a TRELLIS.2 workflow: load one image, toggle the Boolean "Switch to Trellis2" node to true (the same workflow also contains the Pixal3D pipeline), and out comes a model with full PBR textures. This is the least-friction official route for anyone doing node-based production work — you keep your existing pre/post-processing nodes and swap only the generation core.

License boundary: MIT code, research positioning

Here's the nuance. The official README says: "This model and code are released under the MIT License" (microsoft/TRELLIS.2, accessed October 2026) — commercial use of your outputs is permitted. Two footnotes matter. First, the render dependencies (nvdiffrast, nvdiffrec) carry their own license terms. Second, the project page frames the work as "purely a research project" in its Responsible AI notes. Those statements coexist: the license governs what you may do; the positioning tells you the model ships without production SLAs. We treat MIT as decisive for commercial questions and the research framing as a reason to QA outputs before shipping them.

Method 1: Image-to-3D Generation

Image-to-3D is the most popular workflow. You upload a single photo and TRELLIS 2 reconstructs a full 3D model.

Step 1: Prepare Your Input Image

The quality of your output depends heavily on the input image. Follow these guidelines:

Ideal input images have:

• A single, clearly defined subject
• Clean background (solid color or transparent PNG)
• Even, diffused lighting
• Resolution of 512x512 pixels or higher
• The subject centered and fully visible

Avoid:

• Multiple overlapping objects
• Blurry or low-resolution photos
• Harsh shadows or extreme lighting
• Significant occlusion (parts hidden behind other objects)

Step 2: Upload Your Image

On our platform:

1. Navigate to the 3D generation page
2. Upload your image via drag-and-drop or file picker
3. The system automatically detects the subject and prepares it for 3D generation

Running locally from the official repo, the README's minimal example loads your image and calls the pipeline directly in Python — no wrapper script needed.

Step 3: Configure Generation Parameters

TRELLIS 2 exposes several parameters that control output quality:

Recommended settings by use case:

Step 4: Generate and Review

Click generate and wait approximately 3-10 seconds (depending on resolution and hardware). After generation, review your model:

1. Rotate the model — check all angles, not just the front
2. Check the back side — AI estimates the unseen portions, which are less accurate
3. Inspect fine details — look for symmetry issues or distorted features
4. Evaluate texture quality — textures should match the original image

If the result needs improvement, try:

• A different source image angle
• Increasing sampling steps
• Adjusting the guidance scale

Step 5: Export Your Model

TRELLIS 2 supports multiple export formats:

Choose GLB for game development, STL for 3D printing, and Gaussian Splatting for web-based 3D experiences.

Export your 3D model in any format — try it free

Method 2: Text-to-3D Generation

Text-to-3D lets you describe what you want in natural language and TRELLIS 2 generates it.

Step 1: Write an Effective Prompt

Good prompts are specific and descriptive. Here's a formula that works well:

[Subject] + [Key Features] + [Style/Material] + [Optional: Color, Pose, etc.]

Example prompts:

Step 2: Generate from Text

On our platform:

1. Switch to Text-to-3D mode
2. Enter your prompt
3. Optionally add a negative prompt (what to avoid)
4. Click generate

Locally, the same pipeline call accepts a text prompt instead of an image.

Step 3: Iterate and Refine

Text-to-3D often requires iteration. If the first result isn't what you envisioned:

1. Add more detail to the prompt
2. Use a reference image — combine text with an image for more control
3. Try different seeds — the same prompt can produce varied results
4. Use negative prompts — specify what you don't want ("blurry", "low quality", "deformed")

Can TRELLIS 2 Take Multiple Images? (Multi-View)

Short answer: the official model does not. The GitHub README and roadmap describe single-image input only — there's no multi-view option in the official pipeline, and nothing on the roadmap suggesting one is coming.

What you'll find instead are platform-level workarounds. Some services (ours included) blend several photos of the same object to give the generator more geometric hints, which mostly helps the back side — the part a single photo can't show. If you have front, side, and back shots of your subject, an online multi-view flow will serve you better than anything you can wire up from the official repo today. One honest caveat from testing: multi-view improves silhouette and back-face accuracy, but it won't rescue a subject whose sides were never photographed.

Official Channels vs Online Tool: Which Should You Use?

Three constraints decide it: queue access, GPU ownership, and how much you value your afternoon.

In our experience running a generation platform, the fork in the road is simple: if you have an A100 sitting idle and want to hack on the pipeline itself, clone the repo. If you're producing assets for clients on a laptop, the 24GB requirement alone disqualifies the official routes — and that's most people reading this. We built our platform precisely because we kept hitting that wall ourselves.

What You Can Do with the Output

For Game Development

Export as GLB with these considerations:

• Polygon count: AI models often generate 50k-200k polygons — decimate to 10k-50k for real-time rendering
• Textures: Check UV mapping quality in Blender
• Scale: Set real-world units in your 3D software

For 3D Printing

Export as STL or OBJ:

• Run mesh repair to ensure watertight geometry
• Check for non-manifold edges and flipped normals
• Verify scale matches your intended print size

For Web/AR

Export as 3D Gaussian Splatting or GLB:

• Gaussian Splatting provides real-time rendering in browsers
• GLB is widely supported in WebXR frameworks
• Optimize file size for web delivery

If you're heading into Blender next, our Blender post-processing walkthrough covers cleanup and UV checks step by step. (A longer video version of a local workflow exists too — Benji's AI Playground's December 2025 test on YouTube is worth 13 minutes if you want to watch someone run it end to end.)

Parameter Tuning Deep Dive

Sampling Steps

Sampling steps control the denoising process. More steps produce cleaner geometry and sharper textures:

Guidance Scale

Guidance scale controls how closely the output follows the input. Think of it as "creativity vs. accuracy":

• Low (1-5): More creative, may deviate from input
• Medium (5-10): Balanced creativity and accuracy
• High (10-20): Strict adherence to input, less variation

Resolution Trade-offs

Higher resolution means more detail but requires more VRAM and time:

Output Format Guide

(Consolidated into "What You Can Do with the Output" above.)

Resources

• Source code: github.com/microsoft/TRELLIS.2 (MIT license)
• Official demo: Hugging Face Spaces — microsoft/TRELLIS.2
• Official ComfyUI workflow: docs.comfy.org/tutorials/3d/trellis2
• Model weights: Hugging Face — TRELLIS.2-4B
• Research paper: "Native and Compact Structured Latents for 3D Generation" — arXiv:2512.14692
• v1 paper: "Structured 3D Latents for Scalable and Versatile 3D Generation" — arXiv:2412.01506

Common Issues and Solutions

TRELLIS 2 vs Other 3D Generation Tools

According to community testing on Reddit and benchmarks published by 3D AI Studio, TRELLIS 2 currently leads in generation speed and overall output quality among open-source 3D generation models.

Best Practices Summary

1. Pick the official channel that matches your hardware — the map above exists so you don't burn an evening on a setup your GPU can't run
2. Prepare your input — a clean, well-lit image dramatically improves results
3. Start with default settings — generate a preview before tweaking parameters
4. Review from all angles — the back side is estimated and may need attention
5. Choose the right format — match your export format to your end goal
6. Post-process — use Blender or MeshLab for final cleanup and optimization

FAQ

How do you use TRELLIS 2?

Pick one of the four official routes from the map above: GitHub (local, 24GB GPU), the Hugging Face demo (browser, free), the official ComfyUI workflow (node graphs), or an online platform. Then it's the same three moves everywhere — feed it one image or a text prompt, generate, export as GLB/STL/OBJ.

What is Microsoft TRELLIS 2?

A 4-billion parameter open-source model from Microsoft Research that turns one image into a textured 3D asset, per the official project page (accessed October 2026). Code and weights are on GitHub under MIT.

Can TRELLIS 2 be used for 3D printing?

Yes — export as STL and repair the mesh for watertight geometry. Since STL drops textures, print-first users often generate GLB first to keep the color version, then export STL from the same model. A textured mesh from TRELLIS 2 imports straight into Blender, Unity, or a slicer of your choice.

Is TRELLIS 2 any good?

For a free, open model, it sets the pace. As Progressive Robot put it in April 2026, it "focuses on compact structured latents that can preserve geometry, appearance, and physically based material details in one workflow" — and our own comparison testing puts it ahead of Tripo3D and Meshy AI on speed and resolution. Where it's merely fine: fully unseen back sides and very thin objects. Budget a post-processing pass for those.

Can I use TRELLIS 2 commercially?

The official README releases the model and code under the MIT License, which permits commercial use of what you generate. Note the dependencies (nvdiffrast, nvdiffrec) carry their own terms, the project positions itself as research, and input images you don't own may carry their own rights. Verify your inputs; the license covers the model, not your sources.

How much VRAM does TRELLIS 2 need?

The official repo requires an NVIDIA GPU with at least 24GB of memory, verified on A100 and H100. Our platform's quality presets trade resolution against VRAM (see the table above) — at 1536 resolution you're in 24GB territory, which is why browser-based generation exists.

What file formats does TRELLIS 2 export?

GLB, OBJ, STL, 3D Gaussian Splatting (.ply), and NeRF (.npz). GLB for game engines, STL for printing, Gaussian Splatting for web viewers.

How does TRELLIS 2 compare to Tripo3D and Meshy AI?

TRELLIS 2 generates in about 3 seconds versus Tripo3D's ~10s and Meshy's ~30s, tops out at 1536³, and is fully open source. The comparison table above has the cell-by-cell detail.

Get Started with TRELLIS 2

The fastest way to try TRELLIS 2 is through our online platform:

Generate Your First 3D Model — Free

No GPU needed. No Python installation. Upload an image or describe what you want and get a production-ready 3D model in seconds.

Start Creating 3D Models Now →

Related articles:

• What is TRELLIS 3D?: learn about the AI model behind our 3D generation
• How to Install TRELLIS 2: set up TRELLIS 2 locally on your machine
• How to Test TRELLIS 2: evaluate output quality with benchmarks and comparisons
• How to Turn an Image into a 3D Model: broader guide covering multiple methods