TRELLIS.2 Review

TRELLIS.2 is worth trying if you want to turn an object image into a textured 3D starting point for a game prop, illustration or scene concept. Its appeal is that you get both a shape and materials that respond to lighting. The catch is the work after generation: checking hidden surfaces, fixing geometry and making the model fit your project.

I’d start with Microsoft’s hosted demo before setting up a local installation. If you need an exact replica, a character ready to animate or a file ready to print, keep those requirements separate from whether the preview looks convincing.

What TRELLIS.2 makes from an image

Microsoft’s TRELLIS.2 takes an image and generates a mesh—the connected faces that form a 3D object—along with its surface appearance. The documented export is GLB, a file that can carry the shape and textures together. You can bring it into software such as Blender to inspect and edit it.

Its internal representation, called O-Voxel, stores shape and material information in small cells in 3D space. That lets it represent features such as open surfaces and hollow structures. It doesn’t mean it will infer every opening correctly from your picture.

Treat the official image workflow as the starting point. If your idea is only text, make a reference image first; the documented TRELLIS.2 pipeline doesn’t provide native text prompting. Nor should you assume that multi-image features from the original TRELLIS carry over. A plausible object generated from one view is still an interpretation, especially around the back.

How well TRELLIS.2 captures shape and materials

A model can look like your reference and still get important details wrong.

In a provider’s three-example comparison, both versions received the same AI-generated reference images, with no text prompts for the 3D step. TRELLIS.2 produced recognizable results for an Energy Sword and a Luffy figure where the original TRELLIS struggled. Both failed to reproduce the more complicated Mount Rushmore reference accurately. Those selected examples show potential improvement, but they don’t tell you how often your own images will succeed.

Materials are a meaningful part of the appeal. PBR, or physically based rendering, describes how surfaces react to light. Roughness controls how sharp or spread out reflections look; metallic settings distinguish metal from other surfaces. TRELLIS.2 also generates base color and opacity information. That gives you more to work with than a colored shape, though convincing reflections don’t prove the shape underneath is correct.

There are limits at small scales. The researchers acknowledge occasional holes and explain that closely spaced surfaces can merge or lose definition when they fall within the same cell. Fine geometry and appearance can suffer together. Higher resolution may help resolve smaller features, but it can’t supply missing evidence about the real object.

For your first attempt, use an image where the object’s outline is easy to distinguish. Rotate the result all the way around and temporarily view it without its textures. Check whether a groove is actually cut into the surface or merely painted on, and whether separate parts have fused together. Pay particular attention to surfaces the reference never showed. For a background prop, an invented underside may be acceptable; for a faithful product model, it may rule the result out.

What you need to run TRELLIS.2

The easiest first stop is the TRELLIS.2 Space under Microsoft’s Hugging Face account. It uses ZeroGPU, which offers free access subject to daily usage quotas and queues. Upload your image, choose Generate, inspect the preview, then use Extract GLB and Download GLB. The demo warns that extraction alone may take half a minute or more. Free access doesn’t mean unlimited generations.

For local use, look for Microsoft’s TRELLIS.2 GitHub repository and the microsoft/TRELLIS.2-4B weights on Hugging Face. Microsoft’s own setup calls for Linux and an NVIDIA GPU with at least 24GB of GPU memory; it reports verification on A100 and H100 hardware. That isn’t a promise of support for Windows, macOS, AMD or smaller cards.

Setup means downloading the repository with its supporting components, creating the software environment, installing dependencies, loading the weights and running the image pipeline before exporting. A local browser interface is included. The default installation uses PyTorch 2.6.0 with CUDA 12.4, and some supporting code has to be built on your computer. Choose this route if you’re comfortable troubleshooting that setup.

ComfyUI is another option: it added built-in TRELLIS.2 support to its main codebase in August 2026. Use a version containing that update and a matching workflow; older custom-node tutorials describe separate implementations. Built-in support doesn’t establish that every GPU, memory size or operating system works, so check the requirements for the exact workflow you intend to run.

Microsoft reports about 3 seconds at 512³ resolution, 17 seconds at 1024³ and 60 seconds at 1536³ on an H100. These are 3D generation resolutions, not texture sizes. They aren’t ordinary-desktop benchmarks or upload-to-finished-file times. Allow separately for loading, queues, export and your own editing.

The model and code use the MIT license, so there’s no model purchase required. Hardware, rented compute and separately operated hosted services can still cost money. A website with “Trellis” in its name isn’t necessarily Microsoft’s service.

Supporting software has its own terms. The published licenses for NVIDIA’s nvdiffrast and nvdiffrec rendering tools restrict use outside NVIDIA and its affiliates to research or evaluation. BRIA’s RMBG-2.0 background-removal weights require a separate agreement for commercial use. Check what your chosen setup includes before using it for paid work. The model license also doesn’t settle rights to your reference image or everything depicted in the output.

What TRELLIS.2 assets need before you use them

For a game, I’d first try TRELLIS.2 on a static prop you can inspect and replace easily. That’s a more manageable trial than building a character animation pipeline around its output.

Start by importing the GLB into Blender or your destination tool. Check orientation, set the intended size and examine the material under your scene’s lighting. Microsoft’s documented export defaults to opaque: transparency information remains in the texture but needs connecting to the material’s alpha or opacity input. Check the application’s transparency settings too, especially if cutout areas appear filled in.

Next, look for holes, stray pieces and surfaces that cross through one another. One artist’s illustrated Blender walkthrough reported importing a detailed model and merging duplicate vertices before making simple pose edits. That shows one way to edit the result, but it doesn’t establish automatic rigging or a repeatable cleanup time.

For games, also inspect the triangle count and texture size. Microsoft’s example export targets one million faces and a 4096-pixel texture; these are example settings, not a sensible budget for every asset. Reducing face count is called decimation. Check the result at the distance players will see it, because reducing geometry can also remove useful detail. After editing, inspect UVs—the layout that places a flat texture onto the model—for stretched or misplaced areas.

If a character needs to bend and move, expect a separate preparation step. You may need retopology, which rebuilds the surface with faces arranged for cleaner movement, plus a rig, the controls and skeleton used to animate it. A successful GLB import doesn’t establish good animation, collision behavior or game performance. Test those in the actual project.

For printing, inspect the solid shape, wall thickness and intended dimensions, then check the slicer’s layer preview. A slicer turns the model into printer instructions. Color and shiny materials can hide problems that become obvious in a physical object. Treat the output as a model to prepare, not proof that it will print successfully.

Count all this human work when deciding whether TRELLIS.2 helps. A quick generation followed by extensive rebuilding may be less useful than a modest-looking result that needs only small corrections.

How TRELLIS.2 compares with other ways to make 3D models

Compared with the original TRELLIS, the stronger reason to try TRELLIS.2 is its combination of shape generation and richer materials. The selected comparisons suggest improvement on some recognizable objects. They don’t justify discarding an older workflow that already handles your assets well. Use the same reference and judge the exported models, including how much correction each needs.

Hunyuan3D is the closest alternative to compare for image-generated assets with PBR materials. The TRELLIS.2 researchers report more plausible textures than Hunyuan3D 2.1 in their comparison. Their image-generation study used 100 AI-generated references, so it doesn’t settle performance on everyday photographs. It also doesn’t rank later Hunyuan versions: Tencent’s hosted API now lists 3.0 and 3.1. Compare the specific version and access route you would actually use.

If you want to copy an existing object faithfully, consider photogrammetry: reconstructing an object from overlapping photographs taken around it. Tools such as RealityScan support measured distance constraints to establish scale. That addresses a different need from asking TRELLIS.2 to imagine the unseen side. You still need to evaluate the scan’s accuracy.

For exact dimensions, controlled changes or parts that must fit together, manual modeling or CAD is the more appropriate route to investigate. Blender is also useful for continuing work on a generated mesh. There isn’t one best 3D program across those jobs. Choose based on whether you need a plausible visual, a measured copy or a deliberately designed shape.

When TRELLIS.2 is worth trying

Try TRELLIS.2 if you have a clear reference image, can accept some interpretation and are willing to edit the result. Concept props, scene layout and still-image work are sensible starting points. Its materials make it particularly interesting when you want to move beyond a flat reference and explore an object under different lighting.

I wouldn’t build a project with a tight deadline around it if you need exact replicas or characters you can animate straight away. What’s still uncertain is how often you’ll get a model you’re happy with without too much repair.

Start with one representative asset in the official demo. Download it, inspect the hidden surfaces and finish the corrections your project requires. If that process is manageable, test a few harder references before investing in local setup. The deciding result is the asset you can use—and the effort it took to get there.