Tools
·
5 MIN
Image-to-3D with TRELLIS: 3D assets from a prompt
TRELLIS turns one image or a prompt into a GLB 3D asset, MIT-licensed. What it nails, and the cleanup it still needs.
LOCATION
Worldwide
SERIES
Computational & 3D Design
AUTHOR
Aashwin Shrivastava
PUBLISHED
TRELLIS, the open 3D generator from Microsoft Research, turns a single image or a text prompt into a usable 3D asset in under two minutes on a consumer GPU, and it ships under the MIT license. For concepting, web and AR previews, and e-commerce mockups, that is a genuine shortcut. For game or CAD pipelines, the honest framing is different: treat the output as a fast starting block that still needs cleanup, not a finished asset.
This piece covers what TRELLIS is, how it works, where it fits in a real asset pipeline, and the points where it stops being a one-click answer.
01.
What TRELLIS is
TRELLIS is a 3D generative model that accepts an image, several images, or a text prompt and produces a 3D asset.
🔸 Structured latents (SLAT). The core idea is one unified latent that fuses a sparse 3D voxel grid for structure with dense visual features for appearance. Decoupling shape from look is what lets the same latent become different outputs.
🔸 Three outputs from one latent. Radiance fields, 3D Gaussians, or a mesh exported as GLB. GLB is the practical target, since it carries materials into web and AR viewers.
🔸 Open and MIT-licensed. The image workhorse, TRELLIS-image-large, is 1.2B parameters. The model and most code are MIT, which matters for commercial use. A newer 4B line adds full physically based materials, worth watching as it matures.
🔸 Trained at scale. Around 500,000 curated objects, with the paper landing as a CVPR 2025 Spotlight.
This sits next to the parametric, controllable end of 3D work we cover in Grasshopper for computational design. Generative 3D and parametric 3D solve different problems.
02.
How it works
The pipeline is short to describe and useful to understand before you commit a GPU to it.
Input. A single image, multiple views, or a text prompt.
Structured latent. A two-stage rectified-flow transformer first generates the sparse structure, then fills in the SLAT features.
Decode. The latent decodes to your chosen format: mesh, Gaussians, or radiance field.
In practice, plan for about 16GB of VRAM as a floor, with 24GB comfortable. Generation runs in under two minutes on an RTX 4090. Community builds push the memory needs lower, which is what makes TRELLIS realistic to self-host rather than rent. For teams already weighing local versus hosted inference, the same trade-off we wrote up in on-premise versus cloud LLMs applies to generative media too.
03.
Where it fits in an asset pipeline
The strongest fit is the front of the pipeline, where speed beats polish.
🔸 One product photo or prompt to a GLB. Drop it into a three.js viewer or a model viewer on the web, convert to USDZ for AR, or stage it in a visualization scene. Fast concepting and previews are where TRELLIS shines.
🔸 Local editing. TRELLIS can regenerate or refine a region of the model, which is rare among single-image generators.
🔸 Honest about topology. The output is generative geometry, not artist-clean. Expect uneven triangles, no clean edge flow, and auto-generated UVs. For games or film you will retopologize, re-UV, and rebake textures. The newer materials-aware line narrows that gap without closing it.
So the value test for a given asset is simple: does this need to look right in a viewer next week, or does it need to be a riggable, optimized game asset? TRELLIS is excellent at the first and a starting point for the second.
04.
TRELLIS versus the alternatives
The open and hosted 3D field moved quickly. A quick orientation:
Model | Type | License | Note |
|---|---|---|---|
TRELLIS | Image and text to 3D | MIT | Truly permissive, multi-format output, local editing |
Hunyuan3D | Image to 3D, high-res texture | Community, with conditions | Strong open texture quality, license is not fully permissive |
Stable Fast 3D | Single image to 3D | Free under a revenue cap | Near-instant, lower fidelity |
Tripo | Image and text to 3D | Commercial SaaS | Cleaner, more game-ready meshes |
Rodin | Image and text to 3D | Commercial SaaS | Polished, production-leaning output |
Where TRELLIS stands: the strongest genuinely MIT option with multi-format output and local editing. Hosted tools like Tripo and Rodin beat it on out-of-the-box clean topology, and Hunyuan often wins on texture but with a more restrictive license. If your constraint is owning the model and the legal terms, TRELLIS is the obvious starting point.
05.
The honest limits, and how we use it
A few caveats we put on the table before anyone wires TRELLIS into production:
🔸 Topology and UVs are not drop-in for games or CAD. Budget for retopo.
🔸 CAD is out of scope. Generative meshes are not parametric solids, so they are not engineering geometry.
🔸 You own the input risk. The model is MIT, but the legal exposure of your input image or prompt is yours. Do not feed product shots you cannot license.
We treat TRELLIS as a concepting and previz engine: a way to get a believable 3D asset in front of a client in an afternoon, then decide what deserves a hand-finished pipeline. That is the same instinct behind where the current wave of AI video models is useful, and the same selection discipline we apply when choosing an image model. The tool is not the deliverable. The judgement about where it fits is.
