Image-to-3D AI will not hand you a finished game asset, but it can get you a textured first draft in minutes instead of days. Knowing how it works, and where it breaks, is what separates a useful draft from a wasted afternoon.
The expensive part of 3D has never been the software licence. It is the skilled hours: modelling, topology, UV unwrapping and texturing a prop, a product mockup or an AR preview object in Blender or Maya. Browser-based generators attack that first stage. You upload a photo or type a description, a model runs on a cloud GPU, and you download a textured mesh. No local GPU, no CUDA, no install.
This article looks at Trellis 3D, one such service, and separates what is verified from what is marketing. It also points you to the free official demos from Microsoft, which many people never realise exist.
What Trellis 3D actually is
Trellis 3D (trellis-3d.net) is a third-party paid web service, not a Microsoft product. It borrows the name of Microsoft’s research model, but Microsoft’s TRELLIS project page states that it is the only official webpage for the project. Keep the two apart:
Microsoft TRELLIS / TRELLIS.2 | Trellis 3D (trellis-3d.net) | |
|---|---|---|
What it is | Open research models from Microsoft Research | Commercial hosted service, credit-based |
Licence | MIT (models and most code) | Proprietary service; see its Terms of Service |
Models | TRELLIS (2024), TRELLIS.2 (4B, 2025) | TRELLIS.2 plus Hunyuan 3D, Seed 3D, Tripo and Meshy |
Cost | Free | Monthly plans from $6.99 (billed yearly) |
The service offers several generation models side by side. Its menu lists Trellis 2 and a cheaper “Trellis 2 Lite” option, alongside Hunyuan 3D 3.1, Tripo v3.1 and Meshy 7. “Lite” is the service’s own label; Microsoft has not published a model by that name.
Try Microsoft’s models for free first
If you only want to see what TRELLIS can do, you do not need any paid service. Microsoft hosts official browser demos on Hugging Face, with no install and no GPU on your side:
TRELLIS.2 official demo — the current 4B image-to-3D model, with PBR materials.
TRELLIS official demo — the original 2024 model, which also exports Gaussians.
TRELLIS in Azure AI Foundry Labs — Microsoft’s own hosted entry point.
Public Spaces share GPU time, so expect queues at busy hours. What a paid service adds on top is convenience: a single interface, several competing models, format conversion (FBX, USDZ, GIF), history and no queue.
Image to 3D: input quality decides the result
The model can only reconstruct what the photo shows, so a clean input matters more than any setting. Trellis 3D accepts JPG, PNG or WebP files up to 10 MB.
What helps:
One object, centred. Multiple overlapping objects confuse the boundary between subject and scene.
A plain background. White or neutral grey stops background shapes from fusing into the mesh. The service has a background-remover if your photo is busy.
Even lighting. Strong directional shadows get baked into the texture.
Enough resolution. The generator page asks for at least 1024 × 1024 px; the homepage FAQ says 800 × 800. Aim for 1024 or more.
A front or three-quarter view. A three-quarter angle shows the front and one side, which gives the model more depth cues. This is sensible practice, not a measured guarantee.
What hurts: cluttered scenes, motion blur, heavy compression, haze and other atmospheric effects.
What the model sees versus what it guesses
The back, the underside and any hidden surface are not observed; they are inferred from what the model learned about similar objects. The hidden side can therefore differ from the real thing. That is a limit of single-image reconstruction, not a bug.
Treat every output as a starting asset. The service itself pitches it for non-hero props, decorations and environment filler, and advises inspecting and cleaning anything meant for production, printing or hero use.
Text to 3D: useful, but image-first usually wins
Trellis 3D also accepts a written description instead of a photo. Good prompts name four things: shape, material, surface detail and style. “A matte black aluminium flashlight with a knurled grip” will beat “a flashlight” every time.
It works best for concept prototypes, game-prop ideation and stylised objects. It struggles with human faces and realistic full-body characters: facial detail gets lost and proportions drift. Stylised, anime and low-poly characters come out more consistently.
One important caveat comes from Microsoft. TRELLIS.2 is an image-to-3D model only. For the original TRELLIS, the README recommends generating an image with a text-to-image model first, then feeding that image to the image model. Its text-conditioned models are less creative and detailed because of data limits.
In practice: if a text prompt underdelivers, generate a clean reference image first, then run image-to-3D on it.
Picking a model: what is verified and what is marketing
The service recommends Trellis 2 for complex topology, Hunyuan 3D for hard surfaces and Seed 3D for organic shapes. Only the first claim is backed by the model’s own developers. The other two specialisations come from the service’s product copy, not from published benchmarks. Use them as a starting hypothesis, and test your own object on two models before committing credits in bulk.
Model | Developer | What the developer actually claims | Licence of the open release |
|---|---|---|---|
Microsoft Research | 4B-parameter image-to-3D; O-Voxel handles open surfaces, non-manifold geometry and enclosed interiors; full PBR materials | MIT | |
Tencent | Image to shape (3.3B) plus PBR texture model (2B); “production-ready” PBR | Tencent Hunyuan 3D 2.1 Community License — excludes the EU, UK and South Korea | |
ByteDance Seed | Single image to simulation-ready 3D assets for physics engines and robot training | Not open weights; offered via API |
Microsoft TRELLIS.2: complex and irregular geometry
This is the one model whose claimed strength is documented by its makers. Microsoft describes TRELLIS.2 as a 4B-parameter image-to-3D model built on O-Voxel, a “field-free” sparse voxel structure. It is designed for open surfaces such as leaves and cloth, non-manifold geometry and internal enclosed structures. It models base colour, roughness, metallic and opacity. Lattice work, foliage and objects with holes are its natural territory.
Tencent Hunyuan 3D: strong textures, read the licence carefully
Trellis 3D labels Hunyuan 3D as “Apache 2.0”. That is not correct for the open release. Hunyuan3D 2.1 ships under the Tencent Hunyuan 3D 2.1 Community License. It does not apply in the European Union, the United Kingdom or South Korea, and it forbids using the model or its outputs outside its licensed territory. Services with more than 1 million monthly active users need a separate licence from Tencent.
Two further points:
The service lists Hunyuan 3D 3.1, a newer version than the open 2.1 release, likely accessed through Tencent’s commercial API under different terms. The 2.1 licence may not be the one that applies, but neither is Apache 2.0.
If you are in the EU or UK, do not assume you can use Hunyuan outputs commercially. Check the service’s Terms of Service and ask the operator which Tencent terms cover you.
On quality, Tencent’s own benchmarks rank Hunyuan3D 2.1 ahead of the original TRELLIS on shape and texture metrics. The “hard-surface specialist” label, however, is the service’s framing.
ByteDance Seed 3D: built for simulation, not proven for creatures
ByteDance presents Seed3D 1.0 as a model for simulation-ready assets: geometry, aligned textures and PBR materials that drop into physics engines such as NVIDIA Isaac Sim for robotics training. That explains the URDF export the service lists for it, since URDF is a robot-description format.
The service’s description of Seed 3D as a fast, multi-view model for organic creatures is not how ByteDance describes it. Seed3D takes a single image; “multi-view” refers to images it generates internally to build consistent textures. Treat “best for creatures” as unverified.
Tripo and Meshy
The generator also lists Tripo v3.1 and Meshy 7, two commercial models from other companies. Both are closed models, so their terms come only through the service.
Export formats and polygon counts
GLB is the right default: one file holding geometry and materials, read natively by Three.js, Unity, Unreal, Blender and WebXR. Pick another format only when the destination demands it.
Format | Use it for | Note |
|---|---|---|
GLB | Web, game engines, general use | Binary glTF, the Khronos runtime standard; bundles meshes, materials and textures |
OBJ | Older tools, 3D-print prep | Universally readable; materials sit in a separate file |
FBX | Maya, 3ds Max, animation pipelines | Interchange format for rigging and animation |
USDZ | iPhone/iPad AR Quick Look, visionOS | Lets a shopper place a product in their room without an app |
GIF | Sharing with non-3D people | A rotating 360° turntable preview |
URDF | Robotics simulators | Listed for Seed 3D only |
Microsoft’s own tools export GLB (both models) and Gaussian PLY (original TRELLIS). FBX, USDZ and GIF are conversions added by the service.
Match polygon count to the destination
The service’s FAQ gives these ranges as guidance, not guarantees:
40K–200K polygons: real-time game assets, WebXR, mobile AR.
300K–700K: meshes headed for retopology and cleanup.
1M–1.5M: high-resolution renders, archival files, detailed product visualisation.
Choose by where the asset will live, not by how impressive the number looks. A 1.5M-polygon mesh in a mobile AR scene is a performance problem, not a feature.
Real workflows: where the meshes earn their keep
The pattern in every workflow is the same: AI produces the first draft, a person does the part that needs judgement.
Game prototyping. Upload concept art or a prop reference, generate a GLB, retopologise and clean it in Blender, then export to Unity or Unreal. Best for background props; hero assets still need an artist.
Product visualisation. A clean product photo becomes a GLB for Three.js, Google’s model-viewer or Shopify’s 3D viewer, plus a USDZ for iOS AR Quick Look.
3D printing. Export OBJ or GLB, open it in Blender or Meshmixer, check that it is watertight, repair holes and non-manifold faces, then slice.
Printing is where skipped inspection bites hardest. Simple solid props and decorative objects usually behave. Thin walls, deep undercuts and complex hidden geometry usually do not. TRELLIS.2’s ability to produce open surfaces and non-manifold shapes is a strength for rendering but a liability for printing, so expect more repair work on those meshes.
Pricing and generation time
Trellis 3D runs on monthly credits shared across all its tools, including image, video and voice generation. Its pricing page lists three plans; prices below are the discounted monthly rate when billed yearly.
Plan | Price (billed yearly) | Credits per month | Yearly total |
|---|---|---|---|
Basic | $6.99/month (list $9.99) | 200 | $83.88 |
Pro | $18.99/month (list $29) | 880 | $227.88 |
Enterprise | $35/month (list $49) | 1,800 | $420 |
The cheapest 3D option, Trellis 2 Lite, costs 5 credits and takes 2–6 minutes. Credit costs for the other 3D models are not published on the pricing page, so check the generator before you run large batches.
Three terms worth knowing before you pay: unused credits do not roll over, all sales are final except where law requires refunds, and all paid plans include commercial usage rights under the service’s terms.
For scale: Microsoft reports TRELLIS.2 takes about 3 seconds at 512³, 17 seconds at 1024³ and 60 seconds at 1536³ on an NVIDIA H100. The service’s minutes include queueing, conversion and its own hardware choices.
The Microsoft research behind it
Both Microsoft models are open, MIT-licensed and free to run yourself, if you have a large NVIDIA GPU. The key facts:
TRELLIS (Dec 2024, CVPR 2025) | TRELLIS.2 (Dec 2025) | |
|---|---|---|
Core idea | SLAT, a structured latent that decodes to meshes, radiance fields or 3D Gaussians | O-Voxel, a “field-free” sparse voxel structure with full PBR attributes |
Inputs | Image or text | Image only |
Model sizes | Image-large 1.2B; text 342M, 1.1B and 2.0B | 4B |
Training data | TRELLIS-500K, about 500,000 curated 3D objects | Objaverse-XL and other sets |
Output resolution | Fixed pipeline | 512³ to 1536³ |
Local hardware | Linux, NVIDIA GPU with at least 16 GB | Linux, NVIDIA GPU with at least 24 GB |
Free demo |
Note the TRELLIS.2 requirement. Trellis 3D’s FAQ says 12–24 GB of VRAM “depending on output resolution”, but Microsoft’s repository states 24 GB as the minimum, verified on A100 and H100 cards. A typical 12 or 16 GB consumer card will not run it with the official setup.
Caveats from the research side
Microsoft’s TRELLIS project page says the model excels at artistic-style assets and is limited at photorealistic real-world objects. It also says the materials on that page are for academic and research purposes only. That disclaimer covers the page’s demo content; the model weights and code themselves are MIT licensed, which permits commercial use.
For commercial work through any hosted service, three documents matter: that service’s Terms of Service, the licence of the specific model you pick, and your own rights to the reference images you upload.
People also ask
Is Trellis 3D made by Microsoft?
No. Trellis 3D is a third-party service that offers Microsoft’s TRELLIS.2 among several models. Microsoft’s official ways to try TRELLIS are its Hugging Face demos and Azure AI Foundry Labs.
Can I turn one photo into a 3D model for free?
Yes. Microsoft’s official TRELLIS.2 demo runs in the browser at no cost and exports GLB. Expect queues at busy times. Paid services add more models, more formats and no queue.
Which model handles complex or non-manifold geometry?
TRELLIS.2. Its O-Voxel representation is designed for open surfaces, holes, non-manifold shapes and enclosed interiors, and Microsoft documents this directly.
Which model is best for hard-surface products?
There is no independent test that settles it. Trellis 3D recommends Hunyuan 3D, and Tencent’s benchmarks rate its textures highly. In the EU and UK, check the licence before any commercial use, since the open Hunyuan3D 2.1 licence excludes those regions.
Can I run TRELLIS.2 on my own PC?
Only with a large NVIDIA GPU. Microsoft requires at least 24 GB of GPU memory and tests on Linux. The original TRELLIS needs at least 16 GB.
How long does generation take, and what does it cost?
On Trellis 3D, most jobs finish in a few minutes; Trellis 2 Lite takes 2–6 minutes for 5 credits. Plans start at $6.99 per month billed yearly for 200 credits, and unused credits expire each month.
The bottom line
Image-to-3D tools do not replace a 3D artist; they remove the slow first step of getting from a photo to a textured mesh you can open and fix. How much time that saves depends on how much cleanup the result still needs, which varies by object and destination.
Start with Microsoft’s free TRELLIS.2 demo to see whether the output fits your work. Move to a paid service like Trellis 3D if you need several models in one place, FBX or USDZ export, or no queue. Feed it a clean image, choose a polygon count for the destination, and inspect every mesh before it ships.
Above all, check the licence of the model you actually use. TRELLIS and TRELLIS.2 are MIT. Hunyuan’s open licence excludes the EU, UK and South Korea, whatever a reseller’s label says.
Additional resources
This article includes content created with AI. Facts checked against the linked sources on 14 October 2026.