Skip to content

AI 3D model generators compared (2026)

A small studio can use AI 3D generators today for props, background objects and prototypes, and should not expect finished characters from them. Among hosted services, Meshy and Tripo cover the most of the pipeline for about 20 a month in euros or dollars, with commercial rights on the paid plans. Among models you can run yourself, Microsoft's TRELLIS.2 has the cleanest licence (MIT), while Tencent's Hunyuan3D weights carry territory limits that matter if you publish worldwide.

By Futigo · Updated · 8 min read

Key points

  • Meshy and Tripo both offer text and image input, remeshing, PBR textures, auto-rigging and commercial use from their first paid plan.
  • Free tiers are for trying, not shipping: Meshy's free output is CC BY 4.0 with credit required, Tripo's is public and non-commercial, Sloyd's is personal use only.
  • TRELLIS.2 is MIT licensed with no revenue or territory limits, but its repository asks for an NVIDIA GPU with at least 24 GB of memory.
  • The Hunyuan3D community licence excludes the European Union, the United Kingdom and South Korea, and that exclusion covers the outputs too.
  • Generated meshes work best as static objects. Blender's own manual says topology for a mesh that deforms still has to be made by hand.

Hosted services side by side

Hosted services run in the browser, charge in credits and need no GPU. These six were live, with public pricing or an official announcement we could read, on 15 November 2026.

Hosted AI 3D model generators compared: input, export formats, extras, free tier and entry paid plan, November 2026
ServiceInputExport formatsExtrasFree tierEntry paid plan
MeshyText, image; multi-view image on paid plansFBX, OBJ, GLB, USDZ, STL, BLEND; 3MF through the APIRemesh to quads or triangles, PBR maps, rigging, animation library100 credits a month, up to 10 assets, CC BY 4.0 licencePro: €19 a month, 1,000 credits
TripoText, image, multi-viewGLB, FBX, OBJ, USDZ and othersRetopology, quad mesh, smart low-poly, PBR, auto rig, animation retargeting200 credits a month (about 13 models), public models, non-commercialPro: $20 a month, or $13 a month billed yearly, 3,000 credits
Rodin (Hyper3D)Text, image (one to five images through the API)STL, FBX, OBJ, GLB, GLTF, USDZPBR materials, UVs, smart low-poly; high-poly quads on BusinessPreview for free, pay per confirmed resultCreator: $30 a month, or $24 a month billed yearly, about 60 models
Tencent HY 3D GlobalText, image (up to four views), sketchOBJ, GLBRetopology with triangle or quad faces, texture painting20 generations a day when it launched in November 2025See pricing page (sign-in required)
SloydText, imageGLB, OBJ, FBXRigging, animation, retexturingOne AI generation a day, personal use onlyPlus: $15 a month, or $12.50 a month billed yearly, unlimited generation
Stability AI APIImageGLBUV-unwrapped textured meshNot confirmedPay per use: $0.10 per Stable Fast 3D model, $0.04 per SPAR3D model

Plans and prices read from each service's pricing page on 15 November 2026, as of November 2026. Meshy's page showed prices in euros when we loaded it. The Tencent free allowance is from Tencent's launch announcement of 26 November 2025; current pricing sits behind a sign-in and was not confirmed.

Credits do not compare directly. On Meshy's API price list, an image-to-3D mesh with textures costs 30 credits on its current model, so 1,000 credits is roughly 33 textured models before remeshing or rigging. Tripo estimates its 3,000 Pro credits at about 200 models, but its game-oriented Smart Mesh P2.0 option costs 100 credits a model, which cuts that to 30.

Open-weight models you can run yourself

Running a model locally means no credits and no upload of your concept art. The cost is a large NVIDIA GPU and a licence you need to read. Most of these take an image and return a mesh, usually as a GLB file.

Open-weight image-to-3D models compared: developer, licence, commercial use, dates, stated hardware and GitHub stars, November 2026
ModelDeveloperLicence and commercial useRepository created / last pushStated hardwareGitHub stars
TRELLIS.2MicrosoftMIT. Commercial use allowedNovember 2025 / July 2026NVIDIA GPU with at least 24 GB; tested on Linux only12,930
TRELLISMicrosoftMIT. Commercial use allowedDecember 2024 / October 2026NVIDIA GPU with at least 16 GB14,275
Hunyuan3D-2.1TencentTencent Hunyuan 3D 2.1 Community License. Commercial use allowed under 1 million monthly active users, not in the EU, UK or South KoreaJune 2025 / October 202510 GB VRAM for shape, 29 GB for shape and texture4,446
Hunyuan3D-2TencentTencent Hunyuan 3D 2.0 Community License. Same limits as 2.1January 2025 / October 20256 GB VRAM for shape, 16 GB for shape and texture15,419
SPAR3DStability AIStability AI Community License. Free commercial use under US$1 million annual revenueDecember 2024 / May 202510.5 GB VRAM by default, about 7 GB in low-VRAM mode1,091
Stable Fast 3DStability AIStability AI Community License. Free commercial use under US$1 million annual revenueJuly 2024 / January 2025About 6 GB VRAM for one image1,934
TripoSGVAST (Tripo)MIT. Commercial use allowedMarch 2025 / April 2025CUDA GPU with at least 8 GB VRAM1,909
TripoSRVAST (Tripo) and Stability AIMIT. Commercial use allowedFebruary 2024 / June 2026About 6 GB VRAM for one image7,392
InstantMeshTencent ARCApache 2.0. Commercial use allowedApril 2024 / January 2025Not stated4,601
Step1X-3DStepFunApache 2.0. Commercial use allowedMay 2025 / September 202527 to 29 GB GPU memory for geometry and texture902
SAM 3D ObjectsMetaSAM License, a custom Meta licence. No non-commercial clause, but it bars military and weapons usesSeptember 2025 / June 2026NVIDIA GPU with at least 32 GB VRAM, Linux7,677

Licences read from the licence file in each GitHub repository and the licence tag on Hugging Face; hardware from each README or setup guide; stars and dates from the GitHub API. All checked on 15 November 2026. Some repositories depend on third-party code under other licences, which their READMEs list.

TRELLIS.2 is the one to try first if you have the hardware. Its README describes a 4-billion-parameter image-to-3D model that outputs a GLB with base colour, roughness, metallic and opacity maps. The first TRELLIS also has text-to-3D checkpoints. The newest Hunyuan3D weights we found on Tencent's GitHub are 2.1; Tencent's own announcement names Hunyuan 3D 3.0 for the hosted service.

Licensing and ownership

Three kinds of terms decide whether you can ship a generated model in a paid game.

  • Free-tier terms. Meshy's terms say Meshy owns free-plan output and licenses it to you under CC BY 4.0 if you credit Meshy; paid customers own theirs. Tripo's free plan is marked public and non-commercial. Sloyd's free plan is personal use, and its $15 Plus plan allows commercial use but not reselling on marketplaces.
  • Open-weight licence limits. The Hunyuan3D-2.1 licence does not apply in the European Union, the United Kingdom or South Korea, and it forbids using or distributing the outputs outside its territory. A game sold worldwide on Google Play reaches those places. The Stability AI Community License is free until your organisation passes US$1 million in annual revenue, then needs an enterprise licence.
  • Copyright in the output. The US Copyright Office concluded in January 2025 that AI output can be protected only where a human author determined enough of the expressive elements, and that prompts alone do not qualify. A model you generate and ship unchanged may not be yours to protect, even where the tool's terms say you own it. Other countries differ.

Some services reserve the right to use what you upload to improve their models. Read that clause in any service's terms before uploading unreleased concept art.

What game-ready means, and where generated meshes fall short

A game-ready model has a polygon count the target device can afford, clean topology, sensible UVs, textures in a PBR set, a correct scale and pivot and, if it moves, a skeleton with skin weights. Generators now address most of these as options. They do not do so by default.

  • Polygon count. Raw output is dense. Meshy's remesh step accepts a target from 100 to 300,000 faces with a default of 30,000, according to its API docs. Tripo's P2.0 model caps output at 50,000 triangles or 25,000 quads. For a mobile prop you will usually want far fewer than either default.
  • Topology. Quad output exists in Meshy, Tripo, Rodin and Tencent's service. Quads are not the same as edge loops placed where a joint bends.
  • UVs and textures. Most tools unwrap automatically and can produce metallic, roughness and normal maps. Automatic UV islands are hard to repaint by hand, and lighting baked into the base colour is a common complaint with image-to-3D in general.
  • Hidden sides. From one image, the model has to guess the back. Multi-view input, which Meshy, Tripo, Rodin and Tencent all support, reduces this.
  • Consistency. Ten props generated from ten prompts rarely share one art style. Generating the reference images in one style first, then using image-to-3D, helps.

The usual clean-up path in Blender

Export GLB where you can. Godot's documentation recommends glTF 2.0 over FBX and OBJ. Then, in Blender:

  • Fix scale, orientation and origin, and delete loose or interior geometry.
  • Reduce faces with the Decimate modifier, which the manual describes as cutting vertex and face count with minimal shape change.
  • For badly tangled meshes, use the Voxel or Quad remesher. The Quad mode uses the QuadriFlow algorithm.
  • Unwrap new UVs if you changed the mesh, then bake the generated textures from the original onto the clean copy.
  • Check the material in your engine's lighting, not only in the generator's viewer.

The same manual page is blunt about the limit: automatic remeshing should not be used to create topology for a mesh that will be deformed, such as an animated character. That work is still manual retopology.

AI rigging and animation

Auto-rigging has become a standard extra. Meshy's rigging API costs 5 credits, returns FBX and GLB with walking and running clips, and its docs say it works well only on standard two-legged humanoids with clear limbs, with an input limit of 380,000 triangles. Its paid plans list an animation library of over 500 clips. Tripo charges 20 credits for Auto Rig in its studio and offers animation retargeting through its API. Sloyd includes rigging and animation in its paid plans.

On the open side, UniRig from Tsinghua University and Tripo predicts a skeleton and skin weights for humans, animals and objects. It is MIT licensed, has 2,130 GitHub stars and asks for a GPU with at least 8 GB of VRAM. Its README now points to a successor, SkinTokens, also MIT.

A rig is only as good as the mesh under it. An auto-rigged generated character can walk across a background. Close to the camera, shoulders, elbows and faces tend to show the lack of deliberate topology.

What works today, and what we do at Futigo

  • Works: static props, set dressing, background objects, grey-box prototypes, concept models to hand to an artist, and items for 3D printing.
  • Works with clean-up: mid-ground objects, simple vehicles, and stylised low-poly sets where you control the reference images.
  • Does not work yet: hero characters, faces, hands, anything that deforms close to the camera, modular kits that must snap together exactly, and anything needing strict consistency across a large set.

Our games are mostly 2D, built in Godot 4.7 with much of the art drawn in code, so this comparison is desk research from pricing pages, docs and licence files, not a hands-on benchmark. If you are choosing an engine to bring models into, see our open-source game engine comparison.

Questions

What is the best AI 3D model generator in 2026?
For most small teams, Meshy or Tripo. Both take text or images, export GLB and FBX, include remeshing, PBR textures and auto-rigging, and grant commercial use from their first paid plan: €19 a month for Meshy Pro and $20 a month for Tripo Pro, as of November 2026.
Is there a free AI 3D model generator I can use commercially?
Meshy's free plan gives 100 credits a month under CC BY 4.0, which allows commercial use if you credit Meshy. For no strings at all, run an MIT-licensed model such as TRELLIS.2 or TripoSG yourself, which needs an NVIDIA GPU.
Can I use Hunyuan3D commercially?
Yes, with limits. The Tencent Hunyuan 3D community licence allows commercial use unless your products had more than 1 million monthly active users, but it does not apply in the European Union, the United Kingdom or South Korea, and it bars using or distributing outputs there.
Are AI-generated 3D models game-ready?
Static props often are after a pass in Blender to cut polygons, fix scale and check textures. Characters and anything that deforms are not: Blender's manual says topology for an animated mesh still has to be made by hand.
What GPU do I need to run an image-to-3D model locally?
It depends on the model. TripoSR and Stable Fast 3D state about 6 GB of VRAM, TripoSG 8 GB, TRELLIS 16 GB, TRELLIS.2 24 GB, and Hunyuan3D-2.1 needs 29 GB to generate both shape and texture.
Who owns the copyright in an AI-generated 3D model?
The tool's terms decide what the vendor grants you, and paid plans from Meshy and Tripo include commercial rights. Whether copyright exists at all is separate: the US Copyright Office said in January 2025 that output is protected only where a human determined enough of its expression, and prompts alone are not enough.