“Free” Can Mean Three Different Things
An AI 3D product may be free because it offers a limited hosted trial, because its code and model are openly licensed, or because a demo is temporarily subsidized. Those are not the same offer.
- Hosted trial: no GPU setup, but credits, queues, export limits, and terms can change.
- Open-source model: no per-job vendor fee, but hardware, electricity, storage, setup, and support still cost money.
- Browser tool: conversion or viewing may run locally, while AI generation still uses a paid provider.
Judge a tool by the cost of an accepted asset, not by the word “free.”
What Is Verified About TRELLIS.2
Microsoft's official TRELLIS.2 repository describes a 4-billion-parameter image-to-3D model based on an O-Voxel representation. The released model covers 512³ to 1536³ generation and full PBR materials. The repository and model are published under the MIT License, with separate terms applying to some dependencies.
Microsoft says the code has been verified on NVIDIA A100 and H100 GPUs. Its published timing table is an upstream H100 benchmark, not a promise for consumer GPUs, community wrappers, or this hosted service.
Source: microsoft/TRELLIS.2.
TRELLIS.2 is not the same release as Microsoft's earlier TRELLIS project. Model names such as TRELLIS-image-large and claims about text-to-3D, Gaussian, or radiance-field outputs belong to the older project unless the current TRELLIS.2 documentation explicitly says otherwise.
Hosted TRELLIS.2 on This Site
trellis-2.com is an independent hosted service, not a Microsoft product. It accepts one image, tracks the provider task, and returns a GLB result. An eligible account receives enough trial credit for one 512 job; later jobs use the balance shown in the generator.
The hosted service is convenient when you want to test the workflow without maintaining a GPU stack. It is not “unlimited”: 512, 1024, and 1536 currently cost 150, 300, and 500 credits respectively.
Other Tools Worth Evaluating
Products and open-source projects such as Hunyuan3D, Tripo/TripoSR, Meshy, Rodin, and research reconstruction models may fit different inputs and deployment constraints. Their model versions, free allowances, licenses, and exports change, so verify each claim on the provider's official page on the day of testing.
Avoid a universal ranking. Instead, ask:
- Does it accept the real input type and subject category?
- Can it export the format and material channels the destination tool needs?
- How many submitted jobs become accepted assets?
- What happens to credits when a job fails?
- What cleanup is required for hidden surfaces, topology, UVs, scale, or printability?
- Can the team meet the product's data-retention and commercial-use terms?
A Fair 20-Image Test
- Choose 20 real images, not vendor showcase samples.
- Include clear objects, thin structures, reflective surfaces, and partial occlusion.
- Record every submission, failure, retry, download, and credit charge.
- Review results without showing the reviewer which tool produced them.
- Import each candidate into the actual destination software.
- Time the cleanup needed to meet a written acceptance checklist.
Calculate:
acceptance rate = accepted assets / submitted jobs
cost per accepted asset =
(credits + infrastructure + operator and cleanup time) / accepted assetsA model with a lower generation fee can be more expensive when failures or cleanup are high.
Input and Output Checklist
Use one centered object with a clear silhouette and even lighting. A single image cannot reveal the true back, underside, internal structure, dimensions, or mechanical tolerance. Higher resolution may improve represented detail, but it cannot recover information that is absent from the source.
For every result, check:
- Hidden surfaces and thin parts
- Mesh holes, normals, and disconnected components
- Scale, orientation, and pivot
- Texture seams and PBR channels
- Polygon and texture budgets
- Source-image rights and target-platform terms
Frequently Asked Questions
Is open source the same as zero cost?
No. Open licensing can remove a per-generation vendor charge, but self-hosting still needs compatible hardware, setup, monitoring, storage, upgrades, and support.
Is TRELLIS.2 text-to-3D?
The current official TRELLIS.2 release is documented as image-to-3D. A text-first workflow can create a reference image separately and then submit that image, but that is a two-model pipeline with separate costs and rights considerations.
Can a generated file go directly into production?
Sometimes it can be useful as a draft, but it still needs destination-specific review. Games may need retopology and LODs; printing needs manifoldness, wall thickness, scale, and material checks; VFX needs topology, UV, naming, and pipeline validation.
What is the safest first test?
Use one clear image at 512, inspect every side of the GLB, and import it into the real destination tool before buying a larger pack or trying a higher resolution.