TRELLIS.2 vs Meshy, Hunyuan3D, Tripo and Rodin: How to Choose

Compare the Workflow, Not a Star Rating

AI 3D products change their models, prices, free allowances, and export options frequently. A static winner table becomes misleading quickly, and one attractive sample does not predict how a tool will handle your images.

This page compares TRELLIS.2, Meshy, Hunyuan3D, Tripo, and Rodin (Hyper3D) by the questions a buyer can verify, and it links the official source for every product claim below. It does not assign invented quality scores, and it repeats a competitor price only when that price is published on the vendor's own page and dated here.

What Each Product Actually Is

Mixing a self-hosted bill of materials with a hosted subscription price is the most common comparison error. Decide first which column you are buying.

TRELLIS.2: The Open-Weight Baseline

Microsoft's official repository describes TRELLIS.2 as a 4-billion-parameter image-to-3D model built on a field-free sparse voxel representation called O-Voxel, with published 512³, 1024³, and 1536³ modes and full PBR materials (base color, roughness, metallic, opacity). The repository is MIT-licensed, and Microsoft states the code has been verified on NVIDIA A100 and H100 GPUs, needs an NVIDIA GPU with at least 24 GB of memory, and is currently tested only on Linux.

Microsoft also publishes a model-runtime table (about 3 s at 512³, 17 s at 1024³, 60 s at 1536³ on an H100). That is an upstream measurement on stated hardware, not a promise for a shared queue, a consumer GPU, or an end-to-end upload-export-download workflow. Do not copy it into a hosted latency claim. The earlier Microsoft TRELLIS release is a different model: do not carry its model names, installation flags, or output modes into a TRELLIS.2 evaluation.

trellis-2.com is an independent hosted workflow around that model. It accepts one image and returns a GLB download; the selected resolution costs 150, 300, or 500 credits. No new first-task credits are granted; existing balances keep their original validity. The live generator and checkout are the source of truth for task cost and purchase price.

Meshy: Vendor-Managed Workspace, API and Plugins

Meshy is a commercial hosted platform. Its official documentation lists text-to-3D, image-to-3D, and a conversational "3D Agent", and states that image-to-3D runs on Meshy 7, its current foundation model. The same page documents six workspace modules (Image, 3D Model, 3D Printing, Animate, Scene, Video), REST endpoints for generation, texturing, remeshing and animation, an MCP server for agents, and plugins that run inside Blender, Unity, and Maya. It also lists export formats — GLB, FBX, OBJ, STL, USDZ, and 3MF — and notes which of them keep textures, and that its free plan includes 100 credits per month.

Meshy's marketing material additionally claims print-oriented output such as watertight, manifold meshes, an automatic printability check, and multi-color splitting, plus a PBR set (albedo, normal, metallic, roughness), HD texture, low-poly mode, and auto-rigging for humanoid and quadruped characters. Those are vendor claims published on meshy.ai; treat them as claims to test on your own inputs, not as independent measurements. Meshy's own pricing and terms pages are the only current source for plan amounts.

Hunyuan3D 2.1: Open Weights With a Territorial Licence

Tencent's Hunyuan3D 2.1 repository publishes two model families — a 3.3B image-to-shape model and a 2B texture model — with full model weights and training code, PBR texture synthesis, and stated memory needs of about 10 GB VRAM for shape, 21 GB for texture, and 29 GB for both. It documents macOS, Windows, and Linux support. Tencent also runs a hosted workspace at 3d.hunyuan.tencent.com, so the same brand covers both columns of the table above.

The licence is the part buyers miss. Hunyuan3D 2.1 ships under the Tencent Hunyuan 3D 2.1 Community License, which excludes the European Union, the United Kingdom, and South Korea from its territory, requires a separate licence from Tencent once a product exceeds 1 million monthly active users, and obliges anyone offering it as a service to disclose the real provider and state that Tencent is not affiliated with or endorsing that service. Read the licence file before you plan a commercial deployment in those markets.

Tripo and Rodin: API-First Services

Both are commercial platforms whose public documentation is aimed at developers rather than at consumers.

Tripo's official developer site describes one API that covers image-to-3D, text-to-3D, and multi-view to 3D, plus texture generation, retopology, segmentation, auto-rigging, and animation retargeting, with output in GLB, FBX, OBJ, and USDZ, per-task-type concurrency quotas, and webhook callbacks instead of polling. That site publishes usage-based pricing (1 credit = 0.01 USD, with image-to-3D listed from 20–30 credits and auto-rigging at 25 credits when this page was checked on 2026-09-28) — check the live page before you budget.

Rodin's documentation describes an HTTP API that creates 3D assets from images or text, with Bearer authentication, a base URL of https://api.hyper3d.com/api/v2, an asynchronous task lifecycle with status polling, published data-retention policy, and a credit price list on its own pricing page. Rodin's product page markets Gen-2.5 together with bounding-box, voxel, and point-cloud controls and 3D editing.

Neither vendor's documentation publishes a measured success rate on a defined image cohort, so a comparison of the two can only be settled with your own test set.

What to Verify for Every Product

For the open-weight row, add hardware and licence: VRAM, operating system, and whether the licence covers your territory, company size, and commercial use. The open-source image-to-3D review walks through those constraints for TRELLIS.2, Hunyuan3D, TripoSR, and SF3D.

Hosted Service vs Self-Hosting

Hosted products can remove GPU setup and maintenance. Compare them using a paid test set and record:

• Successful outputs divided by submitted jobs
• Median time to a downloadable asset
• Credits lost on failed or unusable jobs
• Export and material behavior in Blender or your target engine
• Cancellation, refund, data-retention, and commercial-use terms

Self-hosting the open-weight option is not cost-free either: hardware, electricity, model storage, setup time, monitoring, upgrades, and operator support all remain. Hunyuan3D's published VRAM figures and TRELLIS.2's 24 GB minimum are the entry ticket, not the total cost.

A Reproducible Comparison Protocol

1. Select 20–50 images from the real workload, including difficult and ordinary cases.
2. Use the closest equivalent settings and record the exact date and product version.
3. Keep every failure; do not compare only hand-picked successful results.
4. Blind-review shape, hidden surfaces, materials, and destination-tool compatibility.
5. Measure accepted outputs, total spend, operator time, and cleanup time.
6. Recheck current pricing, licence, and terms on each provider's official site before buying.

Use this calculation:

cost per accepted asset =
  (generation spend + variable infrastructure + operator time) / accepted assets

The cheapest generation is not necessarily the lowest-cost asset if cleanup or failure rates are high.

When TRELLIS.2 Is Worth Testing

TRELLIS.2 is a reasonable candidate when the workflow starts from a single clear object image, needs a textured GLB draft, or requires an open-weight path for internal evaluation. It is not a substitute for measurements, CAD, production retopology, rigging, or safety review.

Start with one 512 job, inspect the GLB in the destination tool, and move to a higher resolution only when the first result justifies the added credits.

Run a 512 image-to-3D test

Review current TRELLIS.2 pricing