# 9 Microsoft TRELLIS alternatives for 3D studios

> Microsoft's releases give you a model and a GPU bill; a hosted workspace gives you the handoff. Hydrilla Studio chains an OpenAI or Gemini reference image into BlueFox 1, then lets you check wireframe, face counts, and parts before export or a Bridge handoff. Stay with Microsoft TRELLIS if self-hosting or Gaussian output is the requirement. Hydrilla publishes this guide and sells Hydrilla Studio.

Checked 20 Nov 2026 · Hydrilla Team · https://hydrilla.ai/compare/trellis-alternatives

## TL;DR

- **If you want the handoff in one workspace:** choose OpenAI or Gemini to make or edit the reference image in Hydrilla Studio, run BlueFox 1 on it, inspect wireframe, face counts, and selectable parts, then export or send the mesh through the local Bridge. Studio doesn't run Meshy, Tripo, Rodin, TRELLIS, or any other vendor's mesh model.
- **Keep the original TRELLIS** for its Gaussian and radiance-field outputs, or to keep an existing pipeline pinned to the same model.
- **For local PBR:** start with TRELLIS.2-4B for image-to-mesh or for texturing a mesh you already have. Compare Hunyuan3D 2.1 if you want separate shape and paint stages and Tencent's license covers your use.
- **For local shape edits:** try Stability SPAR3D if you want to move or remove points before the model turns them into a mesh.
- **For hosted work:** try Meshy if you need its remesh or rigging steps; Tripo for references from several angles; Rodin when the material package is part of the brief; or Hydrilla when you want the reference image, mesh generation, and inspection in one chain.
- **For repeatable families or outsourced production:** try Sloyd for template variants and Kaedim when you need an artist-reviewed delivery against a written specification.

## Where does Microsoft TRELLIS fall short?

The baseline here is Microsoft's **original TRELLIS**, not TRELLIS.2. It offers TRELLIS-image-large, and Microsoft added separate TRELLIS-text checkpoints on March 25, 2025. Even for text-led work, [Microsoft recommends generating an image first](https://github.com/microsoft/TRELLIS). The multi-image conditioning added December 18, 2024 didn't come with a separately trained multi-view model. Test what happens when your front and rear references disagree.

The original can produce meshes, 3D Gaussians, and radiance fields. Its [standard textured GLB exporter](https://github.com/microsoft/TRELLIS/blob/main/trellis/utils/postprocessing_utils.py) bakes a base-color texture and uses a fixed roughness factor. If your material artist needs separate roughness and metallic information to relight a prop, start with [TRELLIS.2](https://github.com/microsoft/TRELLIS.2) instead of assuming the original GLB includes it.

Microsoft reports Linux-only testing and specifies at least 16 GB of NVIDIA GPU memory for TRELLIS and 24 GB for TRELLIS.2. Owning the runtime lets you decide where references are processed; it also makes you responsible for CUDA dependencies, job handling, and export checks. The repositories offer Python code and demos. Any managed endpoint you buy from a third party has that provider's pricing, features, and data terms.

## How did we choose these Microsoft TRELLIS alternatives?

We included tools that can replace a named step: run inference on your own hardware, generate a draft in a hosted workspace, make consistent variants, or commission an asset to spec.

| Criterion | Why it matters | What good looks like |
| --- | --- | --- |
| Runtime and data | An artist's workstation and a machine that can run inference are not always the same thing. | A supported GPU and operating system for local runs, or hosted terms that cover where references go and how long they're retained. |
| Inputs | A prompt, one photograph, and a turnaround sheet describe different amounts of the object. | Accepts the references you actually have. If there is a rear view, check that the result reflects it. |
| Geometry and materials | A textured preview can look right while hiding topology or material problems. | A mesh within the face budget, usable UVs, and the PBR channels the destination shader needs. |
| Handoff | The file has to survive export and import, including scale, material slots, and any parts you need to edit. | The intended format opens correctly in your DCC or engine, by file or supported plugin. |
| Rights | A license to run a model and permission to use or distribute the result are different questions. | The model license, dependency terms, or hosted plan cover the project, geography, and intended distribution. |
| Cost per accepted asset | A download or a base generation charge doesn't capture the cost of the version the team approves. | Count compute or credits, retries, export or processing charges, and artist cleanup on the same brief. |

We checked first-party repositories, docs, and pricing on November 20, 2026. This is a documentation comparison, with no hands-on benchmark or quality ranking. The table starts with original TRELLIS, then groups local releases and hosted workflows. Prices are USD; web subscriptions and API credits are separate when noted. Compare candidates using the same single-view brief and, where supported, the same turnaround sheet. Inspect the result in your destination scene under the same lighting.

## Microsoft TRELLIS alternatives at a glance

| Tool | Best for | Inputs | Output | Export formats | Engine and DCC plugins | API | Paid from |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Microsoft TRELLIS (baseline) [1](https://github.com/microsoft/TRELLIS) [2](https://github.com/microsoft/TRELLIS/blob/main/trellis/utils/postprocessing_utils.py) | Self-hosted meshes, Gaussians, or radiance fields | One or more images; separate text checkpoints | Mesh with baked color, Gaussians, or radiance field | GLB; Gaussian PLY | File import | Local Python | Download; own GPU |
| Microsoft TRELLIS.2 [3](https://github.com/microsoft/TRELLIS.2) [4](https://github.com/microsoft/TRELLIS.2/blob/main/example_texturing.py) | Self-hosted image-to-PBR mesh | Single image; mesh + image for texturing | Mesh with PBR material | GLB | File import | Local Python | Download; own GPU |
| Hunyuan3D 2.1 [5](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1) [6](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/LICENSE) [7](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/gradio_app.py) | Separate local shape and material passes | Image; mesh + image for paint | Mesh; optional PBR paint stage | GLB via local app | File import | Local Python | Conditional license; own GPU |
| Stability SPAR3D [8](https://github.com/Stability-AI/stable-point-aware-3d) [9](https://stability.ai/news-updates/stable-point-aware-3d) [10](https://huggingface.co/stabilityai/stable-point-aware-3d) [11](https://github.com/Stability-AI/stable-point-aware-3d/blob/main/LICENSE.md) | Editing points before meshing | Single image; point edits | Textured, UV-unwrapped mesh | GLB | File import | Local Python reviewed here | Conditional license; own compute |
| Hydrilla [12](https://hydrilla.ai/research#model-card) [13](https://hydrilla.ai/features) [14](https://hydrilla.ai/text-to-3d) [15](https://hydrilla.ai/agent) [19](https://hydrilla.ai/integrations) [18](https://docs.hydrilla.ai/api/text-to-3d) [21](https://hydrilla.ai/pricing) | Reference image to inspected mesh, then handoff | OpenAI or Gemini image; upload also works | Segmented mesh with UVs and PBR maps | GLB, FBX, OBJ, USDZ | Blender, Unreal, and others via Bridge | Hosted REST; BlueFox only | $3/mo (daily window); $5/mo anytime |
| Meshy [22](https://www.meshy.ai/pricing) [23](https://docs.meshy.ai/en/api/image-to-3d) [24](https://docs.meshy.ai/en/api/multi-image-to-3d) [25](https://docs.meshy.ai/en/api/text-to-3d) [26](https://www.meshy.ai/integrations) | Generating, remeshing, and rigging in one service | Text, image, or multiple views | Mesh; PBR optional in API | GLB, FBX, OBJ, USDZ, and others | Blender, Unity, Unreal, and others | Hosted REST | $20/mo Pro at regular monthly price |
| Tripo [27](https://www.tripo3d.ai/pricing) [28](https://developers.tripo3d.ai/en/pricing) [29](https://docs.tripo3d.ai/) [30](https://developers.tripo3d.ai/en/docs/plugins) [31](https://developers.tripo3d.ai/en/docs/models-convert) | Turning several views into one editable draft | Text, image, or multiple views | Mesh; texturing and rigging options | GLB, FBX, OBJ, and others via API conversion | Blender, Unity, Unreal, and others | Hosted REST; separate credits | $240/yr Pro ($20/mo equivalent) |
| Hyper3D Rodin [32](https://docs.hyper3d.ai/en/api-specification/rodin-gen2-5) [33](https://hyper3d.ai/pricing) [34](https://hyper3d.ai) | Specifying material output in hosted jobs | Text or up to five images (Gen-2.5 API) | Mesh with PBR, shaded, hybrid, or no material | GLB, USDZ, FBX, OBJ, STL | Blender, Unity, Unreal, and others | Hosted REST; check access tier | $1.50 per direct credit; subscriptions available |
| Sloyd [35](https://www.sloyd.ai) [36](https://www.sloyd.ai/pricing) [37](https://www.sloyd.ai/api/pricing) | Variations on a shared template | Template, text, or images | Customizable templates or generated textured meshes | GLB, OBJ, FBX | Available on paid web plans | Hosted REST; separate prepaid credits | $15/mo Plus, monthly billing |
| Kaedim [38](https://www.kaedim3d.com) [39](https://www.kaedim3d.com/faq) [40](https://www.kaedim3d.com/plans) [41](https://docs.kaedim3d.com/welcome/get-started) | Commissioning a 3D asset to spec | Brief, sketches, photos, or reference pack | Artist-reviewed geometry or finished assets | Specify in the brief | Confirm your pipeline with Kaedim | Documented for Enterprise | $400/mo Indie subscription |

## The 9 Microsoft TRELLIS alternatives worth testing

### 1. Microsoft TRELLIS.2: best for self-hosted PBR meshes from an image [3](https://github.com/microsoft/TRELLIS.2) [4](https://github.com/microsoft/TRELLIS.2/blob/main/example_texturing.py) [1](https://github.com/microsoft/TRELLIS)

**What is Microsoft TRELLIS.2?** Microsoft's December 16, 2025 release is a separate 4B-parameter image-to-3D model, not a new name for TRELLIS-image-large. It generates mesh materials with base color, roughness, metallic, and opacity attributes.

**Strengths**

- Microsoft provides a separate [texturing example](https://github.com/microsoft/TRELLIS.2/blob/main/example_texturing.py) for an existing mesh and a reference image. Try that if the silhouette is already approved.
- The GLB exporter retains texture alpha but writes the material as opaque by default. For a cutout or translucent prop, test the alpha settings in the destination shader.

**Where it isn't the answer.** The released checkpoint takes an image, not a text prompt. Stay with original TRELLIS for the Gaussian or radiance-field output you already use. Microsoft's TRELLIS.2 setup is Linux-tested, needs at least 24 GB of NVIDIA GPU memory, and recommends CUDA 12.4.

**Pricing.** Downloadable under MIT for Microsoft's model and code; some dependencies have separate licenses. Budget for a GPU and maintenance rather than a per-generation vendor fee.

### 2. Hunyuan3D 2.1: best for running local shape and paint stages separately [5](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1) [6](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/LICENSE) [7](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/gradio_app.py)

**What is Hunyuan3D 2.1?** Tencent released the 2.1 shape and paint checkpoints on June 14, 2025. The shape stage turns an image into geometry; the paint stage takes a mesh and an image to synthesize PBR materials.

**Strengths**

- Approve the untextured silhouette first. With separate pipelines, a rejected material needn't send you back to regenerate the shape.
- Tencent reports 10 GB of VRAM for shape, 21 GB for texture, and 29 GB for both. That makes geometry-only worth testing even if your GPU doesn't fit the full workflow.

**Where it isn't the answer.** Don't read the README's macOS, Windows, and Linux list as proof that the same setup works on each. The published test setup pins CUDA 12.4 PyTorch. Test the paint stage on the machine that will actually run it.

**Pricing.** No metered fee for the local download under the [Tencent community license](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/LICENSE) when your use qualifies. It excludes the EU, UK, and South Korea and has a separate license requirement above one million monthly active users, measured as specified in the agreement.

### 3. Stability SPAR3D: best for adjusting the silhouette before generating a mesh [8](https://github.com/Stability-AI/stable-point-aware-3d) [9](https://stability.ai/news-updates/stable-point-aware-3d) [10](https://huggingface.co/stabilityai/stable-point-aware-3d) [11](https://github.com/Stability-AI/stable-point-aware-3d/blob/main/LICENSE.md)

**What is Stability SPAR3D?** Released January 8, 2025, Stable Point Aware 3D reconstructs a UV-unwrapped, textured mesh from a single image. The intermediate point cloud gives an artist a place to make shape edits before the model creates the final mesh.

**Strengths**

- Move or delete points to correct an unwanted projection on a prop before exporting GLB.
- The local repo documents a low-VRAM mode and a CPU fallback. Windows and Apple Silicon MPS support are marked experimental; the Mac notes specify macOS 15.2 or later.

**Where it isn't the answer.** It isn't a multi-view reconstruction pipeline. If you need authored metallic and roughness maps or want to retexture approved geometry, test TRELLIS.2 instead.

**Pricing.** Weights require accepting access terms. The [repository license](https://github.com/Stability-AI/stable-point-aware-3d/blob/main/LICENSE.md) requires registration for commercial use and a separate license above $1 million in annual revenue, including affiliates. The model card says $1 million or more, so confirm terms directly if you're at that boundary.

### 4. Hydrilla: best for carrying one reference image through to an inspected mesh [12](https://hydrilla.ai/research#model-card) [13](https://hydrilla.ai/features) [14](https://hydrilla.ai/text-to-3d) [15](https://hydrilla.ai/agent) [16](https://hydrilla.ai/integrations/mcp) [19](https://hydrilla.ai/integrations) [20](https://hydrilla.ai/api) [18](https://docs.hydrilla.ai/api/text-to-3d) [21](https://hydrilla.ai/pricing)

**What is Hydrilla?** Hydrilla Studio chains the steps: choose OpenAI or Gemini to make or edit the reference image, run BlueFox 1, Hydrilla's own 3D model, on it, then inspect the segmented mesh with UVs and PBR maps.

**Strengths**

- Check wireframe, face counts, and selectable parts in Studio, so you know whether the shade is separate and what the count costs before an artist takes it.
- Send the mesh through the local Bridge into Blender, Unity, Unreal, or another supported app, or export it. The [REST API](https://hydrilla.ai/api) and its [text-to-3D endpoint](https://docs.hydrilla.ai/api/text-to-3d) take the source-image provider; Hydrilla Agent and [MCP](https://hydrilla.ai/integrations/mcp) cover agent-led work.

**Where it isn't the answer.** Wireframe and face counts are inspection, not retopology, so budget the retopo pass. A local Bridge isn't local inference: if references can't leave your network or you need Gaussians, keep Microsoft TRELLIS.

**Pricing.** Free includes 100 credits; Hourly costs $3/month in a daily window, Starter $5/month without one. Image-to-model takes 30 credits, text 45 with the source image.

### 5. Meshy: best for trying generation and common cleanup steps in one service [22](https://www.meshy.ai/pricing) [23](https://docs.meshy.ai/en/api/image-to-3d) [24](https://docs.meshy.ai/en/api/multi-image-to-3d) [25](https://docs.meshy.ai/en/api/text-to-3d) [26](https://www.meshy.ai/integrations)

**What is Meshy?** Meshy offers hosted generation from a prompt, a single image, or multiple photos. Its product separates generation from operations such as remeshing and rigging, so artists can test more than one stage before configuring an external pipeline.

**Strengths**

- The multi-image API takes up to four views of the same object. Use it to check whether a distinct rear panel survives reconstruction.
- For the image API, switch on PBR to request metallic, roughness, and normal maps alongside base color. The texture model matters: Meshy N1 is base-color-only and doesn't accept that setting.

**Where it isn't the answer.** Don't choose it to pin and operate your own weights. If you're migrating a script, the docs say Meshy 7 geometry generation retires December 31, 2026; Meshy 7 texturing doesn't retire on that date.

**Pricing.** Pro lists a regular $20 monthly price before its first-month promotion. Meshy lists CC BY 4.0 on Free and a private commercial license on Pro; check which terms apply before shipping or sharing a test asset.

### 6. Tripo: best for working from a consistent set of reference views [27](https://www.tripo3d.ai/pricing) [28](https://developers.tripo3d.ai/en/pricing) [29](https://docs.tripo3d.ai/) [30](https://developers.tripo3d.ai/en/docs/plugins) [31](https://developers.tripo3d.ai/en/docs/models-convert)

**What is Tripo?** Tripo offers hosted generation from text, a single image, and multi-view references. Its API also documents texture, mesh editing, rigging, and conversion as processing steps after generation.

**Strengths**

- Use multi-view generation when the rear view contains a feature the front image hides. Check that the shape agrees across the two views before spending credits on finishing.
- Try the Blender plugin for image-to-model work if you want to adjust face limit, texture, or PBR before importing. Tripo documents different capabilities for each plugin.

**Where it isn't the answer.** The plugins use a Tripo API key and don't replace self-hosting. If a character will be animated, check deformation at the joints before treating a generated rig as final.

**Pricing.** Tripo Studio Pro lists $240 billed yearly, or $20/month equivalent. Free Studio models are public and for noncommercial use. The API uses a separate pay-as-you-go rate card, where processing such as format conversion can be a separate charge.

### 7. Hyper3D Rodin: best for choosing the material package on an API job [32](https://docs.hyper3d.ai/en/api-specification/rodin-gen2-5) [33](https://hyper3d.ai/pricing) [34](https://hyper3d.ai)

**What is Hyper3D Rodin?** Rodin is Hyper3D's hosted generator. Its Gen-2.5 API takes up to five reference images or starts from a prompt that it first renders into an image. The endpoint exposes material and geometry settings.

**Strengths**

- Request PBR with base color, metallic, roughness, and normal maps for relighting, or request shaded output for a look with baked lighting. You can also request geometry without material.
- Specify the destination format, such as GLB for a real-time scene or STL for a geometry-only handoff, and check the resulting file in the destination tool.

**Where it isn't the answer.** It doesn't replace a studio-controlled runtime. If you use the API, confirm access and price at your tier instead of extrapolating from a Workspace subscription or preview.

**Pricing.** Hyper3D lists $1.50 for a direct credit and documents full API access on Business. The Gen-2.5 rate card lists 0.5 credits for base generation, with separate charges for extreme-high geometry and texture options.

### 8. Sloyd: best for building a family of props from an editable template [35](https://www.sloyd.ai) [36](https://www.sloyd.ai/pricing) [37](https://www.sloyd.ai/api/pricing)

**What is Sloyd?** Sloyd offers hosted text- and image-to-3D alongside handcrafted parametric templates. If a template already matches the object class, artists can change its parameters rather than ask a model to reinterpret the same reference for every variant.

**Strengths**

- Test the template approach on furniture in several sizes or on related set pieces that should share proportions and style.
- Use GLB, OBJ, or FBX for handoff, and check whether your destination is covered by the paid-plan plugins. For physically based shading, inspect the maps from the exact generator and export path you plan to use.

**Where it isn't the answer.** A template is a poor starting point if none fits the brief. And if your business sells the exported assets as marketplace items, don't assume the Plus commercial license permits redistribution.

**Pricing.** Plus is $15 month to month or $150 billed yearly. Plus and Pro include commercial use; Pro adds marketplace resale rights. Multi-image generation uses credits, and API credits are bought separately from the web plan.

### 9. Kaedim: best for handing off a written asset brief with a review cycle [38](https://www.kaedim3d.com) [39](https://www.kaedim3d.com/faq) [40](https://www.kaedim3d.com/plans) [41](https://docs.kaedim3d.com/welcome/get-started)

**What is Kaedim?** Kaedim is a managed 3D production workflow that pairs AI generation with artist review. A studio can submit sketches, a reference pack, or a written brief, then mark up an asset before approving delivery.

**Strengths**

- Write the polygon budget, UV requirements, material maps, and engine format into the brief. Ask for one sample delivery before sending an asset list.
- The review process makes sense when the studio needs someone else to handle revisions against an art guide, rather than another tool for its artists to operate.

**Where it isn't the answer.** Don't compare it to local inference by cost per attempt. You're buying a delivery workflow with a human review step, so compare turnaround, revision scope, and the final files against the work you would otherwise do in-house.

**Pricing.** Indie lists $400/month for 20 credits: up to 20 geometry-only deliveries or one complete asset with retopology, UVs, and textures. Kaedim documents API access for Enterprise; confirm what is included for your brief.

## Which Microsoft TRELLIS alternative should you choose?

| If you need | Pick |
| --- | --- |
| Keep the Gaussian PLY export or a pinned original pipeline. | Stay with Microsoft TRELLIS. |
| Relight an image-generated mesh or texture an approved mesh locally. | Microsoft TRELLIS.2, if you have a suitable NVIDIA/Linux setup. |
| Approve shape first, then run material synthesis separately. | Hunyuan3D 2.1, if your region and deployment qualify under Tencent's terms. |
| Remove or reposition points before rebuilding the mesh. | Stability SPAR3D, if your hardware and commercial license fit. |
| Go from a reference image to a mesh you have inspected. | Hydrilla: OpenAI or Gemini makes the image, BlueFox 1 makes the mesh, Studio checks wireframe, counts, and parts, then you export or send it over the Bridge. |
| Give an AI agent a 3D tool it can call. | Hydrilla Agent with your own provider key for Blender work, or Hydrilla MCP for an outside agent. |
| Try remeshing and rigging without assembling a separate toolchain. | Meshy. Check topology and a test pose before committing the asset. |
| Use views from several angles as source material. | Tripo for a multi-view workflow; compare with Meshy on the same views. |
| Set material mode and export format in a hosted job. | Hyper3D Rodin Gen-2.5. Use the API rate card for the exact settings you send. |
| Make a family of props with shared proportions. | Sloyd, if an existing parametric template fits. |
| Review a delivered asset against a written art and technical brief. | Kaedim. Scope one sample before ordering the set. |

## Is Hydrilla right for you?

### Choose Hydrilla if

- Your pipeline starts with a reference image. Studio lets you choose OpenAI or Gemini to make or edit it, then runs BlueFox 1 on the result.
- You want wireframe, face counts, and selectable parts in one place before the mesh goes to an artist in Blender, Unity, or Unreal.
- Your team works through an agent: Hydrilla Agent with your own provider key for Blender scenes, or MCP so an outside AI agent can call Hydrilla generation.

### Choose something else if

- Keep Microsoft TRELLIS if inference has to stay on your own hardware, or if you need Gaussians and radiance fields. Pick TRELLIS.2 for self-hosted PBR.
- Choose Sloyd if editable parameters matter more than reconstructing a particular reference.
- Choose Kaedim if you want to review a delivery against a brief rather than run generation and cleanup yourself.
- Go to Meshy, Tripo, or Rodin directly when one of their own controls decides the job: a remesh or face target, multi-view conditioning, or a named material package.

## Questions about Microsoft TRELLIS alternatives

### Are Microsoft TRELLIS and TRELLIS.2 the same model?

No. [Original TRELLIS](https://github.com/microsoft/TRELLIS) has separate image and text checkpoints and can produce meshes, Gaussians, and radiance fields. [TRELLIS.2-4B](https://github.com/microsoft/TRELLIS.2) is a separate image-to-3D release centered on meshes with physically based materials.

### Can I run TRELLIS locally on a Mac?

Microsoft documents Linux-only testing and NVIDIA GPU requirements for [TRELLIS](https://github.com/microsoft/TRELLIS) and [TRELLIS.2](https://github.com/microsoft/TRELLIS.2), so don't plan to run the official setup on Apple Silicon. [SPAR3D](https://github.com/Stability-AI/stable-point-aware-3d) documents experimental MPS support and a CPU fallback. You can also run inference on a separate machine and import the output on a Mac.

### Can I use Microsoft's TRELLIS releases for paid client work?

Microsoft lists MIT for the original [TRELLIS](https://github.com/microsoft/TRELLIS) models and most of the code, and [TRELLIS.2](https://github.com/microsoft/TRELLIS.2) lists MIT for its model and code. Review dependencies separately. Get permission to use client references, and review the host's terms if you aren't running inference yourself.

### Does a textured GLB mean I get separate PBR maps?

No. The [original TRELLIS exporter](https://github.com/microsoft/TRELLIS/blob/main/trellis/utils/postprocessing_utils.py) uses a baked base-color texture and fixed roughness. [TRELLIS.2](https://github.com/microsoft/TRELLIS.2) models attributes such as roughness and metallic. Open the actual export in your renderer and check the connected channels and alpha mode.

### Can I generate a 3D asset from text without a reference image?

Yes. [Original TRELLIS](https://github.com/microsoft/TRELLIS) has separate text checkpoints, though Microsoft recommends an image first, and [TRELLIS.2](https://github.com/microsoft/TRELLIS.2) needs an image. In Hydrilla Studio and its [text-to-3D API](https://docs.hydrilla.ai/api/text-to-3d), the text step builds the reference image with OpenAI or Gemini before BlueFox 1 runs.

### Does Hydrilla run Meshy, Tripo, Rodin, or Microsoft TRELLIS models?

No. In the current release, BlueFox 1 is the supported 3D generation model in Studio and the REST API, and the source-image step uses OpenAI or Gemini. Other names in Hydrilla's [public model directory](https://hydrilla.ai/ai-models) are references, not selectable generation paths. To use a vendor's mesh model, go to that vendor.

### How should I compare self-hosted and hosted prices?

Measure cost per approved asset on one representative brief. For self-hosted runs, count GPU time and the time spent maintaining the setup. For hosted jobs, count the plan or credits, retried jobs, and any finishing charges. For both, include the artist time needed after import.

## Sources

1. [Microsoft: original TRELLIS checkpoints, setup, and license](https://github.com/microsoft/TRELLIS), checked 20 Nov 2026
2. [Microsoft: original TRELLIS material in the GLB exporter](https://github.com/microsoft/TRELLIS/blob/main/trellis/utils/postprocessing_utils.py), checked 20 Nov 2026
3. [Microsoft: TRELLIS.2 release, PBR, setup, and license](https://github.com/microsoft/TRELLIS.2), checked 20 Nov 2026
4. [Microsoft: TRELLIS.2 example for texturing existing geometry](https://github.com/microsoft/TRELLIS.2/blob/main/example_texturing.py), checked 20 Nov 2026
5. [Tencent: Hunyuan3D 2.1 checkpoints, setup, and VRAM figures](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1), checked 20 Nov 2026
6. [Tencent: Hunyuan3D 2.1 Community License Agreement](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/LICENSE), checked 20 Nov 2026
7. [Tencent: Hunyuan3D 2.1 local app and example exports](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1/blob/main/gradio_app.py), checked 20 Nov 2026
8. [Stability AI: SPAR3D setup, hardware notes, and output](https://github.com/Stability-AI/stable-point-aware-3d), checked 20 Nov 2026
9. [Stability AI: SPAR3D release and point editing](https://stability.ai/news-updates/stable-point-aware-3d), checked 20 Nov 2026
10. [Stability AI: SPAR3D model card and weight access](https://huggingface.co/stabilityai/stable-point-aware-3d), checked 20 Nov 2026
11. [Stability AI: SPAR3D community license and commercial threshold](https://github.com/Stability-AI/stable-point-aware-3d/blob/main/LICENSE.md), checked 20 Nov 2026
12. [Hydrilla: BlueFox 1 model card](https://hydrilla.ai/research#model-card), checked 20 Nov 2026
13. [Hydrilla: Studio preview, parts, materials, and exports](https://hydrilla.ai/features), checked 20 Nov 2026
14. [Hydrilla: text-to-3D workflow through the reference image](https://hydrilla.ai/text-to-3d), checked 20 Nov 2026
15. [Hydrilla: Agent with your own provider key](https://hydrilla.ai/agent), checked 20 Nov 2026
16. [Hydrilla: MCP connector](https://hydrilla.ai/integrations/mcp), checked 20 Nov 2026
17. [Hydrilla: public model directory and reference entries](https://hydrilla.ai/ai-models), checked 20 Nov 2026
18. [Hydrilla: text-to-3D API source-image provider parameter](https://docs.hydrilla.ai/api/text-to-3d), checked 20 Nov 2026
19. [Hydrilla: plugins and the local Bridge](https://hydrilla.ai/integrations), checked 20 Nov 2026
20. [Hydrilla: hosted generation API](https://hydrilla.ai/api), checked 20 Nov 2026
21. [Hydrilla: web plans and credit usage](https://hydrilla.ai/pricing), checked 20 Nov 2026
22. [Meshy: web plans, licensing, and production tools](https://www.meshy.ai/pricing), checked 20 Nov 2026
23. [Meshy: single-image API and PBR texture options](https://docs.meshy.ai/en/api/image-to-3d), checked 20 Nov 2026
24. [Meshy: multi-image API inputs](https://docs.meshy.ai/en/api/multi-image-to-3d), checked 20 Nov 2026
25. [Meshy: text-to-3D API and model retirement notice](https://docs.meshy.ai/en/api/text-to-3d), checked 20 Nov 2026
26. [Meshy: DCC and engine integrations](https://www.meshy.ai/integrations), checked 20 Nov 2026
27. [Tripo: Studio pricing, license, and supported jobs](https://www.tripo3d.ai/pricing), checked 20 Nov 2026
28. [Tripo: API credit rate card](https://developers.tripo3d.ai/en/pricing), checked 20 Nov 2026
29. [Tripo: generation and model processing API](https://docs.tripo3d.ai/), checked 20 Nov 2026
30. [Tripo: plugins, compatibility, and supported features](https://developers.tripo3d.ai/en/docs/plugins), checked 20 Nov 2026
31. [Tripo: supported API format conversion](https://developers.tripo3d.ai/en/docs/models-convert), checked 20 Nov 2026
32. [Hyper3D: Rodin Gen-2.5 inputs, materials, formats, and API charges](https://docs.hyper3d.ai/en/api-specification/rodin-gen2-5), checked 20 Nov 2026
33. [Hyper3D: direct credit pricing and plan access](https://hyper3d.ai/pricing), checked 20 Nov 2026
34. [Hyper3D: Rodin workspace and 3D add-ons](https://hyper3d.ai), checked 20 Nov 2026
35. [Sloyd: parametric templates and image- and text-led generation](https://www.sloyd.ai), checked 20 Nov 2026
36. [Sloyd: web plan prices, exports, and license terms](https://www.sloyd.ai/pricing), checked 20 Nov 2026
37. [Sloyd: separate API pricing](https://www.sloyd.ai/api/pricing), checked 20 Nov 2026
38. [Kaedim: asset intake, markup, and review workflow](https://www.kaedim3d.com), checked 20 Nov 2026
39. [Kaedim: artist review and production specifications](https://www.kaedim3d.com/faq), checked 20 Nov 2026
40. [Kaedim: Indie price and geometry versus complete-asset allowance](https://www.kaedim3d.com/plans), checked 20 Nov 2026
41. [Kaedim: Enterprise API access](https://docs.kaedim3d.com/welcome/get-started), checked 20 Nov 2026
