Open Source Image to 3D: The Models Behind This Lane

Open source image to 3D explained: which upstream models are public with their own licences, and what this independent hosted service does or does not provide.

Open source image to 3D is really two questions in one search. Which reconstruction models can you actually inspect, download, and run yourself — and is the service in front of them open too? This page answers both for pixal3d.ai, because the answer differs between the model layer and the hosting layer.

The short version: the reconstruction research this lane is built on is public, the hosted product you are using is not, and the difference decides what you can do locally versus what you pay for here.

Which layers are open, and which are not

Layer	Status	What that means for you
Upstream Pixal3D research model	Public repository and weights	You can read the code, check the licence, and run it locally if you have the hardware
Trellis 2 baseline	Public repository	The implementation Pixal3D's current branch builds on is publicly documented
This hosted service	Independently operated, not published as source or weights	You use it in the browser; there is nothing to download here
Provider infrastructure	Third-party	The hosted lanes run through a model provider, as the wrapper disclaimer explains

That division is deliberate. This site is a workflow layer — accounts, credits, task history, browser tools, and policies — sitting on top of research models, and the Mission page describes it in the same terms.

What the upstream projects publish

The image-to-3D lane here is based on the open Pixal3D research model. Its official repository documents a technique it calls back-projection, which lifts pixel features from the source image into 3D instead of injecting them loosely through attention, and that is the mechanism behind this site's pixel-alignment claims.

Repository: github.com/TencentARC/Pixal3D publishes the code and describes the two implementations it keeps — a main branch built on TRELLIS.2 and a paper branch built on Direct3D-S2 that produced the published results.
Weights: the project's Hugging Face files are published as .safetensors checkpoints for local inference.
Baseline: the TRELLIS.2 implementation the current branch builds on is documented by Microsoft's TRELLIS.2 repository.
Licence: the upstream project is released under the MIT License, as noted in the 3D AI guide. The hosted service you are reading now is a separate product with its own terms — the Terms of Service apply here, not the upstream licence.

Both things can be true at once: the model can be open while the service that runs it for you is a commercial workflow layer with billing and policies of its own.

Self-hosting versus the hosted lane

Question	Run the open model yourself	Use this hosted lane
Setup	Linux, CUDA, and a compiled dependency stack	Browser sign-in
Hardware	A capable local GPU	None; the models run on the provider
Control	Full code, weights, and configuration access	Managed lanes, presets, and settings
Multi-view input	Available in the upstream repository	Not available; one source image per job
Cost model	Your own hardware and time	Credits per job: 180–390 for Pixal3D, 160–310 for Trellis 2
Reproducibility	You keep the exact environment	You keep the task history, not the environment
Best for	Research, custom pipelines, offline work	Getting a GLB from an image without a GPU

If your reason for searching open source is that you want to modify the model or run it in a closed environment, the upstream repositories are the right place to start. If you want the output without maintaining a CUDA stack, the AI 3D Generator wraps the same research line with a task pipeline, credit accounting, and inspection tools.

What you cannot do from here

Download the weights. The hosted service does not republish model checkpoints. Use the upstream Hugging Face page for those.
Access the hosting source. This site's hosted implementation is not published as source code.
Get a GGUF build. The upstream checkpoints are distributed as .safetensors files, not as a GGUF distribution.
Use multi-view conditioning. The open repository provides multi-view inference for local runs; the hosted lane accepts one image per job.
Assume the upstream licence covers your output. Licence terms are the place to check before shipping generated assets commercially, and the hosted service has its own terms as well.

Common mistakes

Mistake	Why it goes wrong	Correct reading
Assuming pixal3d.ai is an open-source project	The models are open; the service is a commercial workflow layer	Read the wrapper disclaimer for the boundary
Downloading locally without checking the hardware path	Local inference starts from the TRELLIS.2 baseline and assumes a CUDA environment	Review the upstream install notes before committing to a local setup
Treating model licence as commercial clearance	Rights depend on the licence and on the terms of the service you used	Check both before shipping
Expecting this site to host the research demo	This is an independent hosted lane, not the research team's page	Use the upstream project page for the official demo
Searching for a GGUF of a 3D model	The checkpoints are `.safetensors` files	Plan for the documented inference path, not a quantised single-file build

Costing the trade-off

Self-hosting looks free because there is no invoice, but it costs a GPU environment, setup time, and maintenance. The hosted route costs credits and nothing else: 180–390 for a Pixal3D run, 160–310 for Trellis 2, with the exact price shown before submission. New accounts get 220 one-time credits valid for 30 days, which covers one default job — see what the free credits cover. The browser tools that inspect and convert the output are free either way.

For background on the models themselves, the Pixal3D complete guide and the Trellis 2 guide go deeper, and the 3D AI guide compares the lanes.

If you want the result rather than the repository, open the AI 3D Generator and run one image through the pixel-aligned lane.

FAQ

Is the image-to-3D model here open source?

The upstream research model is. Pixal3D is published by its research team on GitHub with a documented licence, and the TRELLIS.2 implementation its current branch builds on is also public. This website is a separate hosted workflow layer built on top of those models.

Is pixal3d.ai itself open source?

No. This site publishes a hosted service, not source code or model weights for download, and it is not affiliated with the research teams. The wrapper disclaimer sets out the boundary between the upstream models and this service.

Can I download the model weights here?

No. The hosted service does not republish checkpoints. The upstream project publishes its weights as safetensors files for local inference, and local runs start from the documented CUDA installation path, not from a single-file quantised build.

Should I self-host or use the hosted lane?

Self-host when you need to modify the model, run multi-view inference, or keep the work offline, and accept the Linux, CUDA, and GPU requirements. Use the hosted lane when you want a GLB without maintaining that environment; runs cost 180 to 390 credits for Pixal3D and 160 to 310 for Trellis 2.