Skip to content

Clarification on licensing for commercial use (model weights, DINOv3, nvdiffrast, RMBG, EU) #33

Description

@seriousEnthusi84272

Hi, and thank you for releasing Pixal3D — the fidelity is impressive.

I'm evaluating Pixal3D for a commercial product (a video game sold to end users), and I want to make sure I use it correctly and within the license. I've found some conflicting information across the repo and third-party wrappers, so I'd be grateful for authoritative answers on the following:

1. License of the model weights.
The GitHub repository is marked MIT, but at least one third-party ComfyUI wrapper states the released weights are "academic-only / non-commercial / EU-excluded" (citing a Tencent license). Which is correct? Specifically:

  • Are the released checkpoints under MIT, or under a separate Tencent non-commercial/community license?
  • If a separate license applies to the weights, could you link it in the repo?

2. Commercial use of generated assets.
If I generate 3D meshes with Pixal3D and then edit/retopologize/texture them, am I allowed to sell those assets (e.g. as part of a commercial game)? Are the outputs free of restrictions, or do they inherit any license terms?

3. Territory.
Is there any geographic restriction (e.g. EU / UK / South Korea excluded) on commercial use? I'm based in the EU (Italy).

4. DINOv3 dependency.
The released code downloads dinov3-vitl16 (Meta, gated). The paper mentions DINOv2-Large as the image encoder.

  • Which encoder do the released weights actually require at inference — DINOv2 or DINOv3?
  • Is it possible to run the released model with DINOv2 (Apache-2.0) instead, without retraining and without quality loss? If so, how?

5. Non-commercial dependencies (nvdiffrast, RMBG).

  • nvdiffrast (NVIDIA Source Code License, non-commercial) appears to be used for texture baking. Is it possible to run a geometry-only pipeline that does not invoke nvdiffrast, so the mesh output is free of that dependency?
  • For background removal, is RMBG-2.0 (CC BY-NC) required, or can I substitute a permissively-licensed remover (e.g. rembg / U²-Net) or supply a pre-masked image?

6. Multi-view input.
Does the released main branch support multi-view image input (e.g. 4 views) at inference, or is it single-image only? If multi-view is supported, how are the views passed?

Thank you very much for your time — a clear answer on the weights license (Q1) and EU/commercial use (Q2–3) in particular would help me a lot in deciding whether I can build on Pixal3D. Happy to provide any details.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions