Total Program Failure: Trellis2 DINOv3 extractor cannot load the HuggingFace DINOv3 model

[jimpickrellgit]

I’m trying to use the DINOv3 extractor in Trellis2, but it doesn’t load the HuggingFace model (facebook/dinov3-vitl16-pretrain-lvd1689m). The extractor expects model.layer, but the HF model exposes model.model.layer. It also expects RoPE to return two tensors (cos, sin), while the HF model returns a single tensor, causing ValueError: expected 2, got 1.

Before I assume this is a bug, can you confirm whether the DINOv3 extractor is intended to work with the public HF DINOv3 model, or if a different model is required? Alternately is there a different version of the program I should be using?

Has anyone else been able to get this to work?

This seems like an absolute dealbreaker as far as using this software at all. If this doesn't work nothing works.

[jimpickrellgit]

I was able to fix this but it required edits to the source code to match the software and the model. It's like having an sql program with the wrong schema. Since this was required, I don't think any fix that was done by installing versions of some other software would have helped. I would submit the changes to the project but I'm not actually an author so now sure what to do about it.