Distilling from EUPE rather than DINOv3 as teacher

[elnursahbalayev]

First of all, thank you so much for amazing work.

Recently Meta has released another foundational model with exact same structure as DINOv3 but with better performance called EUPE (Efficient Universal Perception Encoder).

Would you be considering running the same pipeline you did, but with EUPE weights instead of DINOv3. It generates much better feature maps.

[capsule2077]

Hi, thank you so much for bringing EUPE to our attention!

We’re definitely aware of Meta’s EUPE and its improvements over DINOv3 — it’s already on our roadmap and we plan to run the same distillation pipeline with EUPE weights as soon as we can. That said, our team is currently constrained on compute resources, so we likely won’t be able to push an update very quickly. We’ll keep the community posted once we make progress.

Thanks again for the great suggestion!

[XiSHEN0220]

Hi, thanks for your interest in our project.

We did try using EUPE as the teacher, but so far it hasn't given us better performance.
We're still tuning some hyperparameters and running more experiments, but at the moment the results don't look very promising.

We'll update the manuscript once we have more results. Stay tuned!

Best,