Add DINOv3ViTForImageClassification support

[dimidagd]

@yonigozlan @molbap

[molbap]

Thanks! Left two comments

[dimidagd]

Discussed reviewers comments, no immediate points of action. Left for future

• Refactoring to remove redundant code between Dinov2 and DINOv3
• Convert meta checkpoint with classifier and add a regression test

@molbap would you like to approve this PR?

[molbap]

Sure, just need to check get_input_embeddings, as in most cases it's not needed to add it and we prefer to add as little code as possible. It's not needed in Dinov2, why here?
Otherwise LGTM and ok to merge once this is sorted out!

[dimidagd]

@molbap what is the process for merging PRs after they have been approved? Trying to understand better the contribution workflow

Best
DD

[molbap]

I have a pending question on which test was failing for the embeddings, please let me know! apart from that for the review process when reviewers have approved we ping the core maintainers to approve and merge your PR!

This one in particular launched an internal discussion since we're trying to minimize the maintenance surface, see here #41450 but it'll wait a follow-up PR.

[molbap]

Added a missed head_mask, other tham that ping me when you know for the failing test and we can merge this! With #41276 it's nice additions to dinov3

[dimidagd]

Added a missed head_mask, other tham that ping me when you know for the failing test and we can merge this! With #41276 it's nice additions to dinov3

Here is the failing test

https://app.circleci.com/jobs/github/huggingface/transformers/1982122

[dimidagd]

Added a missed head_mask, other tham that ping me when you know for the failing test and we can merge this! With #41276 it's nice additions to dinov3

Here is the failing test

https://app.circleci.com/jobs/github/huggingface/transformers/1982122

on a second thought, I removed the associated test in 1fb6000

Looking forward to your feedback

[dimidagd]

Added a missed head_mask, other tham that ping me when you know for the failing test and we can merge this! With #41276 it's nice additions to dinov3

@molbap upon rereading your comments, I think you implied that adding the get_input_embeddings should not be necessary.

Here is the failing test, which was copied over from dinov2.

https://app.circleci.com/jobs/github/huggingface/transformers/1993707

[molbap]

Hi @dimidagd , circling back to this, this is to support the classification head released by Meta for dinov3, right? In that case, it would also be needed to have a small test to load that head and test it works to avoid future regressions. I'm referring to this https://github.com/facebookresearch/dinov3?tab=readme-ov-file#pretrained-heads---image-classification

That way there'd be more ground to support the implementation of a new head!

[dimidagd]

@molbap I will see if I can find the time for this, what are your thoughts on the `get_input_embeddings` test?

…

On Fri, 31 Oct 2025 at 16:03, Pablo Montalvo ***@***.***> wrote: *molbap* left a comment (huggingface/transformers#41224) <#41224 (comment)> Hi @dimidagd <https://github.com/dimidagd> , circling back to this, this is to support the classification head released by Meta for dinov3, right? In that case, it would also be needed to have a small test to load that head and test it works to avoid future regressions. I'm referring to this https://github.com/facebookresearch/dinov3?tab=readme-ov-file#pretrained-heads---image-classification That way there'd be more ground to support the implementation of a new head! — Reply to this email directly, view it on GitHub <#41224 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALECAUJ5KWWT4LDURHPDCB332N24LAVCNFSM6AAAAACH4V2IVCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTINZTGQ4DGOJWGQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>

-- [image: photo] *Dagdilelis Dimitris | PhD* m:+45 52 60 18 25 <+45+52+60+18+25> | ***@***.*** | a:Gullandsgade 18, 2800 Lyngby <http://us.linkedin.com/in/dimidagd>

[merveenoyan]

hey @dimidagd thanks a lot for working on this!

Meta has three checkpoints for image classification (with linear layer), object detection (DINO3+DETR) and depth estimation (DINO3+DPT). Would you like to load image classification weights and push to Hub for people's convenience?

We have DETR and DPT officially supported in transformers. Would you be down to convert the depth and detector checkpoints too?

[dimidagd]

hey @dimidagd thanks a lot for working on this!

Meta has three checkpoints for image classification (with linear layer), object detection (DINO3+DETR) and depth estimation (DINO3+DPT). Would you like to load image classification weights and push to Hub for people's convenience?

We have DETR and DPT officially supported in transformers. Would you be down to convert the depth and detector checkpoints too?

Hi @merveenoyan! My dev machine is on codespaces, with limited memory, therefore I can't even load their checkpoints. Do you have access to resources I could use?

[merveenoyan]

@dimidagd can you use a Colab? 👀 using small checkpoints to validate implementation is ok!

[dimidagd]

@dimidagd can you use a Colab? 👀 using small checkpoints to validate implementation is ok!

@merveenoyan hi again, meta released the classification head only for the 7B model. I can't fit that on colab resources either (GPU or CPU). I can't imagine myself implementing such tests from a colab env tbh.

[molbap]

@dimidagd Just in case, we have a run-slow CI which we can run, which runs tests with a @slow marker and fits a 7B. If you write a corresponding ...IntegrationTest, we can run it on our large github runner and test out the result, just ping me and I'll run it

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, dinov3_vit

[dimidagd]

@molbap I uploaded the 7b weights with the linear classifier adapter on HF, and wrote a simple test on the cat COCO sample in the repo. Perhaps you could run the slow tests?