André Araujo

@toilaluan_ @GoogleDeepMind yep, our model is working well without it, so it wasn't something we particularly aimed to improve. But it's definitely worth exploring to potentially boost performance even more.

@andrefaraujo @GoogleDeepMind Curious why the model doesnt use rope? I see your model still behaves well in multi resolution without rope, but lots of work use it

True multimodal AI needs to understand the world spatially 🎯 🚀 Excited to release #CVPR2026 TIPSv2 from @GoogleDeepMind, a foundational image-text encoder with spatial awareness, leading to strong overall results and massive gains on patch-text alignment. 🔥 1/N

@tommiekerssies @GoogleDeepMind worth a try for sure! In this paper, our focus was on vision-language pretraining, very clearly this works great there, for both TIPS and CLIP/contrastive recipes.

@andrefaraujo @GoogleDeepMind Great work! Curious if supervising unmasked tokens could be used to fix the dense feature degradation seen in DINOv3 without the need for Gram anchoring.

@cataluna84 @BingyiCao @kmaninis @kfrancischen @arjunkarpur @ye_xia_ai @sahildua2305 @tanmayadabral @GuangxingHan @AlexBewleyAI @washingtonsk8 @kchorolab @mseyed @howardzzh Thanks for the invite, sounds interesting! Please reach out to us by email, we'd be happy to chat about it.

Congratulations on the awesome release! Would you or any of your co-authors like to present your papers in the Cohere Labs Computer Vision community? Cohere Labs Community Page: sites.google.com/cohere.com/coh… Here's the playlist to previous talks, if you are interested: youtube.com/playlist?list=…

@lasitolas @GoogleDeepMind we didn't focus on this in this work, but it's a great idea for an extension!

@andrefaraujo @GoogleDeepMind How about image and audio. ? For instance detecting an insect using audio and image

@gbiziel @GoogleDeepMind oh, we just didn't release ViT-S since for many cases ViT-B seemed small enough. But we could possibly release the ViT-S as well if there is excitement to use it in the community.

@andrefaraujo @GoogleDeepMind Why have you skipped distilling a small version? The results were much worse? I am curious from a mobile use case perspective.

@Hershal0_0 @GoogleDeepMind oh yeah, IMHO these are greatest research findings! A simple change that makes a huge diff in practice and anyone can quickly adopt.

@andrefaraujo @GoogleDeepMind the "few lines of code" fix that saves the whole architecture. peak engineering.

Huge thanks to amazing co-authors! @BingyiCao, Koert Chen, @kmaninis, @kfrancischen, @arjunkarpur, @ye_xia_ai, @sahildua2305, @tanmayadabral, @GuangxingHan, Bohyung Han, Joshua Ainslie, @AlexBewleyAI, Mithun Jacob, Rene Wagner, @washingtonsk8, @kchorolab, @mseyed, @howardzzh

Qualitatively, TIPSv2 PCA maps capture highly semantically focused details compared to DINOv3 – notice the distinct ceiling lamps in row 2, or how it easily contrasts the people from their backpacks in row 3! 13/N