First Gemini release with a piece of my work inside 😄 and all the countless other amazing people 🚀🚀🚀

🚀 Excited to share our new work RefAM: Attention Magnets for Zero-Shot Referral Segmentation, a training-free approach that turns diffusion model attentions into segmentations. By @anna_kukleva_, me, Alessio Tonioni, @ferjadnaeem, @fedassa, @janericlenssen, Bernt Schiele

🎺Meet VIST3A — Text-to-3D by Stitching a Multi-view Reconstruction Network to a Video Generator. ➡️ Paper: https://t.co/sFqbbUiGOO ➡️ Website: https://t.co/QWMLwXyVcB Collaboration between ETH & Google with Hyojun Go, @DNarnhofer, Goutam Bhat, @fedassa, and Konrad Schindler.

Oguzhan is an amazing mentor to work with, apply if you are on the internship market

A big congratulations to the whole Gemini team on pushing this amazing family of models out 😄 Our tech report is out now: https://t.co/5FfTM1LEdN Feels a bit unreal to share the contributors list with all the amazing colleagues

Active Data Curation Effectively Distills Large-Scale Multimodal Models - compute per sample loss with large batch - only backprop (probabilistically) through samples with high loss intuition: these are the samples where there is “something to learn” - if both teacher and

Stop by this amazing work from Vishaal and the team today at CVPR

timm's got a new vision transformer (NaFlexVit), and it's flexible! I've been plugging away at this for a bit, integrating ideas from FlexiViT, NaViT, and NaFlex and finally ready to merge for initial exploration. The model supports: * variable aspect/size images of NaFlex (see

At #GoogleIO, we shared how decades of AI research have now become reality.  From a total reimagining of Search to Agent Mode, Veo 3 and more, Gemini season will be the most exciting era of AI yet.  Some highlights 🧵

📢 We just released the code for JetFormer at https://t.co/Wgiz3tK9S8 Enjoy!

We are presenting JetFormer at ICLR this morning, poster #190. Stop by if you’re interested in unified multimodal architectures!

Google's global PhD Fellowship program will open for applications this week! (on Apr 10th) This supports PhD students in computer science and related fields, also connecting to a Google mentor. Learn more and apply at: https://t.co/ynVQDf5xLi (deadline: May 15th, 2025)

Check out the strongest open-source dense prediction models from our colleagues!

The majority of features in this layer of Siglip-2 are multimodal. I'd expected some multimodality but was surprised that two-thirds of the neurons I tested bind together their visual and linguistic features. This neuron fires for images of mustaches and for the word "mustache"

Fully supportive of this. Machine Learning/ Computer Vision review process is broken with irresponsible reviewers. Glad to see there is some accountability.

Delighted to share that ACED has been accepted at CVPR2025! Check out our work to know how to distill the strongest smol size image-text contrastive models.

📢2⃣ Yesterday we released SigLIP 2! TL;DR: Improved high-level semantics, localization, dense features, and multilingual capabilities via drop-in replacement for v1. Bonus: Variants supporting native aspect and variable sequence length. A thread with interesting resources👇

Excited to share what we have been up to in image text embedding models. SigLIP 2 is the most powerful encoder for most open vocabulary computer vision and MMLLM tasks. Checkpoints are open sourced and we look forward to what the community achieves with these.

Excited to see our paper "Tokenformer: Rethinking transformer scaling with tokenized model parameters" accepted as a spotlight at #ICLR2025 ! Hope our idea of ​​tokenizing everything can inspire the future of AI. Paper: https://t.co/0ofQpsudSH Code: https://t.co/D3rhZzqMwD

Ever thought of training multimodal models with 100 billion 🚀 unique examples? Check out WebLI-100B! The study reveals exciting insights on long-tail tasks, including multilingual and cultural diversity benchmarks. Paper: https://t.co/npNrvPGY53