iiinova — where your digital self keeps evolving

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

… @timdarcet.bsky.social, Theo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, @jmairal.bsky.social, Herve Jegou, Patrick Labatut, Piotr Bojanowski And of course, it’s open source! github.com/facebookrese... 📜 Paper: ai.meta.com/research/pub...

github.com

GitHub - facebookresearch/dinov3: Reference PyTorch implementation and models for DINOv3

Reference PyTorch implementation and models for DINOv3 - facebookresearch/dinov3

020

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

Immensely proud to have been part of this project. Thank you to the team: @oriane_simeoni, @huyvvo, @baldassarrefe.bsky.social, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michael Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, …

110

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

And here’s my favorite figure from the paper, showing high resolution DINOv3 representations in all their detail-capturing glory ✨

110

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

To recap: 1) The promise of SSL is finally realized, enabling foundation models across domains 2) High quality dense features enabling SotA applications 3) A versatile family of models for diverse deploy scenarios So many great ideas (Gram anchoring!) to how we got there, please read the paper!

120

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

Satellite you said? Yes, the same DINOv3 algorithm trained on satellite imagery produces a SotA model for geospatial tasks like canopy height estimation. And of course, learns beautiful feature maps. This is the magic of SSL 🪄

110

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

3) DINOv3 is a family of models covering all use cases: • ViT-7B flagship model • ViT-S/S+/B/L/H+ (21M-840M params) • ConvNeXt variants for efficient inference • Text-aligned ViT-L (dino.txt) • ViT-L/7B for satellite All inheriting the great dense features of the 7B!

110

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

Well, Jianyuan Wang of VGGT fame simply dropped DINOv3 into his pipeline and off-handedly got a new SotA 3D model out. Seems promising enough?

120

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

But what do these great features bring us? We reached SotA on three long-standing vision tasks, simply by building on a frozen ❄️ (!) DINOv3 backbone: detection (66.1 mAP@COCO), segmentation (63 mIoU@ADE), depth (eg 4.3 ARel@NYU). Not convinced yet?

110

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

2) DINOv3’s global understanding is strong, but its dense representations truly shine! There’s a clear gap between DINOv3 and prior methods across many tasks. This matters as pretrained dense features power many applications: MLLMs, video&3D understanding, robotics, generative models, …

110

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

1) Some history: on ImageNet classification, supervised and weakly-supervised models converged to the same plateau over the last years. With DINOv3, SSL finally reaches that level. This alone is a big deal: no more reliance on annotated data!

120

Max Seitzer @maxseitzer.bsky.social · 14/08/2025

Introducing DINOv3 🦕🦕🦕 A SotA-enabling vision foundation model, trained with pure self-supervised learning (SSL) at scale. High quality dense features, combining unprecedented semantic and geometric scene understanding. Three reasons why this matters👇

2269

Reposted by Max Seitzer

Cansu Sancaktar @cansusancaktar.bsky.social · 14/07/2025

✨Introducing SENSEI✨ We bring semantically meaningful exploration to model-based RL using VLMs. With intrinsic rewards for novel yet useful behaviors, SENSEI showcases strong exploration in MiniHack, Pokémon Red & Robodesk. Accepted at ICML 2025🎉 Joint work with @cgumbsch.bsky.social 🧵

1205

Reposted by Max Seitzer

Mehdi S. M. Sajjadi @msajjadi.com · 10/07/2025

Scaling 4D Representations Self-supervised learning from video does scale! In our latest work, we scaled masked auto-encoding models to 22B params, boosting performance on pose estimation, tracking & more. Paper: arxiv.org/abs/2412.15212 Code & models: github.com/google-deepmind/representations4d

0208

Reposted by Max Seitzer

Georg Martius @gmartius.bsky.social · 04/04/2025

Introducing 3DGSim🧩— an end-to-end 3D physics simulator trained only on multi-view videos. It achieves spatial & temporal consistency w/o ground truth 3D info or heavy inductive biases— enabling scalability & generalization🚀 Kudos to Mikel + @andregeist.bsky.social www.youtube.com/watch?v=3Ar3...

mikel-zhobro.github.io

176