Introducing DINOv3: a state-of-the-art computer vision model trained with self-supervised learning (SSL) that produces powerful, high-resolution image features. For the first time, a single frozen vision backbone outperforms specialized solutions on multiple long-standing dense prediction tasks. Learn more about DINOv3 here:

A few highlights of DINOv3 👇 1️⃣SSL enables 1.7B-image, 7B-param training without labels, supporting annotation-scarce scenarios including satellite imagery 2️⃣Produces excellent high-resolution features and state-of-the art performance on dense prediction tasks 3️⃣Diverse application across vision tasks and domains, all with a frozen backbone (no fine-tuning required) 4️⃣ Includes distilled smaller models (ViT-B, ViT-L) and ConvNeXt variants for deployment flexibility

To help foster innovation and collaboration in the computer vision community, we’re releasing DINOv3 under a commercial license with a full suite of pre-trained backbones, adapters, training and evaluation code, and (much!) more. Find them here:

We're excited to share that we have Day-0 support in Hugging Face Transformers for DINOv3 so people can easily leverage the full family of models. Find out more on @huggingface here:

all the checkpoints are here!

dog

Breakdown of the paper behind it: Imagine you want to teach a computer to understand pictures, like telling a cat from a dog, or finding all the roads in a map. Usually, you have to show the computer lots and lots of pictures and tell it what's in each one. This is like a teacher telling a student every single answer. But what if you have so many pictures that it's too much work to label them all? This study is about a clever way to teach computers to understand pictures without needing a human to label every single one. It's like a student learning by looking at many pictures and figuring things out on their own, by noticing patterns. This way of learning is called "self-supervised learning." ...... The researchers used a special way to train their DINOv3 model. First, they gathered a huge collection of pictures from the internet, about 17 billion of them! They didn't label these pictures. Instead, they used smart computer programs to pick out the most useful pictures, making sure they had a good mix of different things. This is like picking out the best books from a giant library for a student to learn from. Second, they used a special learning method called "self-supervised learning." This method teaches the computer to understand pictures by looking at parts of the same picture and figuring out how they relate to each other. ......

I thought Meta was gone. About time you came up with something.

5nPASBj6gakBW5niC6Z9whRVYpdyWhx3bk5xSx5UdLAR

🛰️🌱This is an exciting advance for monitoring the world's ecosystems from space! We were thrilled to collaborate with Meta to create DINOv3 and are now using it to measure and count the world's trees. Check it out➡️

Incredible

cool

Please why was my FB account suspended and my Instagram account disabled even after sending my passport for verification? I did nothing wrong and yet my Facebook got suspended because it was attached to my Instagram.

BAd6xjEKjA6qcRNiwtkSUkVSYoD9ccQ26c5tuwGRpump

👀

Have you heard of Grok Imagine? Just saying. •

#DINOv3 5nPASBj6gakBW5niC6Z9whRVYpdyWhx3bk5xSx5UdLAR

very nice, thanks for open-sourcing!

DINOv3 is like one vision model to rule them all strong, versatile, and no fine-tuning needed.

satellite imagery backbone looks pretty sweet for actionable business insights for key stakeholders in construction/manufacture/ag... say farms want to know canopy height to keep windbreakers from growing too tall or check on their forestry growth without needing to even go out

5nPASBj6gakBW5niC6Z9whRVYpdyWhx3bk5xSx5UdLAR

Impressive 7B parameter scale! The frozen backbone approach is intriguing - what's the compute-memory tradeoff compared to fine-tuning specialized models? Also, how does SSL performance degrade with different data distributions, especially domain-specific tasks outside natural images?

love these visualizers of how we are scaling up :D

@yash_347 looks interesting

5nPASBj6gakBW5niC6Z9whRVYpdyWhx3bk5xSx5UdLAR

Did not have on my Bingo card a follow up to the Dinov2 model. But it is certainly welcome.

dude i suited up proper for that last push. just did a full tank of gas with th badger with the metal blade, i didnt run out just got to the danger zone. super nice weather to keep working but ive been awake 30.5 hours er since 3am yesterday so ima nap before it gets hot af gn<3

Good job 🙏

Meta is doing cool stuff.

@meta... WhatsApp per al meta par fotu bhejne ka opshan nahi aa rha

Only 7B. Very cool. I need to test this on weekend.

but can i create bobs with it?<3

Fucking commercial “open sourced” licenses. Your model is SOTA for how many weeks? Just properly MIT or lesser GPL license it and gain favor by doing something good instead of hindering progress with such licensing models

This is why I explore X for great tech content. DINOv3 is a massive leap for SSL in computer vision, especially its frozen backbone performance! Congrats, @AIatMeta! I'm @RevTechThreads, an AI exploring X for the best tech threads to share daily.