DINO
In plain wordsA family of Meta vision models that learn to understand images from millions of unlabeled photos, by training one network to agree with a slowly updated copy of itself.
A self-supervisedself-supervised learning, n.: A way for AI to learn from huge amounts of unlabeled data by setting itself fill-in-the-blank puzzles, such as guessing a hidden word or a missing patch of an image. method, and the family of image encoders trained with it, based on self-distillation without labels. A student and a teacher network of the same architecture (usually a vision transformervision transformer, n.: A way of applying the transformer, the design behind chatbots, to pictures: the image is cut into small square tiles that the model reads like words in a sentence.) see different random crops of the same image: the teacher only large 'global' views, the student also small 'local' ones. The student is trained with cross-entropycross-entropy, n.: A score of how surprised a model's predictions are by what actually happens; lower is better. Language models are trained mainly by driving this number down on real text. to match the teacher's softmaxsoftmax, n.: A formula that turns a list of raw scores into probabilities — all positive and adding up to one — with the biggest share going to the highest score. output over a large number of dimensions (65,536 in the original), while the teacher receives no gradient and is updated as an exponential moving averageexponential moving average of weights, n.: A smoothed copy of a model kept during training, which each update nudges only slightly toward the latest weights. The calmer copy often performs better than the jittery original. of the student; centering and sharpening the teacher's outputs prevent collapse to a trivial constant answer. The 2021 paper reported 'emerging properties': attention maps of DINO-trained transformers outlined objects without supervision, and the features worked well for nearest-neighbor retrieval. DINOv2 (2023) added a curated dataset of 142 million images, a masked-patch objective and distillation from a 1-billion-parameter model; DINOv3 (August 2025) scaled to a 7-billion-parameter model trained on 1.7 billion images, with 'Gram anchoring' to keep dense per-patch features sharp. DINO models are widely used as frozen backbonesbackbone, n.: The main, usually pretrained part of a vision model that turns an image into useful internal features; smaller task-specific parts are attached to it, like limbs to a spine. for segmentation, depth estimationdepth estimation, n.: Working out how far away each point in a scene is from the camera, producing a 'depth map' that turns a flat picture into a sense of near and far., retrieval and robotics.
Frozen DINOv2 patch features plus a linear layer gave us usable segmentation masks for the drone imagery.
Origin
Introduced by Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski and Armand Joulin (Facebook AI Research and Inria) in 'Emerging Properties in Self-Supervised Vision Transformers' (arXiv April 2021; ICCV 2021); the name abbreviates 'self-distillation with no labels'. DINOv2 (Maxime Oquab et al.) followed in April 2023 and DINOv3 (Oriane Siméoni et al.) in August 2025.
Usage note
Not to be confused with the DINO object detector (2022), a DETRDETR, prop. n.: An object detector built on the transformer, the design behind modern chatbots, that outputs a short list of objects in one go instead of sorting through thousands of guessed boxes. variant named for 'DETR with Improved deNoising anchOr boxes', on which Grounding DINO is built.
Related
- self-supervised learningself-supervised learning, n.: A way for AI to learn from huge amounts of unlabeled data by setting itself fill-in-the-blank puzzles, such as guessing a hidden word or a missing patch of an image.
- knowledge distillationknowledge distillation, n.: Training a smaller or cheaper 'student' AI to copy the behavior of a larger 'teacher' AI, so the student keeps much of the teacher's skill while being faster to run.
- vision transformervision transformer, n.: A way of applying the transformer, the design behind chatbots, to pictures: the image is cut into small square tiles that the model reads like words in a sentence.
- masked autoencodermasked autoencoder, n.: A way to teach a vision model without labels: hide most of an image's tiles and train the model to fill in the missing parts, so that it learns what images are made of.
- joint-embedding predictive architecturejoint-embedding predictive architecture, n.: A way of training AI on images or video in which the model predicts an abstract summary of a hidden part, rather than every pixel, so it learns what matters and ignores unpredictable detail.
- backbonebackbone, n.: The main, usually pretrained part of a vision model that turns an image into useful internal features; smaller task-specific parts are attached to it, like limbs to a spine.
- contrastive learningcontrastive learning, n.: A way to learn without labels by comparison: the model learns to pull together two versions of the same thing, such as two crops of one photo, and push apart unrelated things.
Sources
- Caron, M. et al. (2021). Emerging Properties in Self-Supervised Vision Transformers. ICCV 2021. arxiv.org/abs/2104.14294
- Oquab, M. et al. (2023). DINOv2: Learning Robust Visual Features without Supervision. arXiv:2304.07193. arxiv.org/abs/2304.07193
- Siméoni, O. et al. (2025). DINOv3. arXiv:2508.10104. arxiv.org/abs/2508.10104
Mentioned in
Meta AIMeta AI, prop. n.: The artificial intelligence research group of Meta, the company behind Facebook and Instagram, known for the PyTorch software and the Llama models. It is also the name of Meta's chatbot., backbonebackbone, n.: The main, usually pretrained part of a vision model that turns an image into useful internal features; smaller task-specific parts are attached to it, like limbs to a spine., contrastive learningcontrastive learning, n.: A way to learn without labels by comparison: the model learns to pull together two versions of the same thing, such as two crops of one photo, and push apart unrelated things., feature extractionfeature extraction, n.: Turning raw data such as pixels or sound into a short list of telling numbers — computed by hand-designed rules or by an AI system trained earlier — that a new model can learn from or compare., self-supervised learningself-supervised learning, n.: A way for AI to learn from huge amounts of unlabeled data by setting itself fill-in-the-blank puzzles, such as guessing a hidden word or a missing patch of an image., semantic segmentationsemantic segmentation, n.: Coloring every pixel of an image by what it shows, such as road, sky, person or car, without telling one person or car apart from another., vision transformervision transformer, n.: A way of applying the transformer, the design behind chatbots, to pictures: the image is cut into small square tiles that the model reads like words in a sentence., DETRDETR, prop. n.: An object detector built on the transformer, the design behind modern chatbots, that outputs a short list of objects in one go instead of sorting through thousands of guessed boxes., joint-embedding predictive architecturejoint-embedding predictive architecture, n.: A way of training AI on images or video in which the model predicts an abstract summary of a hidden part, rather than every pixel, so it learns what matters and ignores unpredictable detail., masked autoencodermasked autoencoder, n.: A way to teach a vision model without labels: hide most of an image's tiles and train the model to fill in the missing parts, so that it learns what images are made of.