Large Vision Models (LVM)
Large vision models are vision networks with hundreds of millions to billions of parameters, pretrained on huge image collections to serve many tasks.
A large vision model (LVM) is a neural network for images and video with hundreds of millions to billions of parameters, pretrained on millions to billions of images so that one model serves many visual tasks. It is the visual counterpart of a large language model and a kind of foundation model: pretraining is mostly self-supervised or weakly supervised, and the resulting model is adapted to classification, detection, segmentation, depth, or retrieval with little or no labeled data. Most LVMs are vision transformers, whose accuracy keeps improving with model size and data, following the same scaling laws seen in language models. Well-known examples include Google's ViT-22B with 22 billion parameters, Meta's DINOv2 and DINOv3, and Meta's Segment Anything Model.
How Large Vision Models Are Trained#
Labels do not scale to billions of images, so LVMs learn mainly from the images themselves or from loosely paired text. DINOv2, released by Meta in 2023, uses self-distillation, training a student network to match a teacher's outputs on different crops of the same image; its largest model has 1.1 billion parameters and was trained on 142 million curated images, and DINOv3 scaled the recipe in 2025 to 7 billion parameters and 1.7 billion images. Masked autoencoders take another route to self-supervised learning, hiding most of an image's patches and training the model to reconstruct them. Image-text contrastive learning, as in CLIP and SigLIP, trains an image encoder to match captions and produces vision backbones that many multimodal models reuse. Meta's Segment Anything Model learned promptable segmentation from the SA-1B dataset, with 1.1 billion masks on 11 million images.
The model that gave the term its most literal use, LVM by Bai and colleagues at UC Berkeley and Johns Hopkins in 2023, converts images, video frames, and annotations such as segmentation masks into "visual sentences" of discrete tokens and trains on 420 billion of them with next-token prediction, without any text. At test time it solves tasks through visual prompting: given a few example input-output image pairs, it continues the sequence with the output for a new image.
Adapting LVMs to Tasks#
A pretrained LVM is rarely used as is. Without labeled data, the model works zero-shot through its own prompts, such as clicks for SAM, or through image embeddings compared for retrieval and deduplication. With some labels, linear probing trains a small head on frozen features, which is how DINOv2 and DINOv3 features are typically evaluated on depth estimation and segmentation, or fine-tuning updates the backbone itself. When the target is a phone, camera, or robot, the LVM instead labels data or teaches a compact model through knowledge distillation.
The Ultralytics auto_annotate function takes that last path: it pairs a YOLO26 detector with Meta's SAM to turn detected boxes into segmentation labels for training a smaller segmentation model. SAM-b has 93.7 million parameters and YOLO26n-seg has 2.7 million, about 35 times fewer. Ultralytics Platform applies the same idea to detection datasets, where class-prompted vision models label the images that a YOLO26 model then trains on.
LVMs vs. VLMs and Task-Specific Models#
A vision language model connects a vision encoder, often an LVM, to a language model so it can answer questions about images in text; an LVM on its own outputs features, masks, or images. Task-specific models such as object detection networks are trained for one task and a fixed set of classes, which keeps them small and fast enough for real-time inference on edge AI devices. Large vision models trade that efficiency for breadth: a single DINOv3 backbone, kept frozen, was evaluated on classification, detection, segmentation, depth estimation, and 3D correspondence, and Meta also trained a version on satellite imagery for satellite image analysis tasks such as canopy height mapping.










