DINOv3<T>
Models & Types · AiDotNet.VisionLanguage.Encoders
DINOv3 self-supervised vision encoder scaled to 7B parameters on 1.7B images.
Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → DINOv3<T>
Implements: IVisualEncoder<T>
For Beginners
DINOv3 from Meta scales self-supervised vision training to 7 billion parameters on 1.7 billion images, outperforming even contrastive models like SigLIP 2 on most vision benchmarks. It uses improved training recipes with SwiGLU feed- forward networks and enhanced data augmentation, producing the strongest open-source self-supervised vision encoder available. Default values follow the original paper settings.
How It Works
DINOv3 (Meta, 2025) scales self-supervised ViT training to 7B parameters on 1.7B images, outperforming SigLIP 2 on most vision benchmarks. It uses improved training recipes with SwiGLU FFN, enhanced data augmentation, and longer training schedules.
References:
- Paper: “DINOv3: Self-Supervised Vision at Scale” (Meta, 2025)
Example
// Create a DINOv3 model for 7B-scale self-supervised visual features
// outperforming SigLIP 2 on most vision benchmarks
var architecture = new NeuralNetworkArchitecture<double>(
inputType: InputType.TwoDimensional,
taskType: NeuralNetworkTaskType.ImageClassification,
inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
// ONNX inference mode with pre-trained model
var model = new DINOv3<double>(architecture, "dinov3.onnx");
// Training mode with native layers
var trainModel = new DINOv3<double>(architecture, new DINOv3Options());
Constructors
| Constructor | Summary |
|---|---|
DINOv3(NeuralNetworkArchitecture<T> architecture, DINOv3Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
|
DINOv3(NeuralNetworkArchitecture<T> architecture, string modelPath, DINOv3Options options) |
Properties
| Property | Summary |
|---|---|
int EmbeddingDimension { get; } |
Methods
| Method | Summary |
|---|---|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> EncodeImage(Tensor<T> image) |