RADIOv25Options

Options & Configuration · AiDotNet.VisionLanguage.Encoders

Configuration options for RADIOv2.5, NVIDIA’s agglomerative vision foundation model.

Inheritance: ModelOptions → VisionEncoderOptions → RADIOv25Options

How It Works

RADIOv2.5 (Ranzinger et al., 2025) distills multiple teacher models (DINOv2, SAM, SigLIP, CLIP) into a single student model via multi-teacher distillation. The resulting model produces features compatible with all teacher models, serving as a universal vision backbone.

Constructors

Constructor Summary
RADIOv25Options() Initializes a new instance with default values.
RADIOv25Options(RADIOv25Options other) Initializes a new instance by copying from another instance.

Properties

Property Summary
bool UseTeacherSpecificHeads { get; set; } Gets or sets whether to produce teacher-specific output heads.
int AdapterDim { get; set; } Gets or sets the adapter dimension for teacher-specific heads.
int NumSummaryTokens { get; set; } Gets or sets the number of summary tokens for compact representation.
string[] TeacherModels { get; set; } Gets or sets the teacher models used for distillation.