SigLIPSOOptions
Options & Configuration · AiDotNet.VisionLanguage.Encoders
Configuration options for SigLIP-SO (Shape-Optimized SigLIP), a 400M vision encoder.
Inheritance: ModelOptions → VisionEncoderOptions → SigLIPSOOptions
How It Works
SigLIP-SO (Zhai et al., 2023) is a shape-optimized variant of SigLIP designed for use as a standalone vision encoder in VLMs. The SO-400M version (ViT-SO400M/14) uses an optimized width/depth ratio for the 400M parameter budget, producing high-quality visual features widely adopted in LLaVA, PaliGemma, and other VLMs.
Constructors
| Constructor | Summary |
|---|---|
SigLIPSOOptions() |
Initializes a new instance with default values. |
SigLIPSOOptions(SigLIPSOOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseSigmoidLoss { get; set; } |
Gets or sets whether to use sigmoid loss (SigLIP) for any fine-tuning. |
int NumOutputTokens { get; set; } |
Gets or sets the number of output feature tokens after pooling. |
int TrainingResolution { get; set; } |
Gets or sets the resolution at which the model was trained. |