T5TextConditioner<T>

Models & Types · AiDotNet.Diffusion.Conditioning

T5 text encoder conditioning module (Raffel et al., JMLR 2020). Used as the conditioning encoder for Stable Diffusion 3, FLUX.1, and Imagen pipelines. Pre-LN RMSNorm stack with learned relative position bias (paper-shared across all encoder layers).

Inheritance: NeuralNetworkBase<T> → TextConditioningBase<T> → T5TextConditioner<T>

Constructors

Constructor Summary
T5TextConditioner(ITokenizer tokenizer, NeuralNetworkArchitecture<T> architecture, T5TextConditionerOptions options)

Properties

Property Summary
bool ProducesPooledOutput { get; }

Methods

Method Summary
T5TextConditioner<T> FromPretrained(T5Variant variant, string huggingFaceModelName, string cacheDir) Loads a paper-canonical T5 conditioner with its real pretrained SentencePiece tokenizer from HuggingFace.
Tensor<T> GetPooledEmbedding(Tensor<T> sequenceEmbeddings) T5 pools by mean over non-pad tokens.