CompositeConditioningBase<T>

Base Classes · AiDotNet.Diffusion.Conditioning

Base class for conditioning modules that compose other conditioners rather than owning their own learnable weights. Examples: DualTextConditioner and TripleTextConditioner, which delegate encoding work to one or more inner CLIP / T5 / OpenCLIP encoders and then merge their outputs.

Implements: IConditioningModule<T>

Known derived types: TripleTextConditioner<T>

How It Works

This base intentionally does not allocate token / position / transformer weights (unlike TextConditioningBase, whose constructor sizes weight vectors against a single transformer). Composite conditioners hold their inner conditioners as fields and forward calls to them, so all that is shared between composites is the boilerplate every conditioner needs:

The Engine property is exposed as a protected instance member to match the convention used by the rest of AiDotNet’s base classes (see NeuralNetworkBase, AdversarialAttackBase, AugmentationBase, etc.). Subclasses MUST route all engine-dispatched tensor operations through this property rather than calling AiDotNetEngine.Current directly, so that future per-instance engine overrides remain a single-point change.

Properties

Property Summary
ConditioningType ConditioningType { get; }
bool ProducesPooledOutput { get; }
int EmbeddingDimension { get; }
int MaxSequenceLength { get; }

Methods

Method Summary
Tensor<T> Encode(Tensor<T> input)
Tensor<T> EncodeText(Tensor<T> tokenIds, Tensor<T> attentionMask)
Tensor<T> GetPooledEmbedding(Tensor<T> sequenceEmbeddings)
Tensor<T> GetUnconditionalEmbedding(int batchSize)
Tensor<T> Tokenize(string text)
Tensor<T> TokenizeBatch(string[] texts)