DiffusionCrossAttention<T>
Models & Types · AiDotNet.Diffusion.Attention
Cross-attention layer for diffusion models with Flash Attention optimization.
Inheritance: LayerBase<T> → DiffusionCrossAttention<T>
Implements: IShapeContract
For Beginners
Cross-attention is how the model “looks at” text while generating images.
- Query (Q): Comes from the image features
- Key (K) and Value (V): Come from text embeddings
- Output: Image features enriched with text information
This enables the model to generate images that match the text description.
How It Works
Cross-attention allows the model to attend to conditioning information (like text embeddings) when generating images. This is how text-to-image models like Stable Diffusion work.
Constructors
| Constructor | Summary |
|---|---|
DiffusionCrossAttention(int queryDim, int contextDim, int numHeads, int spatialSize) |
Initializes a new diffusion cross-attention layer. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
|
int ContextDim { get; } |
Gets the context dimension. |
int QueryDim { get; } |
Gets the query dimension. |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<ILayer<T>> GetSubLayers() |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
Derived from this type’s [TensorLayout] declarations. |
Tensor<T> ForwardWithContext(Tensor<T> input, Tensor<T> context) |
Performs the forward pass with context (conditioning). |
void ResetState() |
Resets the layer’s internal state. |
void SetTrainingMode(bool isTraining) |
Propagates eval/training mode to the nested cross-attention sublayer so inference uses the GPU-resident fast path rather than the allocating tape/training branch. |
void UpdateParameters(T learningRate) |
Updates parameters. |