MultiViewAttention<T>
Models & Types · AiDotNet.Diffusion.ThreeD
Multi-view attention module for cross-view consistency.
How It Works
MVDream (Shi et al., 2023) makes the views consistent by letting their tokens attend to one another. Here every
spatial position attends across the views with learned query/key/value/output projections of the channel
vectors it receives: x + W_o softmax(Q K^T / sqrt(C)) V, computed per position over the view axis.
Constructors
| Constructor | Summary |
|---|---|
MultiViewAttention(int channels, int numViews, int? seed) |
Creates multi-view attention. |
Properties
| Property | Summary |
|---|---|
long ParameterCount { get; } |
Gets parameter count. |
Methods
| Method | Summary |
|---|---|
Tensor<T>[] Apply(Tensor<T>[] views) |
Applies multi-view attention for cross-view consistency. |
Vector<T> GetParameters() |
Gets parameters. |
void SetParameters(Vector<T> parameters) |
Sets parameters. |