VideoGigaGANGeneratorLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

Native VideoGigaGAN generator with a style-modulated spatial backbone, temporal inflation, anti-aliased flow-guided propagation, and a high-frequency shuttle.

Inheritance: LayerBase<T> → VideoGigaGANGeneratorLayer<T>

Implements: IShapeContract

How It Works

This layer follows Xu et al., “VideoGigaGAN: Towards Detail-rich Video Super-Resolution” (CVPR 2024). A GigaGAN-style residual generator is inflated with temporal convolutions, SPyNet aligns recurrent features in both directions, a low-pass filter is applied before every warp to suppress temporal aliasing, and a parallel high-frequency path injects genuine input detail at each progressive 2x upsampling stage.

Input is [frames, channels, height, width] or [batch, frames, channels, height, width]. Output preserves the leading dimensions and enlarges height and width by the configured scale factor. Every transform uses engine or registered-layer operations so the complete recurrent generator remains on the gradient tape.

Constructors

Constructor Summary
VideoGigaGANGeneratorLayer(int inputChannels, int inputHeight, int inputWidth, int numFeatures, int numResBlocks, int numStyleLayers, int scaleFactor, int flowPyramidLevels, double hfShuttleWeight) Creates the native VideoGigaGAN generator graph.

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
void ClearGradients()
void ResetState()
void UpdateParameters(T learningRate)