UpBlock<T>

Models & Types · AiDotNet.Diffusion.VAE

Upsampling block for VAE decoder with transposed convolution and multiple ResBlocks.

Inheritance: LayerBase<T> → UpBlock<T>

Implements: IShapeContract

For Beginners

An UpBlock is like a decompression stage in a decoder.

What it does:

  1. Increases spatial size by 2x using transposed convolution (decompression)
  2. Processes the upsampled features through multiple residual blocks

Example: 8x8 input -> 16x16 output (spatial dimensions doubled)

Why use transposed convolution instead of simple interpolation?

  • Transposed conv is learnable (the network decides how to upsample)
  • Simple interpolation (bilinear, nearest) has fixed behavior
  • Learnable upsampling can generate sharper details

Structure:

input [B, C_in, H, W]
│
├─→ ConvTranspose (stride=2) ─→ upsample
│
↓
[B, C_out, 2*H, 2*W]
│
├─→ ResBlock → ResBlock → ... (numLayers blocks)
│
↓
output [B, C_out, 2*H, 2*W]

How It Works

This implements an upsampling block following the Stable Diffusion VAE architecture:

  • Transposed convolution (deconvolution) to increase spatial dimensions by 2x
  • Multiple VAEResBlocks to process features at the upsampled resolution

Constructors

Constructor Summary
UpBlock(int inChannels, int outChannels, int numLayers, int numGroups, int inputSpatialSize, bool hasUpsample) Initializes a new instance of the UpBlock class.

Properties

Property Summary
bool HasUpsample { get; } Gets whether this block performs upsampling.
bool SupportsTraining { get; }
int InputChannels { get; } Gets the number of input channels.
int NumLayers { get; } Gets the number of residual blocks.
int OutputChannels { get; } Gets the number of output channels.

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
IReadOnlyList<VAEResBlock<T>> GetResBlocks() Gets the residual blocks for external access.
void ResetState() Resets the internal state of the block.
void UpdateParameters(T learningRate) Updates all learnable parameters using gradient descent.