Example Built with Kernel FX Ships with Voro as an example. It may change between versions.

Purpose

Latent Compose adds compatible latents together. Each input has a weight. Variance Preserving mode compensates for the combined weight magnitude; Linear leaves the weighted sum unnormalized.

Use

Place Latent Compose under Generate in the Image view. Set Inputs to the number of latents you want to combine, then connect matching Latent Seed outputs.

Start with two different seeds and animate their weights. Send out to a generator configured for the same profile, or through Latent Transform for further shaping.

Inputs and outputs

• latent1 is always present. Selecting more Inputs adds latent2, latent3 and latent4.
• out carries the combined latent with the same profile and dimensions.

The initial one-input declaration is shown below. Inputs changes both the ports and the weight controls.

Controls

• Inputs selects one, two, three or four connections.
• Profile is StyleGAN2 Z with 512 components.
• Weight controls each input’s contribution. Negative weights subtract.
• Mix selects Variance Preserving or Linear.

Notes

Variance Preserving divides the weighted sum by the square root of the sum of squared weights when that value is nonzero. It keeps the scale steadier when combining independent seeds. It does not guarantee that every mixture will produce an equally useful image.

Ports

Inputs

• latent1 image required
  1 x 512 x 4

Outputs

• out image
  1 x 512 x 4

Parameters

Latent Compose

Inputs (Inputs) Menu op('latent_compose').par.Inputs

Selects a frozen program and rebuilds this operator's parameters and exact port surface.

Default:

one

Options:

1 Input, 2 Inputs, 3 Inputs, 4 Inputs