Channel Painters in Action

[Update 1/18/24] Visit Latent Painter for a text-2-image generative AI painter driven by Stable Diffusion

The channel painter decomposes a convolutional layer of the deep neural network, and turns the input image into a painting animation. It works on top of the CNN-based generative networks, including the style transformers.

Run the demo in Colab

In different styles

On different generative networks (very channel-rich at the operation layer)

Moving the operation layer down along the forward path

Toward the final layers (with less channels)

On different painting scenarios

On different photos/objects

Write MNIST digits over VAE

More Detail?

Channel Decomposition into Painting Actions