The current semantic segmentation model for ControlNet provides a list of usable materials, but they are not specific enough for architectural purposes. By generating new tokens (colours) that extend the existing model with new material types, we aim to improve the process by which stable diffusion operates and make it more suitable for architecture applications.
