Consider the output of a diffusion model, which should represent a bunch of nearly one-hot encoded vectors. If I wanted to bias the generation to make certain classes of tiles more likely, I could simply add some constant value to that channel of the tensor output, thus making that tile type more likely. This could be a useful knob at inference time. Worth considering.
Consider the output of a diffusion model, which should represent a bunch of nearly one-hot encoded vectors. If I wanted to bias the generation to make certain classes of tiles more likely, I could simply add some constant value to that channel of the tensor output, thus making that tile type more likely. This could be a useful knob at inference time. Worth considering.