Latent Diffusion

To generate high-resolution image.

The noise predictor is trained in the latent space of AutoEncoder.

Earlier models used U-Net with the attention module at each layer for the noise prediction.