[minimaxir]

The diffusion process is conditioned on CLIP text, which works better (in theory) since the encoded text is aligned with images.