I suspect the trick only works because DeepFloyd-IF operates in pixel space while other diffusion models operate in the latent space. > Our method uses DeepFlo

I suspect the trick only works because DeepFloyd-IF operates in pixel space while other diffusion models operate in the latent space. > Our method uses DeepFloyd IF, a pixel-based diffusion model. We do not use Stable Diffusion because latent diffusion models cause artifacts in illusions (see our paper for more details).