When diffusion models are used, the inference time could be meaningful. But then this is only 64x64 with upsampling, so probably not too bad.