Let's set up a predictable generation method where the model would usually reply with "The animal is a dog." However, we'll force the model to say either "The animal is a cat" or "The animal is a fish ...

We propose Model-guidance (MG) for training diffusion models, remove the commmonly used Classifier-free guidance (CFG), and achieve SOTA on ImageNet-256 conditional generation with FID=1.34. We list ...