Classifier-Free Diffusion Guidance

Train a conditional and an unconditional diffusion model on the same data. Subtract the conditional from the unconditional score vector to get a guidance vector, which can be added to the unconditional after multiplying with a desired strength coefficient.