Diffusion models are also very reliant on high-quality training data input, very much a case of garbage in, garbage out.