49ad23d1ec9fa4bd8d77d02681df5cfa-Supplemental.pdf

Compute isessential tomodern machine learning applications, andmorecompute typically yields better results. It is thus important to compare our method's compute requirements to competing methods. Table 10: Training compute requirements for our diffusion models compared to StyleGAN2 and BigGAN-deep. Underreasonablesettingsforβt andT,thedistribution q(xT) is nearly an isotropic Gaussian distribution, so samplingxT is trivial. In particular, they do not directly parameterizeµθ(xt,t) as a neural network,butinsteadtrainamodel ϵθ(xt,t)topredictϵfromEquation3.