Diffusion noise-prediction loss mathematical notation

Diffusion noise-prediction loss is a recurring research-paper notation family. Trains a denoiser to recover the Gaussian noise that produced a noisy sample at a selected diffusion timestep.

Diffusion noise-prediction loss: Trains a denoiser to recover the Gaussian noise that produced a noisy sample at a selected diffusion timestep. Example: Compare the added noise with the network's estimate at every sampled noise level.