Albis: Noise Scheduling for Diffusion

Training a diffusion model means deciding how much effort each noise level gets. This noise schedule, set by noise sampling and loss weighting, strongly affects sample quality. Yet what makes a good schedule is hard to tell, because DDPM, EDM, and flow matching each bundle it with other training choices. We show that once these choices are aligned, the large differences between the frameworks mostly disappear. So how should we design the noise schedule? Permanently favoring low or high noise levels gives no consistent gain, but shifting the emphasis during training does. This raises the question of how much it matters when noise levels are emphasized, compared to which. We study this question with Albis, a deliberately simple scheduler that shifts samples toward noise levels where the training loss is still high. With sufficient training, Albis reaches a lower loss than fixed logit-normal sampling at every noise level and improves sample quality on CIFAR-10 and CelebA-HQ subsets and ImageNet. When noise levels are emphasized therefore matters as much as which levels are emphasized.