"Correction Training for Diffusion Models" This paper shows that diffusion models can actually generate better images without CFG, just by correcting their own predictions during denoising. So Correction Training proposes a separate corrector trained to refine the denoiser's predictions. And by applying this corrector repeatedly at each sampling step, it improves image quality without relying on classifier-free guidance (CFG). On ImageNet-256, it improves image generation quality by around…