it uses a new technique called "consistency" that lets latent diffusion models to predict images in much fewer steps.

some links here: - https://arxiv.org/abs/2310.04378 - https://arxiv.org/abs/2311.05556