Despite the promising performance of diffusion models on continuous modality generation, one deficiency that is holding them back is their requirement for multi-step denoising processes, which can be computationally expensive. In this article, we examine recent works that aim to build diffusion models capable of performing sampling in one or a few steps.

Multi-modal and multi-function Transformers enables a single architecture to process diverse data types such as language, images, and videos simultaneously. These models employ techniques like vector quantization and lookup-free quantization to map different modalities into a unified embedding space, allowing the Transformer to handle them within the same sequence. Beyond processing multiple data types, these architectures can also combine different functionalities-such as auto-regressive language generation and diffusion-based image creation-within a single model.