Single-Step Diffusion via Direct Models

Hamadi Chihaoui, Paolo Favaro

Keywords: Efficient generative models

Abstract: We introduce Direct Models, a generative modeling framework that enables single-step diffusion by learning a direct mapping from initial noise $x_0$ to all intermediate latent states along the generative trajectory. Unlike traditional diffusion models that rely on iterative denoising or integration, Direct Models leverages a progressive learning scheme where the mapping from $x_0$ to $x_{t + \delta t}$ is composed as an update from $x_0$ to $x_t$ plus the velocity at time $t$. This formulation allows the model to learn the entire trajectory in a recursive, data-consistent manner while maintaining computational efficiency. At inference, the full generative path can be obtained in a single forward pass. Experimentally, we show that Direct Models achieves state-of-the-art sample quality among single-step diffusion methods while significantly reducing inference time.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 24752