One Step Diffusion via Flow Fitting

Hamadi Chihaoui, Paolo Favaro

Keywords: Efficient generative models, Single step diffusion

Abstract: Diffusion and flow-matching models have demonstrated impressive performance in generating diverse, high-fidelity images by learning transformations from noise to data. However, their reliance on multi-step sampling requires repeated neural network evaluations, leading to high computational cost. We propose FlowFit, a family of generative models that enables high-quality sample generation through both single-phase training and single-step inference. FlowFit learns to approximate the continuous flow trajectory between latent noise \(x_0\) and data \(x_1\) by fitting a basis of functions parameterized over time \(t \in [0, 1]\) during training. At inference time, sampling is performed by simply evaluating the flow only at the terminal time \(t = 1\), avoiding iterative denoising or numerical integration. Empirically, FlowFit outperforms prior diffusion-based single-phase training methods achieving superior sample quality.

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

Submission Number: 24339