A Comprehensive Evaluation of Timestep Discretization Strategies in Text-to-Image Diffusion Models

Abstract

Text-to-image latent diffusion models produce unprecedented visual fidelity but remain severely bottlenecked by the computational latency of iterative sampling. While optimizing the discretization of the continuous-time variable offers a powerful, training-free acceleration pathway, the comparative tradeoffs of foundational spacing heuristics remain under-explored. In this paper, we systematically benchmark three primary timestep spacing strategies—leading, linspace, and trailing—across a broad spectrum of Number of Function Evaluations (NFEs) to explicitly isolate their impact on generation quality within the text-to-image task. Through extensive evaluations on MS-COCO assessing structural fidelity (FID, IS) and text-image semantic alignment (CLIP Score), we identify distinct solver-dependent behaviors. With first-order ODE solvers such as DDIM, linspace yields stable, marginally superior outcomes for high-step generation but suffers from noisy outputs under low-step constraints. Conversely, trailing significantly mitigates this low-step performance collapse, while leading consistently degrades image quality and restricts the brightness range. However, with high-order ODE solvers, exemplified by DEIS, the performance gap between linspace and trailing diminishes, though leading remains strictly inferior. Ultimately, our findings suggest leveraging trailing as a highly versatile timestep spacing strategy, particularly for optimizing fast generation in latency-sensitive applications.