Zheng et al. propose Multi-Resolution Flow Matching, a training‑free technique that accelerates text‑to‑image diffusion models by staging sampling across multiple resolutions. Their method improves on existing hardware‑agnostic approaches, such as timestep distillation and feature caching, by addressing latent‑space upsampling and selective region modification, aiming to surpass the typical 5× speedup without requiring custom kernels or system‑level tweaks.