Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling

Authors: Xingyu Zheng, Xianglong Liu, Yifu Ding, Weilun Feng, Junqing Lin, Jinyang Guo, Haotong Qin

arXiv ID: 2607.01642

Problem: Multi-resolution acceleration strategies for text-to-image diffusion models suffer from blurring and artifacts when performing latent-space upsampling with selective partial-region modification.

Key Methodology:

MrFlow is a training-free, staged low-to-high-resolution pipeline for pretrained flow-matching models (FLUX.1-dev, Qwen-Image). It first generates the main structure at low resolution, then performs super-resolution in pixel space via a lightweight pretrained GAN.
Low-strength noise is injected after upsampling to enable high-frequency resampling, followed by final high-resolution detail refinement.
Exploits quadratic token reduction and reduced step count from low-resolution sampling - no training or runtime dynamic identification required.

Key Results:

Achieves 10× end-to-end acceleration while keeping OneIG within a 1% gap of the unaccelerated baseline, significantly outperforming other training-free acceleration strategies.
When combined orthogonally with pretrained timestep distillation, reaches up to 25× generation acceleration.

Applied Context: Builders can drop MrFlow into existing FLUX.1-dev or Qwen-Image pipelines for order-of-magnitude speedups with negligible quality loss, no training, and no custom kernels - making high-quality diffusion viable for latency-sensitive or resource-constrained deployment.