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

MrFlow accelerates text-to-image diffusion by combining low-resolution generation with pixel-space super-resolution and noise injection, achieving up to 25x speedup without training or runtime modifications. This approach exploits quadratic token reduction and reduced step requirements for low-resolution sampling. MrFlow can be combined with pre-trained timestep distillation strategies for further acceleration.

1sec.ai