Researchers propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models that uses a staged low-to-high-resolution pipeline to significantly reduce inference time. The method generates main structure at low resolution, performs super-resolution in pixel space using a lightweight GAN-based model, injects noise for high-frequency resampling, and refines details at high resolution.
- MrFlow achieves up to 10.3x end-to-end speedup on Qwen-Image and 8.25x on FLUX.1-dev while keeping OneIG within a 1% gap relative to unaccelerated models.
- When combined with pre-trained timestep distillation strategies like Pi-Flow, MrFlow can achieve up to 25x generation acceleration.
- The implementation requires no finetuning, learned upsamplers, or custom kernels, relying on standard PyTorch and Diffusers pipelines.
- The approach is compatible with FLUX.1-dev, FLUX.2 Klein Base 9B, Z-Image-Turbo, and Qwen-Image families.
MrFlow allows users to accelerate text-to-image diffusion without retraining models or using system-level optimizations, making it a practical solution for faster generation while preserving visual quality.