Xingyu Zheng, Xianglong Liu, Yifu Ding, Weilun Feng, Junqing Lin, Jinyang Guo, Haotong Qin
4 min
Abstract
Hardware-agnostic strategies for accelerating text-to-image diffusion, such as timestep distillation and feature caching, can reduce inference time without custom kernels or system-level optimization. Among them, multi-resolution generation strategies have recently received broad attention, attaining more than 5x speedup without any training. However, the design of performing upsampling in the latent space, together with the selective modification of partial regions, causes these methods to exhibit noticeable blurring or artifacts. To this end, we propose MrFlow, a training-free multi-resolution acceleration strategy for pretrained flow-matching models built upon a staged low-to-high-resolution pipeline. MrFlow first rapidly generates the main structure at low resolution, then performs super-resolution in the pixel space using a lightweight pretrained GAN-based model, subsequently injects low-strength noise to enable high-frequency resampling, and finally refines the details at high resolution. Quantitative and qualitative results on FLUX.1-dev and Qwen-Image show that MrFlow exploits the quadratic token reduction and reduced step requirement of low-resolution sampling to achieve 10x end-to-end acceleration while keeping OneIG within a 1% gap relative to that before acceleration, significantly surpassing other training-free acceleration strategies, and requiring no training or runtime dynamic identification whatsoever. MrFlow can further be directly combined orthogonally with pre-trained timestep distillation strategies, achieving even higher generation acceleration of up to 25x.
Alex: So to recap the pipeline: fast low-res generation, scale it up with the GAN, add a touch of noise to clear up artifacts, then one final high-res polish.
Sam: That's it. And because the heavy model only runs for that final pass rather than for every step of the process, you avoid the main bottleneck. The result is roughly ten times faster than the standard approach—and it doesn't require retraining the underlying model at all. It's just using an existing model in a smarter sequence.
Alex: That's an elegant solution. Are there any downsides the authors flag?
Sam: They do note that quality can vary slightly depending on how complex the text prompt is. But the gap stays within a very narrow margin compared to the slower, standard method. For most practical uses, the difference is difficult to notice. The paper's argument is that this makes high-resolution image generation genuinely usable in everyday tools, where waiting nearly a minute for a single image simply isn't practical.
Alex: It's a useful reminder that speed and quality don't always have to trade off against each other—sometimes it's just about being smarter with where you spend your computing power.
Sam: That's the core insight. Focus the expensive work on the decisions that matter most—the global structure—and handle the fine details with a lightweight, targeted pass. It's a principle that shows up across a lot of engineering problems, and MrFlow applies it cleanly to image generation.
Alex: Thanks for walking us through that, Sam. And thanks to everyone listening to ResearchPod.