Xingyu-Zheng/MrFlow

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

Official implementation for MrFlow

This repository provides the implementation of MrFlow, a training-free staged sampling method for accelerating pretrained flow-matching text-to-image diffusion models.

MrFlow first samples a low-resolution image, upsamples the decoded result in pixel space with Real-ESRGAN, re-encodes the upsampled image, injects scheduler-consistent low-strength noise, and performs a short high-resolution refinement. The pipeline shifts most denoising cost from expensive high-resolution sampling to cheaper low-resolution sampling while preserving local detail quality.

✨ Highlights

• Training-free deployment. No finetuning, learned upsampler, or model-specific retraining is required.
• No custom kernels. The implementation uses standard PyTorch, Diffusers pipelines, and scheduler controls.
• Strong aggressive-speed regime. MrFlow reaches more than 10x end-to-end speedup on Qwen-Image while preserving visual quality.
• Works with distilled models. The same pipeline can be combined with pretrained timestep-distilled models such as Pi-Flow and FLUX-schnell.
• Compact staged design. The implementation transfers across Qwen-Image, FLUX.1-dev, FLUX.2 Klein, and Z-Image families.

📢 News

• [2026/07] 📰 MrFlow is featured on Hugging Face Daily Papers.
• [2026/07] ⚡ We release the MrFlow ComfyUI plugin.
• [2026/07] 🔥 The MrFlow paper is available on arXiv, and the source code is released.

🛠️ Installation

Create a Diffusers-compatible environment for the target backbone. The demos use:

• PyTorch
• Diffusers
• Transformers
• Real-ESRGAN

MrFlow uses Real-ESRGAN for x2 pixel-space super-resolution. Install Real-ESRGAN from the official project and download the x2 weights:

https://github.com/xinntao/Real-ESRGAN

The scripts contain placeholder checkpoint paths. Replace them with local paths to the pretrained text-to-image model and Real-ESRGAN x2 weights before running.

🚀 Quick Start

The repository root keeps only two minimal reference scripts plus the shared scheduler helper:

Script	Model	Setting	Output
qwen_image_mrflow.py	Qwen-Image	MrFlow 12plus1	outputs/qwen_image_mrflow_12plus1/
flux1_mrflow.py	FLUX.1-dev	MrFlow 12plus1	outputs/flux1_mrflow_12plus1/

Edit the checkpoint paths at the top of each script:

MODEL = "/path/to/Qwen-Image"
REALESRGAN_X2 = "/path/to/RealESRGAN_x2.pth"

Run:

python qwen_image_mrflow.py

python flux1_mrflow.py

Each script saves:

• stage1_low.png: low-resolution generated image.
• stage2_upscaled.png: Real-ESRGAN x2 upsampled image.
• stage3_refined.png: final high-resolution refined image.

⚙️ Core Settings

Setting	Low-resolution steps	Refinement steps	Direct sigma	Typical use
12plus1	12	1	0.12	Aggressive acceleration.
20plus1	20	1	0.12	Higher-quality operating point.

The high-resolution refinement uses an explicit direct-sigma schedule. For example, 12plus1 denotes 12 low-resolution denoising steps followed by one high-resolution step from sigma=0.12 to 0.

📦 Supported Demos

Parameterized variants and additional model-family demos are available in examples/.

Script	Backbone	Notes
examples/flux1_mrflow.py	FLUX.1-dev	Training-free MrFlow.
examples/flux1_piflow_mrflow.py	FLUX.1-dev + Pi-Flow	Combines MrFlow with distilled weights.
examples/qwen_image_mrflow.py	Qwen-Image	Training-free MrFlow.
examples/qwen_image_piflow_mrflow.py	Qwen-Image + Pi-Flow	Combines MrFlow with distilled weights.
examples/flux2_mrflow.py	FLUX.2 Klein	Base and non-base variants.
examples/zimage_turbo_mrflow.py	Z-Image-Turbo	Reduced-step model plus MrFlow refinement.

Run all configured examples with:

bash examples/run_examples.sh

See examples/README.md for command-line usage, FLUX.2 Klein presets, Z-Image-Turbo refinement defaults, and output filename conventions.

Pi-Flow examples are optional and require a separate local checkout of LakonLab. Set LAKONLAB_ROOT to that checkout before running the Pi-Flow scripts.

🧩 ComfyUI Plugin

The repository also includes ComfyUI-MrFlow/, a ComfyUI custom-node extension for Qwen-oriented MrFlow workflows. It provides helper nodes, editable workflow and API JSON examples, a reusable subgraph, and a model-link helper for split Qwen-Image bundles.

To use it, place or symlink ComfyUI-MrFlow/ into ComfyUI/custom_nodes/, restart ComfyUI, and open ComfyUI-MrFlow/examples/qwen_mrflow_workflow.json or load ComfyUI-MrFlow/subgraphs/qwen_mrflow.json.

🖼️ Results

Qwen-Image generation examples. With 12 low-resolution steps and one high-resolution refinement step, MrFlow produces diverse 1024-resolution samples on Qwen-Image while reaching above 10x end-to-end speedup.

Accuracy-efficiency trade-off. On FLUX.1-dev and Qwen-Image, MrFlow offers a flexible trade-off between generation quality and measured end-to-end speedup, and remains effective where other training-free strategies degrade sharply.

Runtime breakdown. For Qwen-Image 12plus1, measured end-to-end latency is 4.77s versus 49.32s for native 50-step inference. The main cost is shifted from high-resolution sampling to cheaper low-resolution sampling, while SR and VAE overhead remain small.

📊 Representative Numbers

Backbone	Setting	End-to-end speedup
FLUX.1-dev	12 + 1	8.25x
Qwen-Image	12 + 1	10.3x
FLUX.2 Klein Base 9B	12 + 1	8.79x
Z-Image-Turbo	8 + 1	21.0x
Qwen-Image + Pi-Flow	4 + 1	up to 25x

Speedups are measured end to end, including text encoding, VAE encode/decode, super-resolution, noise preparation, and diffusion forward passes.

🗺️ Roadmap

• Project README, framework figure, visual results, trade-off plot, and runtime breakdown.
• Implementation code.
• Public paper link.
• ComfyUI extension plugin.
• Demo video.

📝 Citation

If you find MrFlow useful, please cite our paper:

@misc{zheng2026multiresolutionflowmatchingtrainingfree,
  title={Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling},
  author={Xingyu Zheng and Xianglong Liu and Yifu Ding and Weilun Feng and Junqing Lin and Jinyang Guo and Haotong Qin},
  year={2026},
  eprint={2607.01642},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2607.01642},
}

🙏 Acknowledgements

This implementation builds on the Diffusers ecosystem and uses Real-ESRGAN for pixel-space super-resolution.