ComfyUI Extension: ComfyUI-piFlow

Authored by Lakonik

Created

Updated

110 stars

ComfyUI pi-Flow nodes for fast few-step sampling.

Custom Nodes (0)

    README

    ComfyUI pi-Flow Nodes for Fast Few-Step Sampling

    <img src="https://raw.githubusercontent.com/Lakonik/piFlow/refs/heads/main/assets/teaser.jpg" alt=""/>

    ComfyUI-piFlow is a collection of custom nodes for ComfyUI that implement the pi-Flow few-step sampling workflow. All images in the above example were generated using pi-Flow with only 4 sampling steps.

    pi-Flow is a novel method for flow-based few-step generation. It achieves both high quality and diversity in generated images with as few as 4 sampling steps. Notably, pi-Flow’s results generally align with the base model’s outputs and exhibit significantly higher diversity than those from DMD models (e.g., Qwen-Image Lightning), as shown below.

    <img src="https://raw.githubusercontent.com/Lakonik/piFlow/refs/heads/main/assets/diversity_comparison.jpg" width="1000" alt=""/>

    In addition, when using some photorealistic style LoRAs, pi-Flow produces better texture details than DMD models, as shown below (zoom in for best view).

    <img src="https://raw.githubusercontent.com/Lakonik/piFlow/refs/heads/main/assets/piflow_dmd_texture_comparison.jpg" width="1000" alt=""/>

    Installation

    This repo requires ComfyUI version 0.3.64 or higher. Make sure your ComfyUI is up to date before installing.

    ComfyUI Manager

    If you are using ComfyUI Manager, you can load a workflow first, and then install the missing nodes via ComfyUI Manager.

    Manual Installation

    For manual installation, simply clone this repo into your ComfyUI custom_nodes directory.

    # run the following command in your ComfyUI `custom_nodes` directory
git clone https://github.com/Lakonik/ComfyUI-piFlow

    Workflows

    This repo provides text-to-image workflows based on Qwen-Image and FLUX.1 dev.

    pi-Qwen-Image

    Currently supports the Qwen-Image text-to-image base model (and possibly some of its customized versions). Qwen-Image-Edit may be supported in the future.

    Please download the image below and drag it into ComfyUI to load the pi-Qwen-Image workflow.

    <img src="workflows/pi-Qwen-Image.png" width="600" alt=""/>

    Model links

    Base model

    pi-Flow adapter

    Text encoder

    VAE

    Adapter strength

    Increasing adapter_strength in the loader node can reduce noise and enhance text rendering. This may be helpful especially when using a customized base model or additional LoRAs.

    Sampler steps

    The 4-step adapter works well for any number of sampling steps greater than or equal to 4.

    pi-Flux

    Currently supports the FLUX.1 dev text-to-image base model (and possibly some of its customized versions).

    Please download the image below and drag it into ComfyUI to load the pi-Flux workflow.

    <img src="workflows/pi-Flux.png" width="600" alt=""/>

    Model links

    Base model

    pi-Flow adapter

    Text encoder

    VAE

    • Download ae.safetensors and save it to <br>models/vae/ae.safetensors

    Adapter strength

    Increasing adapter_strength in the loader node can amplify contrast, reduce noise, and enhance text rendering.

    Sampler steps

    Use gmflux_k8_piid_4step.safetensors for 4-step sampling and gmflux_k8_piid_8step.safetensors for 8-step sampling. Using other settings may result in amplified or reduced contrast, which could be re-calibrated by adjusting the adapter_strength.

    Guidance

    The adapters only work with guidance set to 3.5. Do NOT modify this value, otherwise the results will be very noisy.

    Training Your Own pi-Flow Models

    Please visit the official piFlow repo for more information on training.

    License

    This code repository is licensed under the Apache-2.0 License. Models used in the workflows are subject to their own respective licenses.

    Changelog

    • v1.0.5 (2025-11-11)

      • Add experimental support for polynomial-based DX policy.
      • Update README.md and pi-Flux workflow (highlighting the FluxGuidance setting).

    • v1.0.4 (2025-11-09)

      • Fix a bug in GM-Qwen when running in BF16 precision.

    • v1.0.3 (2025-11-09)

      • Added support for scaled FP8 base models.