comfyui-kandinsky22

comfyui-kandinsky22
★ 9

Kandinsky 2.2先验合成img2img深度提示
组合Kandinsky 2.2的prior与decoder模型，支持查找文本/图像先验、按权重融合、生成或基于图片准备潜变量，并接入深度提示。
💡 在ComfyUI中按权重融合文本与图像先验以生成或编辑Kandinsky 2.2图像。
🍴 4 Forks💻 Python🔄 2025-04-02
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/9671236b7e59
📦 requirements.txt
accelerate==0.27.2
diffusers
@
git+https://github.com/huggingface/diffusers@31de879fb4530ad203dbdaef993404c924daf35d
torch>=2.2.0
📄 README

Kandinsky 2.2 ComfyUI Plugin

Use the models of Kandinsky 2.2 published on 🤗 HuggingFace in ComfyUI.

Features provided

Nodes provide options to combine prior and decoder models of Kandinsky 2.2.

  • Find priors for text and images.
  • Combine priors with weights.
  • Prepare latents only or latents based on image (see img2img workflow).
  • Use depth hint computed by a separate node.

    • All the weights can be found in Kandinsky Community on 🤗 HF in Kandinsky 2.2 Collection.

    Workflow Examples

  • A simple text-based workflow (source).
  • A workflow based on image prior embeds (source)
  • An Image-To-Image workflow (source)
  • A depth based workflow (source)

    • *Note: Don’t forget to switch to kandinsky-2-2-controlnet-depth in decoder node.*

    Installation

    For the easiest experience, install the Comfyui Manager and

    use it to automate the installation process. The repository is not included in the list at the moment, but

    you’ll need Marigold depth estimation, that can be installed via manager.

    To install the plugin, open the terminal, cd to /custom_nodes, and clone the repo:

    git clone https://github.com/vsevolod-oparin/comfyui-kandinsky22

    Install the requirements using:

    python -s -m pip install -r requirements.txt

    Download Models

    Go to models/checkpoints in ComfyUI directory and run the command

    git clone --depth 1 <HF repository>

    E.g. to download needed checkpoints for the presented pipelines run the following

    git clone --depth 1 https://huggingface.co/kandinsky-community/kandinsky-2-2-prior
git clone --depth 1 https://huggingface.co/kandinsky-community/kandinsky-2-2-decoder
git clone --depth 1 https://huggingface.co/kandinsky-community/kandinsky-2-2-controlnet-depth

    Note 1: Git won’t show much of the progress. You’ll need to wait till the models will be downloaded.

    Note 2: Argument --depth can be skipped, but you’re risking to download a lot of unnecessary data.

    Acknowledgments

    A special thanks to:

  • The developers of Kandinsky Model.
  • Patrick Von Platent for moving Kandinsky to diffusers.
  • Comfyanonamous and the rest of the ComfyUI contributors for a fantastic UI!