ComfyUI-ZImageDit

ComfyUI-ZImageDit
★ 5

ComfyUI插件SDNQdiffusers集成显存优化
在ComfyUI中非官方集成diffusers的SDNQ管线，便于比较质量并节省显存运行SDNQ模型。
💡 在ComfyUI中运行并对比SDNQ模型效果与显存占用。
🍴 2 Forks💻 Python🔄 2025-12-03
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/e98a62d17551
📦 requirements.txt
diffusers>=0.35.2
transformers>=4.56.1
safetensors>=0.4.0
quanto>=0.1.0
📄 README

ComfyUI-ZImageDit

What is this ?

  • an Alpha repo: unofficial diffusers integration of the official SDNQ pipeline to run in ComfyUI
  • …because I wanted to compare quality and be even more vram savy via SDNQ which is not officially supported and experiments with parameters

    • What can I do with this ?

    Check these example LLM “Clones” , credits to the original authors (Civitai) for variety of generes, styles, media.

    Notes:

  • installation

    • you might have to install some pip packages manually, nothing too difficult

    you need: accelerate, the latest diffusers from source to support z-image pipeline

  • install_sdnq.bat might help on windows because it looks like their toml file has an issue with double licensing (open inside the bat and change paths)
  • diffusers to install the latest diffusers manually via git to support the pipeline (from the embedded python folder if using portable comfyui):

    • python.exe -m pip install git+https://github.com/huggingface/diffusers.git

  • for flash attention (optional) find a .whl, if you need you can try these places:
  • seems to be the best place to find them:
  • https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.5.4
  • other places
  • prebuilt wheels https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.4.10 (i ended up using one package from here, it gives a nice speed boost, sage attention makes it slower, not sure why)
  • prebuilt wheels https://huggingface.co/Kijai/PrecompiledWheels/tree/main
  • prebuilt wheels https://huggingface.co/lldacing/flash-attention-windows-wheel/tree/main
  • about compile: does not work, for me.

  • if startup fails check requirements for what is needed (quanto is not needed for these nodes, but for the other broken ones)
  • weights are downloaded by diffusers on first run for sdnq nodes, in you huggingface default cache folder unless you change it
  • some option dont work or I did not finish porting, test.
  • there are other files in the other folders but they are experimental, ignore them (you might need quanto even or other installs)
  • internally sampling happens with flowmatching euler
  • only tested on windows (but linux should be even easier)
  • Platform: Windows
  • Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
  • pytorch version: 2.8.0+cu128
  • xformers version: 0.0.32.post2
  • Set vram state to: NORMAL_VRAM
  • Device: cuda:0 NVIDIA GeForce RTX 3080 : cudaMallocAsync
  • ComfyUI version: 0.3.75
  • ComfyUI frontend version: 1.33.8
  • Total VRAM 10240 MB, total RAM 32560 MB
  • if you are on linux… you are smart enought to know what to do

    • Enjoy!

    Enrico aka ErosDiffusion

    ps.: you might have issues installing, but I have no time to support 😀

    additional notes:

  • this does not use ComfyUI memory management, so use carefully.
  • I have added an option to unload but did not test it not sure it works.
  • the memory footprint is around 7gb vram more or less, you can safely run up to 2048×2048 i can run lmstudio with qwen4 3b in parallel and between ram and vram and this, and never get oom.

    • ´´