3.7 KiB
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 at the hart 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 :D
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 2048x2048 i can run lmstudio with qwen4 3b in parallel and between ram and vram and this, and never get oom.
´´