comfyui-zimage-sdnq
Custom ComfyUI nodes for running Z-Image Turbo with SDNQ quantization (q8/q4). Supports text-to-image, image-to-image, inpaint, ControlNet, and LoRA.
Nodes (12)
The edge map that makes Z-Image build to your sketch, not its own mood
The VRAM reset button for Z-Image SDNQ, disguised as a node
Works, with a quality tax
Loading Z-Image with its Fun ControlNet Union, without fighting diffusers
The 'use MiDaS' advice, packaged as one node
Attaching Z-Image LoRAs without ever reloading the model
Load the whole model in one node
The node that makes Z-Image build to your sketch, layout, or depth map
The refiner the community actually uses
Only the masked pixels move
Z-Image Turbo, 8 steps, one sampler — the whole point of this pack
The tile sampler that cheats the 2MP ceiling
ComfyUI Z-Image SDNQ
ComfyUI extension for Z-Image Turbo with SDNQ quantization.
Layout
comfy/: ComfyUI-facing node classes and the public registry.catalog/: model, LoRA, and control-weight discovery plus control-config inference.session/: typed runtime state, cache ownership, and live LoRA attachment state.loading/: base and control pipeline construction.generation/: seed handling and shared sampler runners.imaging/: image conversion and preprocess operations.integrations/: thin wrappers around ComfyUI, SDNQ, MiDaS, and torch device selection.zimage/: active Z-Image engine fork and LoRA implementation.workflows/: example ComfyUI graphs.
Dependencies
- ComfyUI
diffusers/sdnq(required)timm/opencv-python-headless(controlnet)
Models
Place models where ComfyUI already expects them:
- Diffusers models under
models/diffusers/ - LoRAs under
models/loras/ - Control weights under
models/controlnet/