1.9 KiB
1.9 KiB
Example Workflow
This repository includes sanitized example workflows:
workflows/z-image_5.4_emberframe.json
workflows/z-image_5.4_emberframe_PiD_Experimental.json
z-image_5.4_emberframe.jsondemonstrates the base non-PiD Z-Image Base / Z-Image Turbo split-sigma workflow using EmberFrame prompt and resolution helper nodes.z-image_5.4_emberframe_PiD_Experimental.jsondemonstrates an experimental PiD final-latent decode branch added to the same split-sigma Z-Image workflow.
What Was Sanitized
- private/custom LoRA filenames were removed and LoRA slots are disabled
- custom merged checkpoint names were replaced with placeholders
- local image inputs were replaced with
example_input.png - temporary preview image metadata was removed
- prompt-helper notes were replaced with public-safe example text
- old private wildcard references were replaced with the bundled example wildcard files
- legacy helper-pack node types were replaced with EmberFrame node types where applicable
Before Running
After loading the workflow in ComfyUI, set these to your local files:
- Z-Image Base diffusion model
- Z-Image Turbo diffusion model
- VAE
- CLIP/text encoder
- PiD checkpoints
- optional ControlNet, depth, upscaler, and SeedVR2 models if you use those sections
If you use the img2img path, replace the placeholder example_input.png with your own image.
Required External Node Packs
The workflow still needs external packs such as:
ComfyUI-PiDrgthree-comfyComfyUI-Image-SaverComfyUI-Easy-UseComfyUI-KJNodesRES4LYF- ControlNet/depth helper nodes used by the optional sections
- SeedVR2 / Ultimate SD Upscale packs used by the optional upscaling sections
Install emberframe-nodes once for the EmberFrame helper nodes used by the main sampling and prompt-helper sections.
PiD Reminder
For Z-Image / Flux PiD, keep PiD scale at:
scale = 4
Both the 2k and 2kto4k PiD checkpoints are 4x decoders.