Z-Image ControlNet Compact

A compact streamlined workflow for the Z-Image model.

_________

Required models:

Z-Image Turbo models: https://civitai.com/models/2168935/z-image

Z-Image ControlNet model: https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/Z-Image-Turbo-Fun-Controlnet-Union-2.1.safetensors

(ControlNet model must be placed in the models/model_patches folder)

Required Extension:

AIO ControlNet Aux: https://github.com/Fannovel16/comfyui_controlnet_aux

_________

Usage Tips: The ControlNet model for Z-Image isn't quite as good as the SDXL Union model. If using it at a strength of 1.0, it typically comes out too crispy or with artifacts from the preprocessor image, so it is best to use it at a strength of 0.8 or lower. I think it is more like the behavior of SDXL ControlNet-LLLite model types, rather than a fully fledged ControlNet model. It can have artifacts depending on the preprocessor used, such as line-based preprocessors like Canny or TEED, where the outline of the preprocessor image can be visible in the final image, sometimes even at lower strengths. Depth preprocessors seem to be the better option since it doesn't produce these artifacts. Other preprocessors may or may not work well and will require tweaking the ControlNet strength.

Don't be afraid to experiment with Samplers/Schedulers or increased steps. Step counts such as 9, 12, 15, or 16, can potentially produce better results, but sometimes higher steps are either worse than 8 steps quality wise, or start to lose prompt adherence.

Tested best Sampler / Scheduler combinations:

ResMultiStep - SGM Uniform (smoother skin), Beta (more detailed)

Euler - SGM Uniform, Normal (very good), KL Optimal, Beta, Bong Tangent

LCM - Beta, Bong Tangent

UniPC - KL Optimal, DDIM

Heun - Linear Quadratic, Beta

DPM2 - DDIM, Bong Tangent

DPM++ 2M - Linear Quadratic, Beta (better)

DPM++ 2s - Karras (very good), Normal, Linear Quadratic, KL Optimal, Beta, Bong Tangent (very good)

_________

Shift CFG: works similar to how it works with WAN and Flux. A value of 1.0 means it's off. Use higher values if you want to experiment with potentially different model behaviors.

_________

For those having issues with ComfyUI lately: They have decoupled the backend code from the frontend code, or at least, it doesn't update everything properly with the ComfyUI update bat file. You have to add the startup flag "--front-end-version Comfy-Org/[email protected]" to your ComfyUI run bat file. This will force ComfyUI to use the latest frontend version. Leave the flag there or it will revert, and in the future lookup the latest ComfyUI frontend on Github and change the number in the string to whatever the latest version number is.