Z-Image-Turbo Fun ControlNet Union 2.0 (Patreon)

Video - https://youtu.be/Y-qmJZh1FRQ

Overview

Z-Image-Turbo-Fun-Controlnet-Union-2.0 is here - https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0 !

Update - 13/Jan/2026 - Version 2.1 is now also available :D

See https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1

The newly updated 2.1 file gives us better ControlNet functions AND better InPainting too, plus everything now seems to work much better with LoRAs as well.

NOTE: They now also have 8-steps versions and Tile :)

It supports multiple control conditions; including Canny, HED, Depth, Pose and MLSD can be used much like a standard ControlNet, even though it's loaded as a model patch through a new FunControl Node in ComfyUI.

As the node is NEW, don't forget to update to the latest version if you haven't already!

Also new in ComfyUI is a built-in version of manager - which you can install using "pip install -r manager_requirements.txt"

I've also attached the script I use to start Comfy (comfy_start.sh) which automatically updates everything anyway, so if you're using something similar you may like to add the new requirements as well. To use the new manager also needs a new command-line option as well.

Workflow Information

Even MORE options now because why not :)

The changes here are for Union Control, where you can now opt to use the new Union 2 ControlNet, Union 2 Inpainting (manual mask) or Union 2 Inpainting (prompt mask)

You'll find the three new groups at the bottom:

The first group uses the new Union 2 ComfyUI node and is for depth, canny, pose, etc:

The second group is for Inpainting with a manual mask:

The third group gives you prompt-based masking, and so is ideal for batching up tasks without having to manually mask everything:

A link to the new file and some more information is also included in the note:

To stop the non-masked areas from being so washed-out when using an empty latent, you may need many more steps (20+) or simply use the Image Encoded latent option, which you can toggle using the switch:

By default it will use the image latent, as generally speaking I found that to be better as I could then use the usual 6-10 steps and a lower controlnet strength value (0.65-0.75 is generally best). Lower values will still apply some control, and can actually be fun if you don't mind it not actually being as strict.

By default I'm using a new sampler (because why not!) - change to whatever you prefer :)

Each of the groups in the first column now have both LATENT and IMAGE set nodes, so you can easily pick what you want to send to either the latent or image upscaler nodes:

For example, if you want to use the HiResFix (latent upscale) group, you can pick any of the LAT_* options. You will, obviously, have to generate the latent for whatever you select first so that it can be passed though.

This means you can do fun chains such as:

Split Sigma Detail Demon -> Latent Upscale -> Image Upscale -> Ultimate Upscale

Or whatever you fancy! Just make sure you don't go over the 2k limit until you hit either Ultimate Upscale or SeedVR2.

Have fun :)