Z-Image-Turbo: Multiple Character LoRAs (Patreon)

Video - https://youtu.be/1OIdS-9rDpc

Overview

Every wanted to use more than a single character (or object) LoRA for Z-Image-Turbo and found you got less than stellar results? If so, then FreeFuse is going to be the thing for you as you can now use up to 4 individual LoRAs together without them fusing strangely together!

FreeFuse - https://github.com/yaoliliu/FreeFuse

No more face or colour mixing, meaning you can generate multiple characters easily in a single image.

Workflow Information

There are 2 workflows attached - one for using just FreeFuse by itself, and one with the FreeFuse group also included in my main Z-Image workflow.

The Fuse-Only version has very minimal custom node requirements, as shown in the note:

As usual, custom nodes can be installed via the typical "git clone" or ComfyUI Manager:

Z-Image-Turbo works best:

I've tried to get it to work with Base, but the masking just didn't seem happy.

Be aware than SOME image sizes can cause issues. The default settings are fine:

However, something like 1280x720 will currently generate an error:

FreeFuse uses the prompting to align the tokens, so your prompts must match.

An easy way to make sure you have a good Main Prompt is to copy got Background Prompt into the main prompt first, then the prompt text you have for your other characters.

You can make small changes, such as positioning and lighting details without it getting too confused.

If you DON'T get a multi-coloured mask like the image below:

then the chances are your LoRA and Main prompts are too different. If you want your characters performing actions, it can be best to include those actions on the LoRA prompt as well.

For my Z-Image Mega workflow, there is now a FUSE option available too:

Remember to set whatever options you want BEFORE running the workflow.

The loader has bot ZIB & ZIT, so if you're not using FreeFuse then you could start with ZIB and use ZIT for upscaling if you like... or the other way around ;)

For Upscaling FreeFuse generated images, I think Ultimate Upscale is the best option:

If you don't mind super-bright colours and extra sharpness, then SeedVR2 is also a good choice:

Unfortunately, the typical IMG & Latent upscales will still merge LoRAs to give not-good results:

Though IMG upscale can do a reasonable job so long as you don't load the LoRAs there. Either select the ZIT model in the model GET node, or simple don't load the LoRAs in the main model loader.

Remember you can pick and choose your model/image/latent via the GET nodes in the groups:

Enjoy! :)