Z-Image (base) - Not Turbo. Great for LoRAs! (Patreon)

Video - https://youtu.be/dAnceJ13HP4

Overview

Z-Image not-so-turbo is here and it is GREAT for training LoRAs! Yes, it is slower to generate as it needs CFG 4 and 25-50 steps, but you can also use a negative prompt now. Some image variation is also possible between seeds, and you make get a slightly wider variety of styles compared to ZiT.

Workflow Information

The model files are linked in the workflow itself:

Note that where there are various CLIP models in the Comfy Z-Image directory, the fp8 model gave error for me. You mileage may vary.

As usual, GGUF files are available from huggingface - https://huggingface.co/unsloth/Z-Image-GGUF/tree/main - though you'll need to use the GGUF loader instead.

Just the usual custom node requirements, no different to my Z-Image-Turbo workflow if you've used or built that already:

Before you start, be sure to use the options control panel to quickly and easily enable or disable any groups:

By default, the workflow will generate a 1024x1025 rodent image using only the base Text-2-Image group. For example:

If using AI Enhanced prompting, be sure to have Ollama installed & your choice of model(s) downloaded.

Gemma3 4b is a great vision capable LLM with a tiny footprint.

The Steps & Sampler apply only to the Stage 1 generations:

For Stage 2 generations (detailing/upscaling) set your options in each group for steps, sampler, scheduler, etc:

The GET node can be used to get anything of the necessary type:

Thus, if you want to use IMAGE UPSCALE on the Text-To-Image result, you'd select IMG_TTI

The names can be seen in each group, for example, in the TTI group, you can see both the LATENT and the IMAGE:

This means you can CHAIN loads of groups one after the other if you like! For example:

Stage 1 TTI -> HiRes Fix -> Image Upscale -> Ultimate Upscale

The same goes for the Image Compare groups, so remember to set the TWO images you'd like to compare BEFORE you start your generations:

NOTE: The Union Control Net model for Z-Image (base) doesn't seem to exist yet, and the old Union 2.1 for ZiT doesn't work well with base. Best to not enable those groups until the model exists!

NOTE2: Also added a version 5a with seed & CFG in the stage 1 settings group, as well as other minor fixes.

Have fun :)