Z-Image Workflow Power-Ups! (Patreon)

Video - https://youtu.be/5TQulYkuiw8

Overview

Z-Image-Turbo is a great Free & Open Source Model and now I've powered-up my Z-Image workflow with even more things! The best bits being EVEN MORE DETAIL and HIGHER RESOLUTIONS! This workflow version now includes:

Union ControlNet Support

Power LoRA Loader

Detail Daemon

Detail Demon Split-Sigmas

Ultimate SD Upscale (go beyond 2048x2048!)

Improved AI Prompt Enhancer

Additional compare groups

Personally I wasn't too impressed with the Union ControlNet, so do let me know if you get any amazing results as most things I tried were a bit pants ;) It seemed to be even worse when combined with LoRAs, but it is still early days yet. Who knows, maybe it's a code thing or perhaps we'll get a better model soon? Either way, there's a group for that now!

After testing a lot of LoRAs, only the one from Ostris himself seemed to be any good with all of the other ones seeming to be way over-trained and actually producing worse images than with just prompting alone. Lowering the strength on a number of them helped, but once again it hasn't been long since the model was even released, so hopefully everything will get up to the same level as the one Ostris did (which is indeed a fun LoRA!)

On the flip side, Detail Daemon and Ultimate SD Upscale both work a treat and give you great control over the amount of detail you can get in your generations. Ultimate SD Upscale may take a while, but it's the only way to get over 2K resolution images without everything degrading.

Some of my other favourite examples...

Workflow Information

There are two versions of the workflow attached - the full version and one without AI Prompt Enhance.

For the full version you will need Ollama installed, along with downloading any ollama models of your choice. Mistral 3 is a great option, as is IBM Granite or Gemma 3.

Out of the various LoRAs tested, the Children's Drawings and UltraReal (at low strength) were my favourite:

Do let me know if you find any other you like!

The default Detail Daemon setting are very light:

If you set the bias to 0, you will likely need to lower the detail amount as well. Around 0.25 is a good place to start, and the start value seems to have a variable "tipping point".

For example, this is too much detail (detail 1.5, start 0, end 1, 0 for everything else):

With a detail amount of 0.7 and a start of 0.2. the detail is very high, but it still close to the standard, baseline (TTI) image:

A start of 0.07, however, gives a very different result:

When using the High Res Fix, don't forget you can choose which latent to use:

By default it is set to the Detail Daemon Split Sigma, but you can chose LAT_DD (from the Detail Daemon group) or LAT_TTI (the basic Text-to-Image generation) as well.

You can, of course, also set the scale_by value to your own preference.

The same goes for the Ultimate Upscale:

Again, by default it is set to the Detail Daemon Split Sigma generated image, but you could also pick any of the other IMG options.

Use a fairly low starting resolution (under 1920x1920), set it to IMG_HRF and you can get both the extra detail from the HiResFix option AND the Ultimate upscale ;)

For the Image Compare groups, you can also select which images IMG_*)to compare:

With the Detail Daemon Split Sigmas, you may like to switch outputs:

The "safe" option of denoised_output is the default.

First order samplers DO NOT WORK WELL if you use the plain (not denoised) output option:

If using that option, you'll do best with an *_sde option...

... or seeds2/3/sa_solver (or other options such as those from RES4LYF (res2s, res2m, etc)). Basically, any second or third order sampler should give better results.

Have fun!