2 LORA Characters - Z-Image

UPDATE #1: If you have an image prompt generator, find any image that you want to replicate and have it describe it with 300 words+. Take that description, input it into the first box and then ctrl+b the other three prompt boxes. It will do a better job at getting your characters into that replicated image. Still constrained within the limits of the z-image turbo, of course.

UPDATE #2: See below, it's a longer description.

---

This workflow takes two of your z-Image characters and puts them into one image. Essentially, it produces twins and changes one of them to your other character.

NOTE: If this workflow ever produces an index error or a noisy image (ESPECIALLY THE FIRST TIME YOU RUN THIS), change a single character in the prompts, and it will reset. Rerun it from there. The index error is caused by z-image going rogue and not producing two detectable people in an image and I don't know what causes the first run to be noisy.

Also, it's labeled left and right refinement, Florence2 has a tendency to switch it's detection so it may fix the right side character under the left labeled box. Regardless of label, it's fixing both characters.

Check your settings before running.

I couldn't get the scheduler_selector node to work as a setting so you'll have to change it throughout the workflow if you plan to use anything other than ddim_uniform.

The dpmpp_sde sampler with ddim_uniform scheduler has worked well.

Under the initial image generation, adding more detail can be done by tweaking the deatil_amount input on the detail sampler. Anything above 0.7 I've found to produce a lot of weird image-breaking artifacts.

Character variability is noted in the worklfow but basically, the higher it is, the more it changes the refiner from the original image. Recommend setting 0.85 and trying different things from there.

I won't upload examples of this workflow working because, well, reasons, you know.

I have no issues with anyone using or reproducing parts of this workflow. However, if you see errors, spot inefficiencies, found better settings, see something that will help others, let me know so I can update this post.

Enjoy!

---

UPDATE #2: Only put one of your characters into an image. Combine this with the picture captioner from update #1.

Under settings, pull the front nodes away until you find "Seed Generator.' Pull this out by itself and then set it to fixed.

Under the mask creator, find the two nodes labeled "Florence 2 Coordinates" and pull them out below. They should be 0 and 1 in the index box.

Disable your first person lora node by ctrl+b.

Generate an image. This will be a z-image turbo image without LORAs influencing it. The 2nd person LORA will then be applied to only one of the characters in the picture.

If you want to switch which one your character is converted to, since the seed is fixed, switch the 0 and the 1 under the indexes. This will then produce the same scene but your character will switch positions. To make the scene even more exact, just turn off the seed enhancers. Z-image turbo keeps scenes almost exact unless a setting is changed in the prompt or model loading stages.