UPDATE 02/11/2026 ZIT V2

It helps to specify characteristics like "orange wings" "white one piece body suit" "orange and blue birdlike helmet" "white hair" "blue boots" if you want the traditional look.

V2 is captioned differently. works really well across all mediums

UPDATE 12/4/2025

Sorceress Z Image Turbo

TRIGGER WORD: S0rceress

For Z Image Turbo version, use trigger S0rceress

My first Z Image Turbo LoRa that I feel is good enough to share. Lora strengths around .6-.85 are best.

TRIGGER WORD: S04cer3ss

This is a tribute to my first crush, Sorceress of Castle Grayskull

A few observations and notes:

I made this Lora twice.

I trained the first lora on the same data set with a combination of Joy Captioning and my own captioning. Then I did a second Lora with no captions. Just a trigger word.

After going through both, I ended up going with the no captions Lora although I don't think that's always going to be the right choice. Mainly my choice for going with the no captions was that I trained the Lora on as much of a diverse set of media types and fan's interpretations as I could find and the no caption Lora felt like it had more artistic freedom to push the boundaries whereas the captioned images required more precise prompting. Somewhat hard to explain but I suggest people that play with making Loras try out both. See what works best. It's not a one size fits all for all Loras at least in my opinion.

Something else I noticed, despite specifically training on non ai images that had good hands and feet and no signatures or watermarks, somehow they still found their way in. Perhaps thats just the impact Loras have on it until the technology gets better.

Lastly. Most images were made on ComfyUi however I did make some on Forge so people could see the prompts that were used. I used the full Dev model for the images. It was by far the most vibrant and detailed model. If you try running this on a "realistic" flux model, I found the outputs were extremely plain and dull. I do not suggest it.