Tuktirey Sully from Avatar: The Way of Water, because I love her, rofl.
I saw that there were some other models of her already, but none entirely accurate. As my thing is to try and create LoRAs which are "on model", I decided to give it a shot, with ZiT seeming like the best candidate for accuracy.
I would have done this on regular Z Image, but it's so much slower, and in my experience the quality is not worth the extra wait.
This was trained using 220 1024x1024 crops from The Way of Water, with AI Toolkit. 4k steps (that's just where I stopped it, but it definitely didn't need any more than that, lol).
Minimal captions, mostly just the trigger word "tuktirey"
If you want, the training config is here:
It has default sample prompts which I did not bother to modify, as I disabled sampling.
These images were made using a two-part workflow, that uses a few steps of Z Image Base for extra image variation, then ZiT for refinement. But you can use this LoRA with whatever ZiT workflow you choose.
I had another version before this, but it was totally borked. AI Toolkit defaults to rank 32 alpha 32 for some reason which led to complete destruction of everything that wasn't Tuk. So I set it to rank 16 alpha 2 and retrained.
This LoRA, while it does accurately recreate the character, does have some problems:
1. Backgrounds still get a little bit borked. For example, if you ask for her standing next to a tree, the tree will have braids on it.
2. It usually renders her with a normal amount of fingers rather than the Na'vi four. Sometimes it gets it right, though.
3. Outfits aren't very flexible. You can try lowering the LoRA strength a bit to try and fix that.
4. Little bit of body horror sometimes when rendering more than just a portrait view, especially on hands. Might be a ZiT thing, or a consequence of my two-part workflow, or something to do with my training settings.
5. Only used screenshots from "Way of Water", none from "Fire and Ash". If I train again, I will try to add some from the third film.