A model depicting the character Tyranny from the Critical Role Campaign 4.

There wasn’t a character LoRA for her, so I made one myself.

Is it good? Probably not.

Lora is also likely trained on too little dataset.

Version V1.3 is trained on a dataset roughly eight times larger. It should perform a bit better. Not all styles and checkpoints understand “horizontal pupils” correctly (change strength appropriately). The earrings are still inconsistent.

V1.5

Negatives: cross pupils

I highly recommend add this to negatives, it should help with pupil shape a lot. Many checkpoints still don't recognise "horizontal pupils" and try to go for cross shape.

Tried to simplify prompting, as You should be able to get good eyes only with "tyranny eyes" vs previous "yellow eyes, horizontal pupils, yellow sclera".

Robes should be more consistent.

Headdress should be much easier to prompt in/out.

Earrings/piercing should be more consistent.

V1.51 (ANIMA)

Negatives: cross pupils

I highly recommend add this to negatives, it should help with pupil shape.

It's same dataset as V1.5.

V1.61 (Krea2)

It got similar dataset to previous versions, but it's not tagged, but poetry prompting that Krea2 likes.

It still not perfect, as with some tests it turns out that Krea2 doesn't get at all shapes of pupils or different than standard sclera (only prompt that model react was iris color).

Some of traits are probably bleeded into main trigger (TyrannyCR) and are hard to change with description/prompt. When I would get better with using Krea2, I will fix it.