TRIGGER WARNING (or aesthetic advice): this model connects body parts to make poker cards, putting in a realistic models will result in bodies with severe deformations (I personnally think this is hilarious, I have seen most movies on this page https://en.wikipedia.org/wiki/Body_horror ). You can't post them here (considered graphic violence).
This Lora is experimental. See 'Centrosymmetry' in suggested resources, it looks better.
This makes figures that have (approximate) central symmetry, to use as figure for poker cards.
The recommended aspect ratio is 0.6 (this is approximately the aspect ratio on poker cards figures), so 460*768. You might get two figures if you use a square resolution.
Prompt:
no activation word
helper word: upside down, multiple views, white background, simple background
do not prompt for clothing items or body parts under the belt (feet, trousers, dress, legs, ...), it will break the symmetry
lora weight: 0.5 to get better drawing (but lots of bad cards), 1.0 to get lots of good symmetry (but blurry faces)
Additional Notes:
this was made by rotating the cards in OpenCV and correcting cards by hand (I'm not very good with GIMP so the data isn't top quality). This might require another round of generate/select/train.
dataset included
as you can see on the images (and on the 'CentroSymmetry' Lora), upside down faces are not very good. You can actually tell whether the generation is going to be good during the generation process, the upside down face must be drawn at the same time as the upper face (and not twice as blurry).
I'm stupid, the danbooru tag is rotational_symmetry, but it is mostly 'dual_persona', that is two opposite but related characters
added an example of a 'joker' prompt (no rotational symmetry)