I improved many things but many still require more attention. The main thing that improved is the size of human's body in relation to a car is now more realistic and consistent. I have also reworked the captions so now it's a bit easier to control.
I guess this is the maximum of quality that can be achieved with 32 dim lora (200 mb lora), I will bump it to 64 in the next version.
The key tags to know:
csngs - is a trigger word, I'm not sure what it does since it's always present for 1girl in the dataset, but the idea was to make it trigger cool car posing
photo \(medium\) - add realism to your generations
front view, side view, rear view - helps to get car in the right direction
from above, from below, dutch angle - these I actually tagged
standing near car, squatting near car - useful tags when you don't want your character to be on car
on car - when a character is not just near the car
no humans - when you need just a car
leaning back, leaning forward
vehicle focus
sports car
outdoors, indoors
cabriolet - I tagged all cars that have their top retracted with this tag
photo background + mixed media - if you want 2d anime girl with photorealistic car
Cars from best to worst (yes, use underscores):
mazda_rx-7
porsche_911
toyota_supra_mk_iv
nissan_gt-r - r35 variant
mitsubishi_lancer_evolution - X generation
ford_mustang - GT 500 variant
subaru_impreza - generates grill wrong. You know when it's not wrong? When the car is blue...
mazda_mx5new - (ND variant of Miata) works, but lights shape is distorted and it ruins the vibe
mazda_mx5old - (NA variant of Miata) it just refuses to generate pop up lights, despite me having many pictures with them and what's funny is that it does them in landscape orientation
But also try cars that are not in the dataset:
mazda miata - for some checkpoints it can generate NC variant of Miata
bmw, mercedes-benz, lamborghini, ferrari
Super secret bonus:
My lora has learned the style of these artists:
vinneart - not as good as a separate lora would have, but very decent. I will pin an image with it below.
bokuya - simplistic flat style, very great to have since this lora is biased towards photorealism
Unlike any other loras, here you have to experiment with its strength for each of the generations. If you feel like car looks too detailed, sporty or has artifacts - reduce the weight. I also encourage you not to use any specific car tags if you don't need a specific car.
The output will be very different depending on which checkpoint you use it with. I had most success with Hassaku Style B 1.3 and WAI 11.
And most importantly: your cars don't have to be orange and the girl doesn't have to have "very long blonde hime cut"