My second public LoRA is an artistic style. I love anime but i think more western art styles are needed too. That's why I have decided to make one about the Spanish artist Alfonso Azpiri. He was well known for his erotic Lorna comics and for creating more than 200 Spanish video game covers in the 80s and 90s. He also published his comics in the prestigious magazine 'Heavy Metal', and also made comics for young audiences like those of his character Mot. Its graphic style is very characteristic and attractive.
This LoRA is trained with Anything V3 checkpoint and the Kohya-ss GUI utility. Works well with a strength of 0.6 - 1 and most of the anime checkpoints, but the best results are with Anything V4.5. This LoRA needs the negative embbeding EasyNegative to achieve optimal results. If you don't want to use embbedings, you can instead write '(worst quality, low quality:1.4)' in the negative prompt.
How to use EasyNegative:
First, download it from here.
Put the archive downloaded in the folder 'embbedings' of the Stable Diffusion WebUI folder.
Write 'EasyNegative' in the negative prompt.
It's interesting to note that you have a style 100% Azpiri with strength from 0.8 - 1, but below that, you have some other sort of western style smoother and more detailed, in between 0.6 and 0.8. You can see that in one of the pictures, with a comparision with diferent strengths of the LoRA.
This LoRA is too nsfw, so expect to see a lot of nude or semi-nude womans in your generated images. Although the LoRA has not been specifically trained with the famous character Lorna, it is very easy to generate her without the need for other LoRA. You only have to write your prompt like this:
1woman, blonde hair, long hair, make up, orange jumpsuit, big breasts, science fiction
You can see her in one of the example pictures.
It's very important to use hires fix to upscale , but below 2x, with denoise streght below 0.3 and 10 hires steps, and using as upscaler 'none' if you want to conserve the style. The upscaling process can change a lot the faces in this style even with lower denoise streght, so beware...
Description
This version results from doing some test to optimice the LoRA, using a different method of training and more images. Although I don't consider it a finished version, I find it interesting to share it because of how nice the results are. Now the images are more detailed and soft, with better faces and eyes, and beautiful women. However, the backgrounds are much more generic despite being more detailed, moving away from Azpiri's own style. That's why I don't consider it a legitimate version 2, and I'm going to treat it as if it were a separate experiment.
The features of this version are:
Better detail, better faces and eyes.
Backgrounds with a lot of detail, but more generic.
Now the LoRA is a lot more hot... So you have to put in the negative prompt 'nipples', 'nude' and 'nfsw' if you want to do 'normal' images. Some times it's hard to get women to appear in normal clothes.
Now the LoRA is smaller, only 37 MB vs 144 MB from the previous version.
It is recommended to use hires.fix, but with caution, because the rescaling affects the faces of the characters and can change them a lot. This depends a lot on the model that is used, so it is a matter of trying different resolutions and observing the changes. Use 'none' or 'Remacri upscaler' (you can download it from here), 10 hires steps and 0.3 denoise or less. This time it is essential to use EasyNegative embbeding in the negative prompt to obtain good results.
The best results are with the model Anything 4.5. Yesmix is very good at getting backgrounds similar to the ones Azpiri drew, but it changes the facial features too much in the rescaling. The best option with this last model is to send the images to img2img, change the model to Anything 4.5 and reescaling the images with less than 0.35 denoise and the same prompt. This way you can keep the faces with few changes.