Another request done.
I tried lowering the dim to decrease the filesize of the model and it still looks good I will experiment more with it in the future so your folder doesn't get too big.
This is one of the main girls from the anime "Rent-a-Girlfriend" I only pruned eye and hair color so the outfit is pretty flexibel. I included a lot of different outfits in the dataset and it works kind of also her alternate hairstyle is part of the dataset as well and you can easily get it.
Images all contain metadata and you can look at them for guidance.
I'm pretty happy with the Lora overall.
Have fun generating some pictures with it.
Description
FAQ
Comments (16)
Can you make Marin Kitagawa cosplay as Shizuku kuroe?
Sure I will make one soon I have to retrain an old model beforehand though
It didn't produce the same result for me even though I used same prompt as mentioned in the images
You can literally copy the prompts and get the same image they are not inpainted and dont use another lora so there must be something wrong with your settings. These images are upscaled with 4x Anime Sharp maybe thats the reason why they look different but even if another esrgan filter with a denoising strength of 0.35 should produce similar images
@foxtrott I can show you one of the test runs, you can check the params i used here:- https://mega.nz/file/uqoWFQIL#7Ug6wDZnkdjfFqUug0jXI8gwdcRSdSC3k4pl8xpp2Wo
and i get this image:- https://mega.nz/file/7jIQlaAK#53UZHytu7gf6tttmjokPKkU9wZSbFKZwa6JYKx8DmbI
Is it due to lack of upscaling?
You used different params. Sadly the resolution is not shown in the metadata but you can download the preview pictures and see it in your png data. Also do you use a vae ? Your picture didnt look like you used one. I would recommend the wd vae also known as kl-f8-anime2.ckpt. And the picture you wanted to recreate was made in 640x640 res it should look similar on low res but it was upscaled and that can change it but shouldnt be noticeable. If you upscale use the esrgan_4x or any esrgan based upscaler. I used anime sharp and set denoising to 0.35. You can find other upscaler on the wiki sadly i dont have the link anymore but shouldnt be difficult to find.
@foxtrott isn't this waifu diffusion vae? I will test and let you know if it works
hello, very good, your loRA wanted to know how you do it, do you have a link where you learned from, thank you very much
The model is good for only generating chizuru. If you tried generating anyone other than her like 1 male or 1 female, it will fail
The dataset had like no picture of multiple female and only very few were male chars were present that could be why
@foxtrott That might be the reason
While creating the images the eyes looks little weird, https://drive.google.com/file/d/1_QoNThN1bsk1EG3wfKEFLOqCiYZdiw7K/view?usp=sharing Can you tell me what am I doing wrong?
Which model are you using this shouldn't happen I think it's either the model that you are using is not compatible or maybe there were some update and stable diffusion changed how it reads Loras and the old Loras have to be updated. Did you use the additional network extension while making it or the default Stable Diffusion function of loading Loras ?
@foxtrott I'm using AbyssOrangeMix2_hard model and no i didn't use any extension
The picture is full body and the face will be deformed
Have you tried high res fix ? Full body pictures are normally not really good at portraying the face if it is not in high res because the character is far away from the viewer i don't think that's a problem with the Lora but more a problem with stable diffusion in general