Consider joining my Discord or my Telegram (if you speak Russian). That's the only way to support me at the moment. :)
Artist's resources:
https://www.pixiv.net/en/users/60588249
Uhh... I don't know what to say about this one except that it works damn amazing. Don't forget to put "bar censor" to negative and "uncensored" to positive prompt. Also consider following and subscribing the artist. Yup.
Description
FAQ
Comments (2)
Hi,
> I forgot to check if Noob already knows the style of the Artist itself before training the LoRA.
There is absolutely no way the model would perform better at any style than a mid or well-trained LoRA.
LoRAs easily outperform styles that the model already knows - especially if you use high-quality, uncompressed, artifact-free source images (resized to 1024xXXX) and good captions. At the very least, having accurate captions is crucial.
Additionally, adding more images to your dataset can significantly improve style representation. For example, if you currently have 100 images, expanding it to 150–200 could give you a noticeable boost in flexibility, quality, and stylistic accuracy.
Anyway, I hope you’re happy with your results. If not - well, there’s always room for improvement.
Только в конце написания текста заметил, что ты знаешь русский. Мне было крайне лень переписывать всё с телефона, но надеюсь, что описал мысль внятно.
Hi! Thank you a lot for clarification. I've always thought that there is a little point making dataset larger than 100 images (there wasn't even a hundred in this LoRA dataset to be honest).
Yeah I got you just right, but anyway it's better to communicate in English here since other people who don't speak Russian may find the info useful as well.