Due to Taylor's success as a youngling the base model training are tainted by too many young pictures creating a weird old/young affect when generating Taylor Swift images. So this Lora is trained on only recent images, a lot are taken from the recent era's tour.
Few quirks of note of this Lora:
In terms of clothing a lot of the training images included Taylor's iconic bejewelled leotard so even normal clothing is sometimes bejewelled or leotarded if you are not specific with clothing
Lovers of Taylors goofy smile will be happy as I made an extra effort to include a few in the dataset
Nearly every picture of Taylor has her holding a microphone, this sometimes pops up but is no worse than the usual AI hand qualms
Lighting plays quite an affect as it was difficult to find any natural lighting images but this just means you may need to be specific in this area if you don't get the desired affect
Makeup is not defaulted on images but trained in so do specify if desired
Any full body images, I fully recommend using aDetailer for best affect but nothing new there.
If you use this for NSFW images you will break my heart.
Trained on 1024 x 1024
I am still learning so any advice is greatly appreciated but this is the result of months of fun work!
Description
Ignore that this is V5, I waited till I actually had a good version before I released one, though I will likely do anouther one in the future
FAQ
Comments (2)
Thank you
Always train images from a time period. For sure if you train Madonna from 1980 with Madonna from 2024, you're gonna end up with an alien LoRA.
You gotta stick to images from only a few years of people's lives ;)
This being said, she looks cute in there. if you're ok sharing the dataset somewhere, I'd like to adapt the captions to booru and train her on Pony XL. I did a few tests and this (amazing !!) model actually generates impressively photo realistic images when fine tuned on realistic photographs. I train Swift for example, and just the LoRA loaded suddenly chairs, cars, bricks, anything in the image looks realistic. Great!
Anyway. Thanks for the training, and yeah share the dataset on pixeldrain or something eventually. I'll train that.
Cheers ;)