This was a test to burn off my Tsubaki 2 coupons. I generated a batch of usable photo real (fictional non-existent persons) sets of a 'photobooth' shoot with a hot topic styled woman. The dataset is not entirely the same person through out. Ideally what this lora would give is a range of different person types in realistic to semi real styles depending settings.
I do not know how well this will work, consider it experimental, this is my first photoreal lora.
The lora was trained on 53 images varied from SFW fully clothed and covered to fully nude NSFW. Version 1.0-PhotoR was trained on SDXL Photopedia. The images are all adult buxom proportioned and all female (no futa or otokonoko in this one, which isn't to say it won't be compatible with that if given supplemental resources and settings/prompts). The images are simple photobooth background like ones you would find at a shopping mall.
Recommended settings:
lora weight -1.0 to -0.4 on inversion. 0.4 to 1.0 on posative
cfg: 1.5 to 12.0 pending rescale and shift
steps: 20 to 60
VAE: SharpSpectrumVAE
The lora is also built to be tested on hyperanime and anime styles. (Will see if it translates or not).
For the data set, I did best to eliminate any with bad hands or bad limbs or bad details. Result should be somewhat good but expect some face drift to illustration unless adjusted and some errors to adjust.
