V2.5 for ZIT is trained on base and plays along well with other loras on ZIT. Improvement focused on accurate female genitalia.
Added version trained on Z-Image "Base". All samples are genned on Z-Image Turbo, where the main benefit is that it seems much more resilient on stacking loras and has less of an effect on overwriting features or look from other loras.
Added version trained on ZIT, the effect is pretty subdued but cuts away from the cinematic color with very little impact on detail on my tests, also steer models preference bit away from asian women. Works without trigger phrase, but was trained with "Low quality cellphone photo of". For accurate representation of genitals, look for some other lora.
Added WAN 2.2 T2V models, trained on the same dataset as V1. All gens done with Lightning 4 steps loras @480p, so likely there's possibility of improved quality by improving the workflow.
V1.5 is more consistent and degrades Flux quality less, but the nude selfie might not be as authentic or perform as good as V1.0. Art nude dataset was added and nude selfie dataset was expanded considerably.
Flux lora trained on non-specific set of 50 or so images of amateur nudes, mostly selfies and mirror selfies. Knows "nude selfie" and to lesser degree breast sizes and body shapes, prompt following for body sizes is limited but it's definitely not as stuck on the model body as base Flux.
Anatomy is not perfect by any means.
Description
Trained on Base, added set focused on genital generation.
FAQ
Comments (6)
the tits are vanished and not having a depth of field or any volume, it's flat but in a unrealistic way, also noticed some square pixelated pictures do you train with high res pictures?
Which model are you talking about?
@konflikti 2.5 zit same of the zbase
@carlitajohn89931 Then do you mean generating yourself or the examples? I run the same set of prompts so you can see the difference, but it is hardly optimized for the best performance.
The lora is based on amateur nudes so high definition is not the greatest of focuses, but I can see if I can improve the breast definition.
@konflikti i mean i don't know where you choose your dataset but for example see this one, the tite are not flat are visibly dephts its not 2d so i guess need to rework on that as well as the underpart it's not that realistic, and if you can also remove the blurriness you will have a good release and i'll be the first to use and share it. https://civitai.com/models/2366964/best-breast-and-nipples-for-z-image-base?modelVersionId=2670771
Just wanted to comment that this is a very good and probably underrated LoRA. Keep up the great work!