An experimental LORA and my first. The idea was to be able to produce more realistic female faces with greater variety by a LORA trained on beautiful woman with longer noses.
Works suprisingly good on a variety of ethnicities but requires different LORA strenghts (see images). Best variety is at low strength (0.5-0.7). High values, >1, will create an average woman with long nose.
Flux2-Klein-9b
Trained on 512, 768 and 1024. Wide range of strength on distilled 9b model (see examples). Works much better on distilled than on base model.
Z-Image
ZiT model v.1.0 now trained on 1024x1024 and is an improvement. Best strength is now around 0.7 with better variability and detail. When using higher strength more detailed prompt is necessary and additional prompts to avoid blur and artefacts (see examples).
Description
First version.
FAQ
Comments (7)
This is wonderful, if you go on to make one for say, mouth width, lip thickness, eye size, eyebrow shape, chin shape, we can (((( FINALLY )))) start to make truly randomized NSFW ish females that don't all look related. Let me be the first to praise you for this, variability equals hottness, that's just biology 101.
I am following you now in the hope you can fix Flux identical (mostly) female faces.
Thank you! Very much appreciated. I was first trying how ZiT compares. It shows some more realism, but LORA training seems to be more fragile and prompting behaves differently (more precise).
1024x1024 only required if training fine detail like skin texture, z image alrerady have skin detail/ 512p is enough for likeness
@iluvlamia The higher res LORA has definetily higher skin details now, but also improved the range of usable LORA strength. My gut feeling is that precise and manual captioning the training set could improve versatility.
@hanswurst13 bigger nose too plz. really big :D
Correclty prompted you can already get all the facial features you want.
@GlowingGuardianGirl Yeah, tend to agree for Flux2 which has vastly improved facial diversity and prompt adherence. In any case my models are just products of experiments and curiosity.