Adult Actress, Lora was trained with about 90 pics
Trigger: j3nnif3r (not required)
Additional tags for enhancement and to show her acting skills:
"low angle shot / shot from above / feet up / hairy / buttplug" etc...
Recommended settings:
Distilled CFG scale: 3 -3.5
Steps: 20+
Sampler: Euler / DPM++2M
Scheduler: Beta / normal / simple / SGM uniform
Strength: 1.0-1.2
(my preference: Dist.CFG: 3.5 / 22 steps / Euler - Beta / strength 1.2)
Description
FAQ
Comments (2)
I would love to know your training workflow, I tried to train my own character and the lora works (depending on what checkpoint I'm using) but all of th eimages i get have some soert of texture/grain shapet like a grid of noise on them..
Hi, could be you issue is with the quality of the training data. For a person Lora, I create the training pics with Flux (or SDXL) itself, I describe the bodyshape, hair, skin color, etc. and apply a faceswap. Then select the best pictures with most likeness and uspcale them to 2048x2048, with Topaz Gigapixel. I use Fluxgym for the training, with 14 epochs, 4-6 repeats and training resolution of 768 (1024 would be better, but I am limited by my Vram of 16gb), in total like 3-4000 steps. Usually Epoch 14 or 12 will do the job. Hope this helps.