Strenght:0.5-1. Trigger words: futa, futanari, penis, huge penis, testicles.
V2
Main trigger word "futa", but it can work without TW. Different settings can affect on variations of composition and can help to reduce errors. You can try different samplers, steps, cfg etc. See the example images for settings.
Steps 4 - 20.
Cfg 1-2.
Samplers - SEEDS 2, SEEDS 3, res 3s or res 6s works slightly better, Euler in my tests tend to produce more errors.
Scheduler - simple, beta, SGM uniform, DDIM uniform
Description
FAQ
Comments (3)
samples are looking really good. maybe the most coherent looking anatomy from all so far just by judging the samples.
I am curious to know how the 613 steps and 10 epochs look like as a total number?
if I understand this correctly, it means the total training steps were 6130?
i've started just today experimenting with ostris ai-toolkit, and I have been trying to find the sweetspot for total steps - since ai-toolkit doesnt work on the old kohya equation of img repeats x epochs = total steps. and when you choose to save the increment, that is your epoch...
and if 613 steps per "epoch" produced that good of a lora, this lora, then 600 seems like a good number to work around
what do you think?
i'm not sure but i think, that In ai-toolkit it's equal 6130 steps. But 3000-5000 also can work. The final result depends on number of images also and their quality, variantion, image promt quality and settings. This one i tested with default settings on CIvitAI lora trainer with 77 images.
@VergilSardos i think you shared a key important point - the amount of images in your dataset. I did a trial run with anatomy with only 30 images and I was dissatisfied, but with 17 images on an original character, it worked almost perfectly on the first run.
thank you for your insight, i will try increasing my dataset instead of adjusting for more steps. this is a good start.