It's focused on women playing breastfeeding eachother
Description
FAQ
Comments (7)
what strength you using? and any chance of getting the showcase prompts? Or are they the same as the others below
Hey! Thanks for being my first comment!
This is new to me, so I am still trying to see how stuff goes.
And after releasing this I have been messing around with different models and prompts, so I think I understand what I should include as description and how I should improve the Lora.
Anyways... I do have really promising results with:
3:2 as an aspect ratio. This is the most important factor I believe so far.
Also I use Eular A... I dont have that much experience, but this scheduler have done miracles for me.
The strength of Lora should be appropiated based on model. Really. If it's low, then it has greater results and more creativity... But on some models, It needs a great strength. So fitting models like 0.3 is wonderful, but on hard models 1.0-2.0 is great. If you catch my drift.
I also usually use a low CFG Scale, it's usually on 2.5... but pro's know when to scale this up and down.
Usually it has problem girl's trying to kiss eachother, and both having the same expression or tongue out, so i would put prompts to counter it.
Also typical shit is that the flying dicks appear, so try to counter the flying dicks. This is pretty much why i made this Lora... to counter flying dicks.
I think the best way to show prompts and result is by sharing a post with some details, take a look at this... these are all using the same negative and positive three in a row: https://civitai.com/images/66296182
My strength in these images are very high. But recommend lower it to get more interesting results, tho the failure rate is increasing.
@dabbit hey I appreciate the detailed response! So yeah, I tried those settings and that pretty much nailed it with my own model I've been training. 81 steps felt high to me but it worked. I found this chart recently thats helped a lot and Euler A is still the best at that level for accuracy so you're right there, you might find it useful:
https://lemmy.world/pictrs/image/c84d458e-1a7b-4cf2-8f74-ea929406e2f1.png
and here's a link to the whole post if you want to read the process they used to gather this data and what the different bordered lines mean (mostly in reference to how some samplers perform in half the time, etc)
https://lemmy.world/post/6034170
But there is some 'beauty is in the eye of the beholder' aspect to it. Like on Pony Realistic models my favorite was always UniPC for the almost grain like effect it gave, but not many people use it. I usually prefer just Euler myself because of the way it handles lighting, some of the samplers really fight natural lighting looks.
What system are you using? I dont ever see +++ or --- kinda things in prompts. It doesnt seem to do anything for me in ComfyUI, I think +++ would be equivalent to something like (cat:1.3) and --- would be equivalent to the same thing just in the negative side.
For the CFG, I usually use around 4.5 myself but from some quick testing with these settings I seem to be getting the best results with 7 and 80 steps and strength 1 on the LoRA. I never thought the steps made a much of a difference but I can see it here mostly in the trees, its a huge difference into the realism of the background.
Also it's kinda funny to be having this CFG/Steps theory convo in the comments of a breastfeeding Lora.
@Fenn Woww! I'm so happy it worked out!
This data you shared explain so much! Thanks a lot! This make so much sense! I have just been running around, testing and been guided by my intuition... So it was good to get some insight.
This "+++" stuff is maybe because I am used to using Invoke. I am just learning about how to work prompts and strength. Here is where i learned it from: https://support.invoke.ai/support/solutions/articles/151000096723-advanced-prompting-syntax
I do have a lot of great results using steps. But sometimes it backfire... And as you show with the data, it's does not seem random that I fell in love with a lot of steps.
When it comes to making the scene better, I usually believe more steps do a great job as you point out with the trees!
Hahahha agree, having this discussion on a Breastfeeding Lora feels kinda crazy - but atleast there is some method to our madness :D
Just out of curiosity I tried on Wan2.1 and it sort of worked. You might consider adapting to a Wan Video LORA
Thanks man! Will look into it!
I had to delete it. It just wouldn't play nice with Juggernaut or Lustify checkpoints. It always gives me girls kissing, or girls that have a double ended dick stretching between their mouths. I tried from strengths 0.25 to 0.98 and no go.
It was vaguely successful with CyberRealistic, but only by typing in the keyword 5 times in a row, and only then maybe 1 in 5 attempts, and even those had degraded faces and skin because of the LoRa strength.