I this V2, tongues were added to the Blowjob Sandwich.
This LoRa was trained on material made in V1.
Training Caption was "TWO STUNNINGLY gorgeous 18-year-old girls
performing a blowjob sandwich, a cooperative fellatio by kissing opposite sides of a cock and sliding it forwards and backwards between their mouths, their wet, fluttering their tongues up and down on the cock, sticking out their tongues, drooling and licking the cock."
STR=1.
I tried getting the tongues fluttering like butterfly wings on the cock and girls smiling sweetly at the camera, but no luck so far.
I noticed that the skin tended to get a bit too shiny, so I suggest adding a realism LoRa.
Also, I recommend giving the girls a description, as they tend to look very much alike otherwise.
I have been experimenting with lower str. I get ok results, sometimes with str=0.65 w fl2va.
Lower is still licks a cock, but no longer a sandwich. Also, I used reference instead of just text for this.
I used this Prompt for the latest post:
"Shot 1: TWO STUNNINGLY gorgeous 18-year-old girls
in <Picture 1> AND <Picture 2>
performing a blowjob sandwich, a cooperative fellatio by kissing opposite sides of a cock
in <Picture 3>
and sliding it forwards and backwards between their mouths, their wet, fluttering their tongues up and down on the cock, sticking out their tongues, drooling and licking the cock.
Shot 2: They give the camera a sweet smile with smiling eyes."
I had better luck with fl2va than ref2va, even when I treated fl2va as references.
Description
FAQ
Comments (18)
All these women are twins making the exact same movement at the same time?
Is it possible to have slightly different movements or something?
You are right. I have tried, and there were more differences in the original training material for V1. I want to give it another go, but first I want to try to see if I can make them flutter-tongue & smile. This time, I was just set on getting the missing tongues out.
this is just difficult in h3 in general. I really tried to get characters to speak in a canon, for example, and I just couldn't get it. Definitely different, but similar
I think this could be just a version update to the other one and not a new model.
And with the other one the women on both side always look the same. This model looks like it does the same thing. The women need to be different.
You are right. The girls were specifically told to be different, so I will have to think about how to go about it.
@CyberCream its a great concept and i really hope you can figure it out! really want this lora to work! :)
i've never trained a video lora before (my comp would burn lol) but is your dataset flux generated images or something? could you not just train it with different girls on each side? or use real videos from the web to train it on?
@Knoxxxonk I had different girls on each side in the original training material.
Is it a prompting issue? Have you tested describing the girls with different appearances?
@d_A_v_I_d I certainly have. It did not make a difference.
@CyberCream Okay. Must be a dataset or captioning issue then.
This lora works with ref2va...just use two ref images with different girls to get different girls. I've done that, works great.
@ltcdrbroccoli I tried with the previous version and it always just chose 1 of the 2 girls and put her on both sides, or really mixed them together. I'll try with this new version maybe will have better luck
@Knoxxxonk @Knoxxxonk I am afraid they both suffer from the same problem
@Knoxxxonk hm, that's interesting. I'm using wan2gp, btw, not comfyUI. I had zero problems just prompting this girl on the left, this girl on the right, boom, no identity mix ups. I can't post cuz they are private use likenesses but I can attest this lora works with 2 different characters. But your mileage may vary. I'll try the new one out also...
@ltcdrbroccoli Yes, that is very interesting! Wan2Gp is new to me. This needs a follow-up
@ltcdrbroccoli So... You are saying that a LoRa that I trained for Minimax H3 works BETTER with WAN22, despite it not even being supposed to work, since it is a different model?
@ltcdrbroccoli Could I see an example? There must be non-related material somewhere, perhaps you could take some from here on CIVITAI to test?
I just tested a WAN22 with this LoRa and it did not work.