Experimental proof of concept for lesbian sex with strap-on dildos. Created at the request of a user. Final checkpoint 24 epochs ~5220 steps. Still hit or miss videos. Trigger word is 'str4p0n'. Other words used in training ' 2girls, nude, black harness, strap-on, sex toy, dildo, object insertion, vaginal insertion, vaginal object insertion, anal object insertion, two topless women engaged in a sexual act. breasts, pussy, yuri, vaginal, sex toy'. Use at your own risk.
NOTE: while the content is present. it is very prompt and seed dependent. i'm not very proficient at prompting this style of content. so, i'll keep posting anything that turns out semi-indicative of training clips.
v2.0 changed workflow, 27x3sec clips 256x256 resolution, added some hair pulling and close-up clips to spice up the generation, blurred faces for better lora interaction, 24 epochs 5228 steps.
v1.0 seems a little better. Still not impressed. Maybe i can't prompt for this content... I may redo it with not so verbose captions? Let me know if anyone gets anything nice...
Sample videos can be saved and dragged into comfyui to see prompts and workflow examples. YMMV.
I used this process for local training:
https://civarchive.com/articles/9798/training-a-lora-for-hunyuan-video-on-windows
46x2 sec video clips sampled at 430x430 resolution.
Description
12 epochs 3960 steps. last installment. still having trouble coaxing out content. i can see it's there, i just don't have prompting skills to bring it out consistently.
FAQ
Comments (7)
Nice work. Weirdly I've been finding lower res training (256pt, or [256,320]) seems to get fewer random limbs etc. Maybe worth a try - also trains faster!
Interesting. I may give that a try.
that sure would speed up training. have you experimented with verbose captions vs. sparse captions? my first 2 loras i used simple 20-30 word captions, this one i used a tool to create very verbose captions.
@tedbiv I'm using joycaption so it's medium length i'd say (like 3 to 5 sentences). It needs some manual fixing up though for anything complex, gets relative positions completely wrong sometimes and can't tell anal vs not. I think couple stuff rather than solo is just hard, the model gets confused about which limb belongs to which person haha.
@logenninefingers888 yes, i used joycaption also. there was a captioning workflow that took 3 stills from a video clip, use joycaption and two wd14 taggers, concatenates them, strips out duplicates. unfortunately it adds a lot of 'scene' content, but it is verbose. here's an example used in 'strapon dildo' lora -
'str4p0n, 2girls, nude, black harness, strap-on, sex toy, dildo, object insertion, vaginal insertion, vaginal object insertion, anal object insertion, In a brightly lit room, two nude women engage in a sexual act. The woman on the left has curly blonde hair, light skin, and small breasts. She wears black leather straps around her thighs and a playful expression. The woman on the right, with straight platinum blonde hair, light skin, and medium-sized breasts, is bent over a table, her face in pleasure. The background shows large windows with sheer curtains, greenery visible outside, and modern decor. The image is a high-resolution photograph.,long hair, breasts, open mouth, blonde hair, multiple girls, holding, navel, 2girls, jewelry, medium breasts, nipples, closed eyes, white hair, ass, nude, small breasts, teeth, pussy, indoors, sex, black footwear, yuri, vaginal, high heels, window, uncensored, tattoo, piercing, ring, letterboxed, sex toy, anal, realistic, dildo, strap-on, mole, curly hair, object insertion'
thanks for the tip... it works much better.
@tedbiv great, look forward to trying it!