Don't @ me I've been training pixel art for 3 years and just joined the dark side now
I don't think civitai picks up the metadata? Anyway strength 1.0 prompt for example vid was:
integrated_multimodal_description: [Shot 1] Modern smartphone footage from a corporate office, a medium shot frames a cute, 24-year-old brunette office woman (S2) enters the frame. She has a fit build, big ass, big breasts, wearing a white button down-blouse, paired with a short dark pencil skirt and sheer pantyhose. At 00:02.000, she stands up and the camera pans down with medium amplitude at fast speed as she reaches down and pulls her skirt high up to her waist, fully exposing her bare thighs and her innie pussy. She smirks and says: <d>[English] What do you think?</d>
At 00:04.000 a male voice says <d>[English] Can you turn around?</d>
She then turns around and shows her nude ass, completely naked, she is wearing no panties, with her innie pussy peeking out from behind
overall_soundscape: Quiet office ambience with the distant hum of electronics.
non_diegetic_music: N/A
Description
FAQ
Comments (14)
thank you for blessing us LORD
Any trigger? You didn't include an example prompt.
for almost all H3 lora there will not have any trigger word. Due to the strict prompt template this model need to follow.
Just analyze the video with Gemini or any video vision analisis, then copy paste the genrated prompt to another llm that will follow H3 strict prompt templatestructure.
Sry I expected civitai to pick up the metadata from the vid but I guess that doesn't happen. I've edited the lora description to include the prompt for the lora showcase video. It's 0.7 megapixels, 35 steps, beta. Lora strength was 1.0
@Pat3dx well, in this case, they trained "innie pussy" as a concept. Gemini wouldn't have come up with that.
what's the weight + trigger word + prompt?
Having good luck with this with Ref2V at 0.6 strength. Too high and it destroys the quality.
I found out if u using a turbo lora the turbo lora needs to load always at first and then a lora like this here because else it destroys the quality really hard.
Very good! A lot of ordinary vagina LoRAs easily produce ugly and deformed results, but only the Innie Pussy one doesn’t break easily. Looking forward to further optimization in the future!
The Goat!
This is pretty damn good. I've tested it with a lot of different angles and concepts using T2V and it holds up for all of them. It also doesn't seem to alter the prompted facial/body features of the subject in any noticeable way. And as a bonus the actual innie itself can be prompted for additional details, such as a freckle, mole, pubic hair, a piercing, covered by underwear etc.... This was never possible with wan 2.2 pussy loras, that particular detail of your prompt would basically just get ignored.
Has somebody tried it with I2V?
why should someone answer your question when the answer is obvious?
it works with First Last Image as well as References.
Thank you for your service