UPDATE 1/31/2025
Last merge had to be removed due to model permissions. Remerged with different model.
Merged Loras:
My Edge of Reality Lora - https://civarchive.com/models/1129944
HunyuanVideo POV Titjob / Paizuri by KaptainSisay - https://civarchive.com/models/1114844/hunyuanvideo-pov-titjob-paizuri
Trigger words: [amat_petite], titfucking
Likes the phrase: "giving a titfuck, She squeezes both of her breasts together and puts the mans penis between her breasts she moves her body up and down titfucking the penis with her breasts."
Example prompt in version notes
UPDATE 1/23/2025
Merged Loras:
My Edge of Reality Lora - https://civarchive.com/models/1129944
Hunyuan Video handjob Lora by cavemanman333 - https://civarchive.com/models/1118309/hunyuan-video-handjob
Trigger words: [amat_petite], handjob, stroking
Likes the phrase: "a woman giving a handjob to a man stroking it up and down with her right hand" to generate a handjob.
Example prompt in version notes
A LoRA designed to help style amateur videos with unique options for body types, clothing, and small details. Trained off of an old dataset of from a FLUX LoRa I left unfinished. I didn't change any of the captions from the FLUX dataset and honestly I'm shocked how well it came out. I ran it as an experiment expecting to throw it away and needing to re-caption.
This LoRA is INTENDED to give a slightly lower quality output more similar to cellphone videos in general.
Civitai Video compression did a number on these videos for some reason. Weird coloring coming through... They look better if you click the image vs just the preview window.
Example of non-strange coloring: https://files.catbox.moe/m5zxii.mp4
Main Trigger Word: [amat_petite] (trained with the [] so they may be necessary)/
See all training captions here: https://files.catbox.moe/rwj3po.txt
What it's good at:
Amateur video styles such as cellphone video and selfie angles
Petite body types
Skinny, thin, shirt, tall, can get others out of it but it will skew towards petite.
Wide range of breasts
sizes but skews towards small breast sizes, has a harder time with large breasts
Various poses
Kneeling, Standing with arms above ahead, reclined on bed/couch, running hands along body through hair, squeezing/holding breasts, hands on hips, general photography model posing, side profile
Tan lines
Very good at tan lines, lots of examples in the training set
Skin Shades
Trained all on Caucasian subjects but you can get a good range of skin tone from pale, average to tan.
Braces
The original dataset had a few training images with braces and honestly they work pretty well, can be a little noise if the video shot is too wide
Clothing in pulled down/up states
But it's not great at the actual motion of pulling it
Hair
In general the motion of hair is pretty good, even if the subject is interacting with it
Different lighting types
Best at natural light, should understand: "outdoor light, natural light, indoor light, overexposed
What it's okay at:
Facial/Cum on Face
There were a couple training images of it, IMO it's a little too thick and doesn't flow well but it's an option
Ahegao
Not captioned as Ahegao but "mouth open with tongue sticking out," gets you there
Rear Facing
Okay with the subject facing away from the camera but there wasn't much in the training data
Adjusting clothing/underwear
Can have the subject slightly adjust their underwear without unrealistic movement
Closeup shots of breasts with head out of frame
What it's bad at / Issues
All the backgrounds were photoshopped out of the training images which leads to the model outputting the subject with a plain white background (which could be interesting for chroma keying the subject for video edits)
This can be avoided by describing the background in detail
It also sometimes leads to the subject looking like it's unnaturally dropped into a static background
If you use the word "photo" you will get an static video or in very slow motion.
This is likely because I captioned the images with things like "professional photo", "cellphone photo", "home photo"
including the word video in the prompt helps
Actually removing clothing with motion
Tends to make the clothing really twisted and strange.
Occasionally puts in tattoos but I don't recall any of the training data having tattoos, none of the training captions included tattoos. I've noticed this effect in other LoRas I've trained as well. I think Hunyuan training occasionally interprets bracelets, clothing logos and necklaces as tattoos, especially when small and not captioned
I've included a ZIP with a TXT that has all the training captions so you can see what you're working with in terms of prompting
Description
Epochs: 30
Total steps: 25200
Steps per epoch: 840
Trained on 168 Images with backgrounds removed.
FAQ
Comments (25)
Civitai Video compression did a number on these videos for some reason. Weird coloring coming through...
Example of non-strange coloring: https://files.catbox.moe/m5zxii.mp4
Yes, I believe it mainly happens with portraits that are longer than the native thumbnail aspect ratio.
ooh, nice. can't wait to try it out. hunyuan really doesn't like to have that low-quality/amateur look regardless of prompt so this could help.
Come back and share some results if you are willing
I agree. I like to generate realistic people and sadly Hunyuan always gives me that AI look everytime, especially if i try any of the Asian ethnicities i get often an AI specific look with plastic like skin and absolute perfect doll like bodies which honestly turns me off. i don't like it. I have had to play around with prompts a lot to get it to generate more normal looking women. But i'll try this lora too to see how it does. hopefully it can be combined with others that do things other then HJ's.
@zengrath If you plan on mixing it with other Loras and don't want to generate Handjobs, I recommend using the standard version of the lora. The merged modal is heavily influenced.
so is this trained on images ore video, out of curiocity?
Images.
To expand on this. My first attempt on training Hunyuan Loras with videos. I got much better results of the same concept trained just via images. That's not to say videos can't be better but so far images are much easier. I am about to do another training on videos that I'm sure wont work with images so we will see how that goes.
Which trainer did you use for this?
@TheAIDoctor Thanks
You should add the prompt to some videos
There is a file in the description that has all of the training data captions, which I used to generate all of these examples.
@TheAIDoctor Yeah but to make it easier and see directly the prompt of the video you like. It's a recommendation because a lot of people will download this LORA. I'm testing it right now, thanks for uploading it <3
@AIVideoEnthusiast Fair enough. Problem is I generated about 200 of these overnight using the random prompt feature which was just all of the caption prompts combined so I'm honestly not sure which prompt was used for which video at this point.
I think this is partly overtrained. I often have the case that the animation freezes. But it's still good thanks
As stated in the description, I believe the problem comes from the captioning as I re-used the exact captions from a FLUX training set that included the words. "photo" repeatedly in various contexts. Even when the animation is frozen you will often notice subtle movements.
But I have some lower epoch versions I could try out to see if it's a training issue.
All that being said I plan on training a new version of this with improved captions, the original images without the backgrounds removed and some fresher more HD images in the coming week or two.
I've been able to get some excellent movement using this Lora. Have you tried lowering the strength? I've had good output at 0.8.
@voyeur19 normally I keep at 1.0 but it helps to lower it if combining with other Lora's or trying to get more motion.
I have same problem with nearly every lora including no lora at all. every now and then Hunyuan produces a frozen video that has 0 movement at all. I doubt it's specific to this lora.
I also have seen the motion freezing issue more often with this Lora in particular. I ran some comparisons to be sure. I still like the effect though and use it often, just not at full strength.
@vim_brigant I'm working on a retrain that has improved captioning and some new training settings so hopefully we will see an improvement going forward.
@TheAIDoctor Thanks for your work on this, and for continuing to improve it!