LoRA for two women performing POV oral sex on the viewer.
It's not quite as good at perfectly following directions as I'd hoped (though it's about what I expected). However, it usually produces something decent even if it gets some details wrong. Sometimes it'll do something weird, but I'm mostly happy with how it turned out.
Strength seems to be OK from 0.7 - 1.0. Maybe try something like 0.9 to start.
Some phrases you can try, though it won't always listen perfectly:
blowjob. Two women are {kneeling|leaning over} and positioned on either side of a man's erect penis.The woman on the left ...
The woman on the right ...She is moving her head up and down as she takes the penis into her mouthShe has her tongue out and is licking the shaft of the penisShe is licking the man's testiclesBoth women are sticking out their tongues and licking the shaft of the penisThe man's lower body is visible, showing his nude stomach and legs{The view is POV|The view is from a low POV angle|The view is POV overhead from above}Description
FAQ
Comments (54)
I don't even have to try it. Looks amazing. Thank you.
wow glad there are people making good loras!
still waiting for an ass-to-mouth
double-penetration
cumswapping
lot of positions left just scratching the surface here
This one is basically perfect. Good job.
I agree. i ran some tests and i am getting perfect result like 95% of time. doesn't get better then that.
Probably the best lora I've seen so far. Just instantly great results. Whatever was done here, people should copy and build on it.
Edit: One issue I do notice, other than now-and-then double-cocking: floaty mouths. The cock will just kind of glide into the mouth of the girl like it's clipping in 3D.
But the girl quality is pretty great.
wdym? a lot of ppl complaining about earrings and tatoos on my lora and this is the best of the best? cmon! XD
this lora is so great that even duplicates penises! XDDD
i guess youre just getting lucky with the seeds youre using,, i bet this can produce bad results, as every other lora ..
@K3NK I'm sure it can produce bad results, but I've just been having very good luck. Maybe it's the seeds. Maybe it's the lora stack I'm using -- I'm not running it solo, I'm combining it with a couple others.
Not to knock other loras, yours included, which I've also used and liked the results of. What mostly impressed me here is that I instantly got two girls, one cock, in some creative settings/with good outfits and looks. I remember when that was incredibly challenging just to get properly on stills, so seeing it in video is a plus.
@Lurkmoore the author did a great job captioning the dataset thats for sure
it's the fact that i can type in exactly what kind of woman is on either side and the output adheres to that 99% of the time. it's really solid
i agree best i've seen so far here. i have only maybe 5-10% at most have the double issue, other then that i haven't seen a lora do this well. All the other single bj ones i've tested also has a very low success rate at anything realistic and smooth. This one is perfect at this specific job.
no hardcore clips on the dataset? -5 points
good work on captioning, any advice, tools or workflows youre using to make those?
So I'm still captioning manually, but I did a couple things to try to improve them. One, I extracted the 15th frame from every video, and then ran joycaption against those. I then referenced it when doing my manual captions. They're still largely unusable directly, but sometimes it would come up with a phrase or notice something that I wouldn't have thought to put into my manual captions, so it can help.
As an example, I think the phrase "positioned on either side of a man's erect penis" came from joycaption.
Another thing I tried to do was run prompts without any lora present to see what words would get closest to what I was going for. The less your lora has to fight against the base model, the better. An example of this was using the word "testicles" instead of "balls", etc.
@dtwr434 i just downloaded joycaption-batch, and LLaVA Video Captioner but no success with the last one, the joycaption does quite well as far as i saw, the only thing is the tendency it has to describe everything as a photograph , a picture,, and so on.. 😂
i assume you used video for training? what resolution and duration? i have to try with 322x322x24 10sec clips with my 4090
@K3NK Yeah, that's why I'm saying you can't use the joycaption stuff directly. It's just a reference to help you word certain things, or get an idea of what things to include in the caption.
My video files are 3 seconds at 480p, however I specify 244 as the resolution, and 50 frames in the config file. The important part is making sure the movement you want is clearly visible in the first <50 or whatever> frames for each clip.
I added "Do not mention that it is an image or photograph" to the joycaption "content" prompt and it does adhere to that. Still gets a lot of other stuff wrong. I tried out Janus-Pro-7B also - it's not censored against NSFW but joycaption is still better, Janus will just say "sexual act" with no detail.
@logenninefingers888 Yeah, another thing I should mention is that when picking a phrase, I look at several joycaption outputs and then decide on something. And then I try to use consistent phrasing across all of the videos. If you used straight joycaption for everything, you'd get much wider range of phrases for the same basic thing.
I don't know if that's actually a good or a bad thing, but I figured being more consistent would make it easier to prompt for later.
@dtwr434 Thanks for sharing all this info. Did you have the training data hosted for a bit? Can't find it.
Also, curious. What is the actual frame rate you're doing for the 3 second clip, if you're putting 50 in the config?
@KinkMaster I didn't include any training data for this one since it's basically the same as other ones I've done. I use 24fps videos, so I basically am not using the last bit of the 3 second videos. I suppose I could use something a little over 50, but I'm not sure it'd be worth the extra training time. At least for these simple movements.
@logenninefingers888 you mean editting the .py file on this setting?
# Captioning Options and settings
# Prompt used for captioning
PROMPT = "Please provide a detailed description of the image. Do not mention that it is an image or photograph."
this one is nice. works straight away without having to wrangle with prompts. good shit
This LORA sucks twice as much as usual... keep it up!
Anyone have an example of hunyuan with multiple people loras? I can't seem to get the hunyuan to create two different women using triggers.
I've had a little bit of success with training two separate women in one lora, but that's about it. Seems like no one else is interested in figuring this out besides you and I haha.
All the techniques you find out there for getting consistent two characters from multiple loras use inpainting after generation. Its a shame, when I put two loras together I ended up getting a mix of the two characters with terrible noise. Ive tried training on a specific word too omitting things like "woman". I wonder if there is an inpaint tool for video to video.@ablestable69420
@0226 Just tried my double woman trained lora with this and it works perfect tbh, the key is to have them in the same lora training with separate triggers and descriptions
@ablestable69420 Do you have any training examples of how to train more than 1 character in a lora? what does dataset preparation look like? Do the people need to be in the same image? Do you merge two trained loras into one?
@0226 x2
@thelastofus94 @0226 so when I did it, I had about 17 images of the one person and about 9 of the other. The biggest thing is just the captioning, and making sure you have at least a trigger and a hair color in the immediate caption using joycaption. It was all done together in like 50 epochs, which was about 2.5K - 3K steps.
not bad but the heavy makeup is impossible to get rid of. make sure you caption about makeup so its optional, thanks!
I personally don't see heavy makeup in my test of this. just a normal amount of makeup. All women use a small amount of makeup like eyeliner and such. However if your looking for less AI looking or less perfect characters there is edge of reality lora which may help, i do wish there was more amateur lora's, though honestly i think hunyuan does a better job then img models for realistic looking characters. just with a basic prompt
@zengrath for realistic porn check BigLove checkpoint for pony, the best I seen
prompts like "natural face" etc helped me to reduce makup, expecially remove these giant black eyelashes. not sure if it coninstent to, try on high guidance
such awesome lora! please make more!!
can you please consider making lora for live2d idle animations. You know like in gacha games whem character is moving slightly, with hairs and clothes moving. It would be such a present for live wallpaper community
I think LTXV works great for such slight movements
As many have pointed already this lora is top quality, but sadly it seems to be almost impossible to get hairstlyle shorter than bob cut/ponytail/bun, (something like pixie cut etc). Any shorthair prompt that works easily in other models or even without loras does not work in yours at all and mixing loras is not helping much. If you are going to work further on lora and if you mind this issue, I can help gather related data, just dm me
Hmm, did you try lowering the weight of the lora? I tried a few times with a weight of 0.84 and it seemed to work decently enough, though maybe I don't understand exactly what you're going for. I uploaded some examples in the gallery.
@dtwr434 yes this seems to help. I did lowered weight to 0.8 after many attempts, but at that point I was desperately trying some weird prompts in order to get short hair so it wasnt working because of them. Big thanks for reaching out with examples! However, doesnt that mean that problem is still with lora itself?
@TurboCoomer Do you mean it's a problem that you have to lower the weight? I'm not really sure if I'd consider that a problem. It might mean that the lora was potentially trained with too many steps, but since you can lower the weight, people can choose what works best for them. Over training it slightly with the ability to lower the weight gives people the most flexibility, I think.
However, if there's some good videos out there with short hair, I could consider revisiting this. I may trim some videos at some point in the future and try again to see if I can get rid of the double dick issue. Finding examples of really short hair seems like it'd be the hard part, though. Maybe just captioning the hair better in the videos I already have would improve it.
@dtwr434 I mean if theres a need to lower weight to achieve result, that means that lora is blocking the result, no? So lowering weight allows to get shorthair from base hunyuan checkpoint? Also I think at 0.8 motion starts to degrade?
@TurboCoomer I think a strength of 1 is just kind of arbitrary, and isn't necessarily the "correct" strength that loras need to be best at. Having said that, the prompts I was using still work fine at a strength of 1, so you don't actually have to lower the weight if you really don't want to.
I'm curious if you were actually noticing degraded movement at 0.8 though, as I still figured that'd be a high enough weight for it to work pretty well. I think 0.9 is kind of the baseline I use for this lora.
@TurboCoomer Just my two cents, but LORAs are rarely if ever perfect. I think that changing the weight is a really really common thing to do when trying to achieve different results with LORAs. I've never heard of anyone retraining a LORA when simply entering a different weight does the trick...
This guy here has some of the most versatile HV LORAs I've seen so far btw... some others are extremely rigid regardless of weight.
Question - is there any way to generate animations on own pc? If yes, how?
if you are asking how to generate ai images/videos locally on you own pc. then look into stable diffusion - it's probably the easiest software to generate images on your own pc imo. then Stability matrix, with this you can download a lot of different programs for all your AI needs. But probably the best software imo is comfyui, but this is quite hard to get to grips with as a total newcomer. Be prepared for a lot of research and trial and error to get things going. good luck
@jgillservices377 Yeah, I made it to the guides. Resources eat so much that so far decided not to do animation.
@Twostep Recources eat so much? That's normal. If your rig can handle it, go for it. Installation isn't that complicated, just follow the guidelines.
Would love to see a Wan2.1 version!
Man, I love this Lora, but this conjoined dick is getting on my nerves
I use 2 characters LORA on left and right, but the gentle slow motion bothers me.It is hard to generate as fast motion as the example
Hey iv just made 2 videos, for some reason im getting a error " SamplerCustomAdvanced
Allocation on device" unsure what this means haha, im still very new to making videos. iv just not a clue what it means. iv added a link with screenshots
Finally, a Hunyuan LORA that works on the first try.
Any advice to avoid getting two penises besides reducing resolution? I use 720p. Thanks