Brutal Dildo
The field where I am writing this right now has a label, simply "Description*". The label also has a description... you know, that little line of text below a title or label. The description for Description reads:" Tell us what your model does"...
Well... I'd really rather not, but you know what they say! A picture says a thousand words, and videos, I assume, are at least as wordy... so just look at the showcase, okay?
If you download one of the showcase videos and drag it over ComfyUI, you'll get the prompt and a workflow.
Versions
WAN22-HIGH-LOW
I have no idea what I am doing.
Three.WAN
You thought I was done?! HA!
I put a better action phrase into the captions and simplified them for a win win of better tighter... captions. I tested a few epochs and got up to 47, haven't tested past it yet, this thing is magical.
Revised Wildcard Prompt
A beautiful {Russian|French|German|South Korean|Dutch|Swiss|American|Swedish|Austrian|Ukrainian|Portuguese|American|Canadian|Colombian|Australian} woman is using a brutal dildo inside her {vagina,vaginal dildo riding|anus,anal dildo riding},
She is naked{ except for {{black|pink|red|blue|green|orange|yellow}socks|a bracelet|a {black|pink|red|blue|green|orange|yellow} {bra|hat|scarf|shirt|tube top|tank top|sports bra|}}|},
{front view, she is {moaning from pleasure|screaming from pleasure|smiling|gritting her teeth with her eyes {closed tight|open wide}}|front view, she is {making lewd faces|smiling seductively|moaning from pleasure|screaming from pleasure|smiling|gritting her teeth with her eyes {closed tight|open wide}}},
she is {barefoot|wearing {black|pink|red|blue|green|orange|yellow} {thigh|knee}-high socks|wearing a {black|pink|red|blue|green|orange|yellow} shirt},
She is squatting, bouncing her hips up and down vigorously, pumping the shaft of the {red|green|blue|orange|yellow|black|white} {horse-penis|scaly|notched|tentacle-shaped|ribbed|spiked} dildo inside her,
{bright|soft} lighting,{ medium close-up framing,||} in a {doctor's office|bedroom|kitchen|living room|basement|bathroom|bathroom stall|shower},
Two.WAN
Same run as One, just a different epoch. I extracted many, tested all, cried, in part from the beauty and in part from the failures and came up with epochs 24, 48 and 53. They kept coming back with interesting renders. 48 seemed hellbent on proper dildo penetration, so that became ONE. Today I work-shopped better prompts and found a real winner. It'll probably unleash 48 as well, but 53 got this massive habit of creating massive dildos, but it was missing some consistency. Well! Pop in (pun intended) this huge (pun intended) Wildcard template update and BAM (pump intended), whoa, holy f--k that's not possible!
Epoch 24 is not uploaded, but I kept it. It has a certain flexibility that 48 and 53 lack, but also produces much smaller dildos, hence it not making the cut. If you're curious though, maybe I'll upload it.
One.WAN
I automated some caption transformations to save time so this run had a couple caption issues, however this epoch was strangely capable. I'm weary of my dataset but finding clean material where models aren't dressed weirdly or covered in tattoos is tough, so I'll need time, precious precious time! :P
HunyuanVideo
Wildcard Prompt
A beautiful {Russian|French|Swedish|Swiss|Latina|Austrian|German|Dutch|English|Irish|Portugese} woman moves her hips up and down while an upright dildo remains firmly inserted inside her vagina throughout the motion.
{low angle, ||} {front view
She is {happy and blushing|wincing|furrowing her brow and grinding her teeth|screaming from pleasure|extremely surprised|moaning from pleasure}. She has {large|medium|small} breasts and a {hairy|shaved} pussy.|back view}
She has {very long|long|medium length} {blonde|black|brown|dirty blond} {straight|curly|wavy|pony-tailed|pigtails|styled} hair and is {squatting|kneeling} {against a simple {space|black|white|red|blue|pink|teal|rainbow|multi-color|tie-dye|mandala} background|at a look out point above a night cityscape|in a modern bedroom|in a white tile bathroom|in a shower stall|in a tent with open door flap showing snowy mountain peaks|in a living room|in a dirty basement}.
She is completely naked {with glistening wet skin from oil and profuse sweat with no tattoos|except for {a colorful crop top|a beret and dark thigh-highs|a vivid color bra|a beanie|a collar around her neck|bracelets and hoop earrings}}.
While the dildo is mostly hidden inside her the whole time, it is {giant|huge|thick|massive} {black|red|white|beige|multi-color|pink|purple|blue|glass} and {notched|ribbed|horse penis shaped|dragon penis|tentacle shaped|smooth|irregular|spiked|bumpy}, evident when her hips rise.Brutality
I'd say this is the final version but I thought Three was going to be the penultimate LORA for this concept. Truth is that even this one has some weaknesses, namely that it doesn't do anal very reliably, especially from the front. However, it will render anal from the front if you add "leaning back" to the prompt, I guess that one video showing that pose made it a rule.
-You lean back, straight up your ass!
This version was trained on a cleaner set with better captions. I changed to 24FPS from 25PFS. 6e-5 was my learning rate with a smoothed loss just under 0.03 for the best performing epoch (90, although 80 was doing really well too). 21 videos total, equally distributed between front and back views for 50 frames per video. Total training time was ~18 hours.
Curve

Other versions
4.2.0
The runner up! As I zeroed in on optimal captioning strategies, this run did the trick nicely with much better faces. Enjoy!
front
All front view, mostly low angle, 5e-5, 256x256 and 432x240 for 50 frames. This one delivers brutally. Try "notched, scaly, horse, tentacle" with various bases. No back views trained at all. Works very well and seems very flexible. The lower resolution might be hurting things like hands though...
back
I finished training this a couple days ago but my tests were not delivering the results I was hoping for. Multiple concepts are harder for LORAs to hone in on and this run didn't impress.
However, this version does the back view very well, while front does front views better, so if that's what you're going for, you can probably skip back. It is what it is...
Three
Because I found that back will place the dildo in front of the woman and not always inside. and not a single one of my videos was not 100% focused on the penetration, clearly something went wrong.
I studied the concepts and dived deep into the subtle effects captioning and prompting provide to this process. I learned even more about captioning, including some very non-intuitive things, all detailed at length in my training guide. I think the fruits of my labour and patience with training a multi-concept LORA have mostly paid off, or maybe my prompts are just better... (Narrator: His prompts were better.)
Prompting Tips
I changed the Wildcard template:
I had to remove "in and out". That was counterproductive.
I had to add "inside her vagina" right after the key phrase, that seems to really have the right effect.
Mentioning that the shaft is "pumping back and forth" disconnects it from the description of the dildo itself, which I think also helped things get rendered in the right places.
Prompt in natural language. Describe the most important parts of your concept right up front, leaving deeper tokens as modifiers rather than scene definers.
Anything you write, the model will try to render, so make sure you actually want to see what you are describing. Sometimes you want to imply things that aren't actually visible or think about how the model sees it, versus what your understanding of the scene is.
Training
My training techniques are discussed at length here.
Disclaimer:
Run
Description
FAQ
Comments (16)
everybody wins!
just tried... holy sh--
I've been testing this out for a couple days now and it rocks! the flexibility keeps me coming back.
@Psk0 seems to work with missionary as well
v4 is awesome! I lowered the lora weight since the bodies get kinda specific at full strength, but the adherence is improved big time. the faces improved dramatically too, thank you!
Hitting that sweet spot between adherance and overfitting can be tough for sure, but I rarely run a good LORA at strength 1.0, typically we want the model to have some wiggle room, so 0.9 or 0.8 with solid prompts should be best :)
Glad you're liking it :D
@az420 hell yes I am! move aside walter white, we've got the real goods here hahaha
v4's videos wont show, so cannot download them to get the workflow
try now!
I was looking through your training data. Can you tell us or everyone following you. What tool did you use to crop and set the duration on the videos? I'm assuming you just downloaded some nice videos and edited them.
I typically cut to the best 4 seconds I can find and then tend to only use 50-65 frames. However I've also been cutting close-ups of those videos for further detail support and to let me crop away from faces for a portion of the training set. Close-ups don't require as many frames... 30 is usually plenty. Stay tuned and happy training y'all!
I found this works pretty well as a LoRA in FramePack Studio, for local generation, at 0.7, if you use the F1 generation mode. I use SD to make an image to base the video on first, then use the image as a starting point for the video generation. Of course, as with all of these models, the seeds are hit and miss for quality/action in the clip. Here are some of the seeds from other LoRAs/generations I've used that worked ok/really good with this LoRA, and those settings, doing a 6 second test clip with whatever gender you want, without changing the gender/intended content part way through.
11948 - 17184 - 18798 - 31337 - 31373 - 654632654 - 1410209313 - 623244689606540
Some seeds work better than others, and your results may vary. Hopefully those seeds and methods help people enjoy this LoRA to it's full capability.
Thank you for making and posting this. I just started messing with local machine video generation a short bit ago now that I have a system that can handle it, but I've found very little in the way of community support for expanding on the base models. People like you making compatible LoRAs has expanded the possibilities of local generation so much, and made it more interesting to boot!
Alright folks, so I've made a discovery. The main reason the model is rendering the dildo outside the body is because we're describing it. Yep.
Removing the description of the dildo itself generally fixes this... but that's not good enough, so...
Move the description to the end of the prompt, using this language:
While the dildo is mostly hidden inside her the whole time, it is {giant|huge|thick|massive} {black|red|white|beige|multicolor|pink|purple|blue|glass} and {notched|ribbed|horse penis shaped|dragon penis|tentacle shaped|smooth|irregular|spiked|bumpy}, evident when her hips rise.
The tokens being further down and made to be conditional on movement, ambiently talking about what it CAN look like, is how we get the model to stop trying to render the whole thing. And in hindsight this makes sense, because before we were saying
"this dildo is inside!" and then also "here's what it looks like!"
So the model was like "ok I guess it's not really inside then if we know what it looks like" and voila, contradiction applied! XP
sweet thanks az, will be testing v soon!