Not everyone agrees that size doesn't matter. Would you rather date someone with a paddle board or a yacht?!? but I digress...
Fisting
When a woman loves a woman... they show it to each other with an iron fist! I couldn't find videos showing that though so I settled for regular fists.
I tried training this one with nary more than a key phrase and generalization tokens, which helped me understand a little bit more about how training works.
Minor seed hunting, although it's pretty darn stable and the mutations tend to be slip under the radar at first glance. Only other thing I've seen is the fisting itself not quite happening but this is rare and probably fixable with prompting. I haven't prompt-workshopped this one too hard yet.
Versions
WAN22
The WAN2.2 version has a solid "from behind" modifier you can put right after the key phrase. I don't have it here because it interferes with facial expressions, putting heads on backwards if you prompt both "from behind" and an expression. One of the reasons I've moved to WAN from HunyuanVideo is the lack of mutations. I really don't like mutations! I've seen things... things no man should see lol
Wildcard Prompt
A {Russian|French|German|South Korean|Dutch|Swiss|American|Swedish|Austrian|Ukrainian|Portuguese} woman with {blonde|black|brown|dirty-blonde|dark red} hair is being vaginally fisted by another person {from behind|},
# Don't use from behind AND prompt to see the face...
She is sitting on a {red|blue|yellow|white|black|brown|teal|pink} {chair|couch|bed|dentist chair} {with {large|medium-sized|small} breasts|wearing a {red|blue|yellow|white|black|brown|teal|pink} {shirt|crop top|tank top|bra|jacket|sports bra}} and a shaved vagina.
{the person fisting her is rubbing her clitoris.|}
She is {happy and in love|moaning from pleasure|gritting her teeth and furrowing her brow|closing her eyes with her mouth wide open|extremely surprised and shocked|giggling and laughing|smiling seductively}.
The person fisting her is a {man|woman|doctor|nurse|firefighter} and is thrusting her entire hand back and forth inside her vagina.
{bright|soft} even lighting, {high-angle|} viewThree
Removed the slower clips, so motion is improved. A caption clean up seems to be giving better stability as well. Enjoy!
Two
Blurred faces, 24fps, 50 frames, 5e-5, dropout=0.1 and I can't remember the rest because I trained this weeks ago and only now decided to get showcase materials and release. This one is more solid than version one for sure.
Wildcard Prompt
A {slim|fit|sporty|cute|skinny|beautiful} {Russian|French|Swedish|Swiss|Latina|Austrian|German|Dutch|English|Irish|Portuguese|Indian|Russian|Swiss|Swedish|Danish|Italian} woman reclining with her legs spread as a person on the the left is rhythmically fisting her vagina. The motion is {quick|steady|slow and rotating inside}.
{Wide angle view showing the {male|female|Doctor} partner|An arm extends from the {bottom left|left}}.
She has long {blonde|light blonde|dirty blonde|dark|black|brown|light brown} hair{ in a ponytail| tied back| in pigtails}.
She {has {large|medium-sized|small} breasts|{is wearing a bra|is wearing a nice shirt|is wearing a colorful tube top} covering her nipples}.
#The other person is a {lover|Doctor|friend}. {caressing her chest|wearing a {yellow|red|blue|gold|silver|pink|teal|orange|purple|rainbow|multi-color} {shirt|bra|tube top|tank top}|wearing a {yellow|red|blue|gold|silver|pink|teal|orange|purple} headband}
She is {screaming from pleasure|happy and smiling|surprised|closing her eyes} with her hands {in her hair|on her chest|at her sides|hidden|holding her own legs}, {sitting in a gynecology chair|sitting on a chair|lying back on a bed|sitting on a couch}.
Lighting is {bright|soft} and evenOne
The original attempt, works pretty well but surpassed by Two.
Training Tips
A LORA is basically a subset of a model that rewrites "weights". The weights represent pathways into the model's architecture in terms of how tokens call up visuals. The training process reads your tokens and scans the visuals in order to nudge the weights in such a way that they more closely reproduce what your training material shows it.
If you never caption for hair color. No weights involving hair or color will be written into the LORA, and as such during training, you don't get access to those weights and thus they either get baked in, or reduce your ability to prompt for that kind of thing in the base model.
If all your models are fat, but you never caption for body type, you're not going to be able to prompt for fit woman, they'll all be fat. The LORA will assume that's what the concept is, fat women, and no weights related to body type will be touched or touchable. So... you caption 'fat woman' just a couple times to get those weights involved in training and thus involved in prompting. That way, the LORA/Model pair have something to work with.
If you don't plan on prompting for things you see, you can generally not mention them at all, but then they might show up unintentionally. This is why a varied dataset is important, because if visually a few videos show the exact same model or background, the training process will lock onto those features and you'll have trouble escaping them.
If your LORA involves poses, you want a dataset that does not deviate much, otherwise you get body horrors and mutations, no amount of captioning is going to let a small file that modifies a massive file fix anatomy.
Disclaimer
Apply lots of lube and go slow.
Description
FAQ
Comments (12)
Nice! Any plans to expand this one or just a one and done for the community?
What do you have in mind? I did chop a few doggystyle fisting examples, but, you need multiple videos of almost the exact same pose to get a really good generalizing LORA, in my experience :/
@az420 I had nothing in mind specifically but with this particular act there's lots more than just position, such as number of insertions, etc etc... it also seems very limited in motion, although as always with your loras it's super high in quality and I wouldn't presume to add any critique as I am currently incapable of doing the same as I don't have the equipment, experience, or knowledge 🤣
I suppose I was asking because you made several variations of the brutal lora and this seemed slightly limited.
Hey, so definitely not a one and done. I trained a new one but it was worse... since I've added a couple nice clips and modified captions/key phrases, training now.
I have a lot more footage as well so I might be able to expand this to allow for some clitoral stimulation. It's a mixed bag finding new good clips because it's like "damn, ok, how many other better clips are out there" and then finding and clipping them is slow going.
I'll also probably do a doggystyle fisting one.
training note:
fingers are poison! I had one clip where there were some fingers appearing behind the thigh, that run rendered extra fingers everywhere. Clips need to be SO clean to be usable, one bad clip can be toxic.
@az420 I wish I was able to train but have only worked with Pony training before and that was a nightmare as well, never seemed to take or was just a mutated mess, and I'm such an amateur I haven't the foggiest where I went wrong lol 🤣. I do have access to a 3060 12GB with 64GB of RAM and an intel Core-i7 12700KF, but wouldn't even know where to begin lol. I'm pretty technically minded but getting ComfyUI even setup was a challenge because all the guides and stuff I've found are either intermediate or advanced so it's been a struggle to understand ground-up concepts since I started with the gradio based SD, that stuff is far easier to learn than the spaghetti mess that is Comfy...
@az420 if I get some time on my hands I'll attempt to read your training guide and try to make heads or tails of it, If memory serves I did look at it before and "it was all Geek to me", not how the saying goes but is a version I have used with my customers for years being a computer service technician with a love for wordplay 🤣 Not holding out much hope, 3 months ago WhatTheGuy tried to explain single and double blocks to me and I had no idea what he was talking about, I tried looking it up but there wasn't anything I could find to explain it.
It wasn't until I loaded your workflow and played with it for a bit around maybe Brutal version 3 or 4 where I finally noticed the lora loader had blocks settings unlike the basic lora loader I was using on the fast hun gguf workflow I use. I do prefer your workflow because it upscales but my 1080 TI is pretty lackluster in the speed department so unless I'm getting good result from a prompt it's not worth me using it even if I disable the upscale portion. I did change it to use a GGUF loader and used fast hun model but I don't understand half the stuff it does anyway lol. If I try using it with the native setup you work with I can't have even a single other thing running on the system 🤣🤣🤣 I need to do a crash course on Comfy from the ground up, then the same with Hunyuan, with all my non-existent free time... Thanks for coming to my TED talk 🤣🤣🤣
@Pity_the_Foo
Yeah, my 4090 makes a lot of the awesome possible. At some point I just said f!ck it and bought a liquid cooled one, no regrets!
If ever you want to dive deeper, I've found ChatGPT quite knowledgeable about AI model architecture, it has sped things up for me in my understanding, even though not everything it says is true.
Single blocks: bring the concept into the rendered image using a single strategy
Double blocks: brings the concept into the rendered image using two strategies (and thus more likely to work with other LORAs)
az420 one thing I noticed is that you must have had at least one double fisting image or video in your training data as it has appeared once or twice while I was testing the model. Any chance you are going to expand both the variety of this concept as well as having increased movement since even version two seems to be movement deficient on the part of the arm and hand?
Pity_the_Foo I can double check clips for any with less movement. No double fisting in my dataset, but my missionary LORA sometimes gives blowjobs, so, I think the base model knows more than we think! ;)
az420 to be clear, looking for more movement, not less lol.
Any plans for a WAN version? 🥺
someday, probably... it takes me too long to render videos with WAN, validation would be a nightmare :S