They sit, we tie, it's all in good fun!
I present... drum roll
Tied Sitting
WAN2.2
Now more... of the same!
ONE.WAN
Found this old set, updated it, then found the rest of it and updated that. Trained, tested, printed, published.
A beautiful {Russian|French|German|South Korean|Dutch|Swiss|American|Swedish|Austrian|Ukrainian|Portuguese|American|Canadian|Colombian|Australian} woman is in a tied sitting pose,
She is sitting on {a chair|a padded bench|a sofa|{a chair|a black pad} in the corner of a room} with her wrists {tied above her head|}{ and knees|} and ankles tied with rope attached to {the wall behind her|a metal pipe frame|the chair}.
She is {completely naked|wearing {a shirt|a white blouse and plaid skirt|a bikini top and naked from the waist down}|wearing bondage gear|wearing jeans and a tube top}.
She is happy with a smile on her face.
{bright|soft and even|dim} lighting in {a doctor's office|an alien laboratory with white walls|a bedroom|a modern loft apartment|a bondage dungeon}. Three
Third times a charm! Trained on new material(s) and an all around cleaner and better captioned set, now you can just sit back and relax, and not move until someone unties you.
Prompts matter for this one and some seeds do not agree with the concept but when this version gets it right, as it often will, it gets it so right.
A beautiful {Russian|French|Swedish|Swiss|Latina|Austrian|Dutch|English|American} woman is tied sitting {in a chair, {her legs down and open with her knees held apart and her ankles tied to the chair legs|her legs spread wide and lifted, restrained at the {ankles|ankles and knees|ankles, knees and thighs} with {rope|straps} attached to {the wall|a horizontal wooden beam on the wall|a metal frame}}|on a black padded bench with her legs spread {and lifted |}tied with rope to {the wall|a horizontal wooden beam on the wall|a metal frame|hooks on the wall}|a wooden panel on the floor with her legs spread wide held at the ankles by rope tied to the panel in the bottom corners of the frame|bound to a metal chair and tied with rope to a black metal pipe frame|on a {black|red|white|blue|yellow|teal|multicolor|corduroy|tan|orange|green|cyan} couch tied with rope coiled around her ankles, knees, wrists and arms secured tightly to various points around the couch} {in a bondage room|in a living room|in a basement|in the corner of a {red|white|blue|green|rainbow|teal|pink} room}. She is held stable in her pose attached to the fixtures.
Her arms are {at her sides with her hands hidden behind her thighs|up with her wrists bound above her head with rope attached to fixed point outside the frame at the top|straight out to the sides with ropes coiled around her wrists{ and arms||}}.
The view is from the front and the pose is symmetrical.
She is {wearing a {red|blue|black|yellow|orange|multicolor|rainbow|gold|leather|silver|latex} {blindfold|pair of panties|bra|dress|bodysuit|hoodie|pajamas set|costume|volleyball uniform}|naked with {huge|large|medium-sized|small} breasts and a {hairy|shaved} pussy}.
She has {very long|long|medium length} {blonde|black|brown|dirty blond} hair.
She {has a blindfold over her eyes|has a big smile on her face|has a {red|black} ball gag in her mouth|has a shy expression|looks concerned and lost}.
Lighting is {bright and even|soft and even|dim in the background with a spotlight on the subject|provided by lots of candles}.Two
Better performing version of this concept, fed with partials, close-ups and more accessories. 5e-5 for ~13 hours. ~ 256, 256, 48. I test the loss curve at the beginning, middle and end of obvious plateaus, in this case the smoothed loss at epoch 70 was ~0.031.
Wildcard Prompt
Front view of a beautiful {Russian|French|Swedish|Swiss|Latina|Austrian|Dutch|English|America|California} woman with {very long|long|medium length} {blonde|black|brown|dirty blond} hair tied sitting with her legs spread on {a chair|a bench|the floor|a pad|a pillow|a gynecology chair|a dentist chair} {in a bondage room|in a living room|in a basement|in the corner of a {red|white|blue|green|rainbow|teal|pink} room}.
She is restrained at the {ankles, knees, wrists and thighs|ankles and wrists|ankles and thighs|wrists and thighs with her feet outside the frame at the sides} with {red|beige|black|white} {rope|chains|straps|cuffs} holding her tightly in her pose.
{Her arms are at her sides with her hands hidden behind her thighs.|Her arms at up with her wrists bound above her head|The restraints wrap around her {belly and thighs|ankles and wrists|ankles and thighs}, securing her firmly restricting any movement}.
She is wearing {{red|blue|black|yellow|orange|multicolor|rainbow|gold|leather|silver|latex} panties and a {red|blue|black|yellow|orange|multicolor|rainbow|gold|leather|silver|latex} {bra|top|shirt}} and {red|blue|black|yellow|orange|multicolor|rainbow|gold|leather|silver} socks.
Her skin is pale and her body glistens with moisture in the light.
She {has a blindfold over her eyes and {red|black} ball gag in her mouth|has a big smile on her face|has a blindfold over her eyes|has a {red|black} ball gag in her mouth|has a blindfold over her eyes|has a shy expression|is terrified and screaming}.
Lighting is {dim in the background|bright and even|natural|soft and even}.One
~8 full body videos, ~6 ones including the Hitachi magic wand (reproduced with limited success) and a bunch of close ups to teach the right details, this training run is worth the post. I may add more videos and retrain this at some point to try to improve the details, but until then, enjoy the one, the only, the... One.
add the second trigger term "v1bed" and a phrase "holding the h1tach1 magic wand vibrator to her vagina" for the secondary concept.
Potential Future Directions
I have a dream now, of training on my entire BDSM training sets, expanding to various positions, restraint types, settings, clothing, etc... some sort of BDSM Uber LORA. I think that my footage and captioning skills are where they need to be, but it's likely that I'll need to run a very low learning rate for 2-3x as long as I usually do to get something stable. Meanwhile, I'm going back through the footage I set aside looking for more good clips to bonify the datasets; this will take lots of time and captioning needs to be accurate and strategic for results to be good, so I can't rush this at all.
I may eventually attempt some of theother conceptspeople have suggested (sex positions++), but many of these already exist and I'm less keen on producing more of the same.There are many other BDSM concepts that I'd like to train, but finding high quality footage of clear and focused examples of the more obscure ones is difficult at best.I also may eventually train forWAN, but I don't use that model and the time it will take to prep and run those trainings will cost me lots of time and energy, ultimately for LORAs I won't personally use at all, so we'll see. *What I am open to is sharing my datasets. I'm not the possessive/selfish type at all. The only thing is that I have spent tens of hours collecting and captioning these sets, so I reserve the right to be be selective about who I share stuff with.
HunyuanVideo doesn't love being trained on multiple concepts. In my experience this is more likely to result in Frankenstein's BDSM LORA.
Apparently WAN does though. 🤔
Description
FAQ
Comments (14)
Awesome lora. do you also have a plan for pulling hair? which is extremely difficult to collecting dataset. 'pulling hair' never works on any video model, especially T2V ones.
I thought my side doggy LORA had some hair pulling action in some seeds. TBH I haven't come across much hair pulling in vids, but then it's not something I've searched for either.
What exactly did you have in mind for this, matched with another concept like doggy or just hair pulling on its own?
@az420 I intend to match 'doggy' motion with 'hair pulling' while significantly improving the quality of the hair pulling motion itself, including SFW contexts. Most existing results only depict characters touching their hair, lacking the physics where the head is actually thrown back. Even prompts like 'their hair is pulled as their head is thrown back' have proven ineffective.
I have several SFW and NSFW clips for this, but since I’ve only trained character LoRAs using images, I’m unsure how to train these specific motions without breaking character consistency. Regarding the 'blurred face' or 'apple-covered face' techniques used in Flux body LoRAs, how should those be captioned? For instance, would you use 'a person with a blurred face is dancing'? I am trying to gather as much information as possible before starting, as testing on RunPod involves costs
@fluxxes I use 'face censored and blurred' exclusively
My rough caption template now is super minimal but focused
[subject] [key phrase]
[factual description of subject(s) position and/or action IF relevant to what you want to train,]
traits, clothing, face censored and blurred,
[]adjective lighting
i.e.
A woman in a tied suspended upside down pose,
Her ankles are strapped and attached to the ceiling, additional coiled rope around waist,arms behind back implying restraints,
naked, face censored and blurred,
bright lighting
@az420 Thank you for info, Does this mean I don't need to manually blur the faces in the dataset with other vid-editing program if I use these captions?
@az420 I apologize for the frequent questions, but I have one last inquiry. I recently attempted training using the RTX 6000 Pro on RunPod with the AI-Toolkit template, using about 15 video clips, each less than 5 seconds long. When I set the Num Frames to 81, I encountered an OOM (Out of Memory) error. Lowering it to 65 allowed the training to start, though the GPU load reached 99%.
I previously saw a tutorial suggesting 5-second video clips for training with an RTX 5090, but I find it difficult to understand how such short clips are processed without crashing. When you prepare a dataset for video LoRA training, what is the typical duration of your clips? Also, if 5-second clips are the standard, what kind of GPU or specific settings are required to handle that load without memory issues?
@fluxxes You do need to blur the faces, yes. I use https://github.com/ORB-HD/deface
@fluxxes The number of frames and resolution combine to sum up to a MegaFramePixel value. I've found value thresholds that coincide with 24GB or 32gb of VRAM
I created this script to take a folder of clips/captions, convert to 16fps and copy to a subfolder called auto_dataset, it creates the dataset.toml with acceptable values.
https://github.com/alienzed/training/blob/main/autoset.py
@az420 you are a godsend. thank you sir. have a nice day.
There is a lora called "adapter_model_epoch50.safetensors" in your Workflow, where it come from? I can't find that lora.