Trigger words:
Spank (or any of its variations like Spanks, Spanking, Spanked, etc)Example prompt:
You can download any of the showcase videos and drag them into ComfyUI. They all have the metadata attached.
A gorgeous 25-year-old redhead woman with her hair in a ponytail and with her round butt towards the camera. She is wearing a a red bra and red lace thong. She is bent over a bed. She is being spanked. Her butt jiggles with each spank.
Realistic. Rear-view shot.TIPS:
This is being re-trained. it has only been trained for rear-view shots as of right now. It is still being trained, but is already more consistent than the previous version.
I will introduce other angles (Side-view, Front-view) in the future as training progresses. I'm hoping breaking it into separate training sessions helps.
Description
#Attention!
If you create or have a video that you think would be stellar to include in training, please let me know via DM or a link under the comments here. I'd love to continue retraining/refactoring this model using better and better videos. Anime videos don't show much jiggle (and therefor look weird), so good videos of those would be very helpful.
Trigger words:
Spanks, Spanked, Spanking, SpankTIPS:
This model works best with "bent over" positions "lying" positions. Overexplaining spanking (you don't need to describe what a spank is. Just use the word "spank") too much can also work against you.
The V1 INITIAL version of the released spanking model. Hands/paddles/etc are more clear, contact is much more consistent, and impact is clear, albeit not as strong as I'd like. This model will continue to be further tuned, and a new model for v2 will be created from a curated list of outputs. The butt jiggles, but does get reduced when combined with other LoRAs.
This version plays better with other LoRAs than the preview version, but it will need likely need more finetuning to accommodate other positions and styles that other LoRAs include.
CHANGES:
If you want to see the difference between versions, I use the same exact prompts in each showcase videos so you can see the progress.
Hands are more consistent
Butts don't sway involuntarily
Clothing can be changed more easily (Still heavily prefers exposed butt cheeks)
More positions
Larger training dataset (25 videos as opposed to 15)
FAQ
Comments (14)
Ah shit, was going to start one of these for training too. Looking forward to finished lora.
You're more than welcome to! I'm not an expert, so I'm sure having more options out there wouldn't hurt.
Did you use Musubi Tuner?
@midiaplaay No, I set up a custom docker container for diffusion-pipe
I considered making one of these myself but I suspected others might so I hadn't started yet! Thank you, looking forward to seeing it finished!
you three should share datasets!
I hadn't made one yet! Making datasets is an undertaking and I expected others to tackle this one XD
@blyss Making video datasets is very trial and error so if you want to trade datasets privately, I'm more than happy to.
@comfyOverflow I wouldn't mind sharing but I really don't have anything as far as spanking goes unfortunately. I have the dataset for Amorous Lesbian Kisses, and I have a partially processed dataset for a nipple play model I was working on as far as video datasets! I've only been trying to make models to release for a short time, really HunyuanVideo is what energized me to do so but I've been having issues perfecting the second version of ALK >_<
The model seems to still want to "rub" and "grope" in a lot of cases, despite the training dataset very clearly spanking. I'll continue training further to see if it gets better.
Training on videos is a bit harder, because a single frame with a hand on an ass "looks correct", but the issue is in the movement and not the image.
I intend to attempt to train this behavior out very aggressively.
In my attempts at training motion it definitely helps to train longer frame buckets. For instance when training my kissing model, it didn't really start to understand the motion until I included a 129 frame bucket. I don't know what your hardware is but I saw on your other post you mentioned using a docker with diffusion-pipe. Musubi has better options for making use of "consumer level" VRAM e.g. 16-24GB that can enable longer frame buckets or higher res training. If you've got professional hardware or multiple cards you're working with then diffusion-pipe might be more suitable but I just thought I'd mention it!
@blyss That's a good shout. My workflow and hardware can definitely accommodate larger buckets. My assumption was that since a slap is usually such a short temporal action, it wouldn't need longer buckets. I'll try a long epoch (200+) run on the short buckets, and then a shorter epoch (Maybe 50-60) run on longer buckets and see which one understands the action better.
@comfyOverflow Good luck, my initial theory was the same with kissing but the longer bucket really helped the overall coherence!