Both hands sensual nipple play (self/other)

Both hands sensual nipple play (self/other)

Update: Finally got a competent Hunyuan version! I couldn't fit it in the short nametag box but it should be face agnostic too! It was trained at 1.6e-5 with LoraPlus of 4 for 2000 steps on just 480x272x105 clips from the same dataset as below except this time normalized from the source to 24 fps instead of 16. Other than that everything was the same. Usage tips for Hunyuan version: All the same triggers and stuff apply, a little bit of real CFG can help any LoRA (not just this one) work(I do guidance scale 1.4 every second step for the first 20 of 50 steps myself, with embedded guidance at 6), I find dropping LoRA strength to 0.9 is best for partnered nipple play otherwise it thinks every hand in the scene has to be on those nipples. Also mentioning where the passive subject's hands are can stop them from playing with their own nipples at the same time as someone else. Now that I've dialed in the settings(I retrained this model over a dozen times before I was happy with it, once per night. My poor 4070 Ti Super never rests XD) I hope to improve my other Hunyuan models for AmorousLesbianKisses and MagicWand!

So this is a particular fixation of mine that nobody had done in quite the way I really like so... I made it! This model makes people play with their nipples or the nipples of others with both hands. It's focused on stimulatory movements like rubbing, stroking, twisting, etc. The kind of stuff that will get you off really well if you have sensitive nipples and like that sort of thing (me! I do! It's my biggest turn on...) Like my usual releases it supports "wide shot/medium shot/close up" for framing the camera distance, in addition to now "medium close up" for in between medium and close up. Recommended strength is 0.8-1.0(I use 0.85-0.92), if you get weird nipples or eyes try turning it down.

Like my MagicWand model, the faces in the dataset were blurred to produce as little face alteration as possible! I actually initially tried without that because I wanted to capture the expressions of pleasure, and while it DID achieve that the faces kind of all look the same so I redid it with blurred faces. The dataset consisted of 16 females, 3 males, and one transfemale as the recipients of the nipple play, about half solo and half partnered, so it should be capable of a wide range of expression. You might add "blurred face" or "face is blurred out" to your negative if you have any bleed through of that!

Triggerable words/phrases: Main trigger: "playing with her/his nipples"; Camera distance: "Wide shot/medium shot/medium close up/close up"; Breast size: "small/medium/large/very large breasts"; Hand motion: "twists her nipples/rubs her nipples/strokes her nipples/stimulates her nipples"; Lighting: "The scene is well lit/the lighting is low/the lighting is warm and sensual"; Clothing "bra pulled down/shirt pulled up/nude" (limited primary genitals in dataset so use a helper if you want good genitals)"

Training captions(so you know how to prompt it! Note these were from the non-blurred version because I don't think you want blurred faces but they are identical except the blurred ones say blurred face instead of mentioning expressions.):

Solo:

"medium close up of a young woman with a sweet face and large breasts playing with her nipples. She has short dark brown hair and is wearing eyeshadow and mascara, as well as a turquoise necklace around her neck. She appears to be about 19 years old. Her face shows an expression of sensual pleasure as she slowly twists her nipples. The lighting is low and sensual and the mood intimate, with the camera focused on her upper body. Behind her a purple paper lantern and a painting can be seen."

"medium close up of a woman in her early 30s playing with her nipples. She has medium breasts and straight brown hair and she's flushed with arousal. She's topless and wearing dainty silver jewelery. She moans with pleasure and looks at the camera as she twists her nipples. A white door is visible behind her."

"medium shot of a man with a shaved chest whimpering while playing with his nipples. He's laying on his back on a blanket and has an erect penis. His white shirt is pulled up to reveal his chest and tummy and his face is out of frame. The camera is focused on the man's torso and the scene is well lit by natural light"

Pair:

"wide shot of an older woman sitting in a modern chair with her legs spread wide. A man standing behind her is reaching over the top of her and playing with her nipples. She has straight black hair and appears to be about 45 years old. She's nude except for a black bra which is pulled down to expose her small breasts. Her face shows a mixture of discomfort and pleasure as he twists her nipples. The man is wearing a tweed shirt and the setting is a well lit livingroom. She has a gemstone butt plug in her anus."

"medium close up of a woman with her dark blue camisole pulled up to expose her small breasts as she lays on a towel on the grass. An off screen man is playing with her erect nipples roughly. She has dark brown hair and appears to be about 40 years old. Her eyes are closed and her mouth is open as she squirms in pleasure. The scene is well lit by natural lighting."

"wide shot of a nude man with a shaved chest reclining in a woman's lap. She's reaching under his arms to play with his nipples. The woman is wearing fishnet lingerie and looks at the camera with a knowing smile as she slowly teases his nipples. Her fingernails are painted white and they are sitting in a leather chair. His face is slightly out of frame"

Training methodology:

20 videos of the subject were collected from various freely available sources online(16 female subjects, 1 transfemale, 3 male), about half solo and half paired, minimum res 720p(most were 1080p before crop). (Side note: Finding enough high quality videos of this kind that weren't just two or three individuals repeated was non trivial!) These were cropped to remove watermarks and provide aspect ratio variation and preprocessed to 5 seconds long clips at 16 fps showing the best examples. The dataset was then fully manually captioned by me using verbose, descriptive language in my usual style. Body parts are referred to with anatomical terms(e.g. vagina, penis, breasts, not pussy, cock, tits) where applicable. Faces were blurred using one of the tools available in Blissful Tuner. Training settings were exactly like my WanKisses model:

Date: 2025-05-14T07:46:39 Title: WanNipplePlay
Network Dim/Rank: 16.0 Alpha: 16.0
Module: networks.lora_wan : {'loraplus_lr_ratio': '4'}
Learning Rate (LR): 2e-05
Optimizer: came_pytorch.CAME.CAME(weight_decay=0.01,eps=(1e-30, 1e-16),betas=(0.9, 0.999, 0.9999))
Scheduler: constant_with_warmup Warmup steps: 100
Epoch: 30 Batches per epoch: 80 Gradient accumulation steps: 1
Timestep sampling: Shift Discrete Flow Shift 3.0
2400 total steps. Base model was T2V fp16, training was done in fp8_scaled.

For training the inputs were bucketed to 1x 480x272x65f, 2x 640x360x33f, and 1x 848x480x21f. I originally tried at only 480x272x65f but this caused very poor quality nipples and loss of fine nipple motion. I still think there is room for further improvement but it's much better than before. All of my showcase videos as well as training were done with my own advanced, extended Musubi Tuner: https://github.com/Sarania/blissful-tuner which contains quite a few advanced and extended features for improving the quality and speed of your generation and training(Latent previews, prompt weighting, upscaling, VFI, beautiful colorful logs and much much more!) You can see my hardware specs in my profile if you're curious. Enjoy those nipples!

Afterthoughts:

This was actually the /first/ model I tried to train for Hunyuan(at the time Wan wasn't out yet). But it's a fairly difficult one to get right because of the dexterous hand movements etc and I was not successful. I've learned a LOT since then and even developed my own version of Musubi, so I decided to tackle it again. I'm much happier with it this time around. I've found Wan to be much easier to train than Hunyuan, likely due to the latters embedded guidance. I still wanna try to make a Hunyuan version of this. I've also improved the "expressive" version of this model (trained without blurred faces) and will upload that soon!