Version 3 Update
This version was completely created from scratch for Wan 2.2 using dozens of videos for motion and over a hundred images. It's my most ambitious Wan 2.2 LoRA yet and it took several tries to find the right settings for training. It's still not perfect, but it's pretty good, and I think it's better than Version 2 of this LoRA, so I'm releasing it for you all to give it a try.
Important: As always with my models, this LoRA does not represent any real person in any way. If this model produces a video with any resemblance at all to a real person, that is purely an accident of AI random generation.
This model includes:
motion hints for unbuttoning and unzipping tops, see sample videos for examples
motion hints for breast play and nipple play
general purpose motion guidance for Wan. None of the sample videos used Boob Physics or any other motion LoRA, only Busty Women and Lightning LoRAs were used.
topless and clothed including tight and loose clothing on women of varying bust sizes
a wide range of bust sizes from large to absurd
limited motion guidance for titfuck scenes, but not enough for this LoRA to stand on its own in Text-to-Video. For those scenes, you should use another dedicated LoRA in the Low Noise stage, or maybe Busty Women by itself may be enough in Image-to-Video, I haven't tried that yet.
This model does not include:
full nude images, only topless
any specific faces. All dataset images and videos were cropped to remove the faces or had the faces blacked out to ensure there was no face bias. This LoRA should play well with any character LoRA you want to use.
What doesn't work well:
Some body positions and angles don't work well with the LoRA, I'm not certain why. It's tough to train enough knowledge of all the motions and how clothing behaves on different body types while also leaving the model lightweight and flexible enough to handle scenes outside the training. I'm clearly still learning in this respect.
The low noise model does not produce realistic nipples in Image-to-Video. At all. They're really bad, so bad that I will probably need to train a specific I2V version of the low noise model to fix it. For any other purpose, this model works fine in Image-to-Video scenes, just as long as the woman doesn't start off clothed and then take her top off.
Sometimes a topless woman's areola will have an extra bump or two. If this happens, and you really like the scene, try reducing the low noise strength to 0.75 or 0.5 and see if that helps. You will lose a little bit of fine motion detail, but it will help with the nipples. One of the sample videos uses this trick, so I've verified that it works.
The bust sizes are not very consistent from the prompts. In the dataset, I used three size words to describe the women in the images and videos: large, huge, and gigantic. The model seems to have learned the difference for the most part, but sometimes a "large" woman in one video will be slightly bigger than a "gigantic" woman in another video. It seems to come down to the pose, the clothing, and the seed. See the prompt guidance section below for usage advice.
Version 3 Usage, Strength Values
When using this model by itself, use 1.00 for high and low noise. You can get good results with only the high noise LoRA or only the low noise LoRA, and of course you can use both together for maximum effect. If you do not use Busty Women in either high or low noise, then I recommend using Boob Physics in whichever stage does not have Busty Women so you can still get good natural motion in your videos.
When using with a Wan 2.1 character model, use 1.00 for high noise and 0.50 to 1.00 for the low noise, while doing the opposite for the 2.1 character. For example, if Busty Women is 1.00, use 0.50 for the character. If the character is 1.00, use 0.50 for Busty Women. Or put both at 0.75. Try not to go above 1.50 for the two added together or you will get oversaturated results that don't look as good. You will need to decide for yourself what settings you like better depending on whether you want to prioritize the look of the character or the qualities given by the Busty Women LoRA. It's up to your taste.
When using the model with a Wan 2.2 character, then there are even more options and flexibility. Again, try not to overload the strengths in either high or low, keeping both totals below 1.50 if you can for the two models added together. In many cases, you can get away with only Busty Women in the high noise model and then use your character in the Low Noise model.
Of course you can reduce the strength of the Busty Women LoRAs to reduce the overall effect, but keep in mind you will also be limiting the natural motion given by this model. If you reduce the strength below about 0.75, I recommend adding Boob Physics to compensate for the motion loss, unless you're also using some other motion LoRA in your workflow.
Version 3 Usage, Prompt Guidance
I included my full prompt text with each sample video, so you can use those as examples.
The model was trained with three size descriptive words: large, huge, and gigantic. The model will not affect the woman's bust size very much unless you use words to describe how big you want her to be. For example, "a woman with large breasts is..." or "She has huge breasts."
Wan understands "very" and "extremely" so you can use that to your advantage with phrases like "very large" and "extremely huge" and so on.
The dataset included both tight and loose clothing. This model does a good job with both, which was one of my goals. Most of the busty models available for Wan, Pony, or anything, are either topless or all the clothing is skin-tight. I wanted a model that did a good job with normal clothing that isn't super tight. Of course it can render tight shirts too, but you have the option for a loose t-shirt or whatever if you want it.
The training images did not have much descriptive text about nipples, so the size and shape of what you get will be mostly random.
Breast play motions included descriptions of squeezing and lifting, so both should work well.
Nipple play included phrases like "she uses her fingers to pull and rub her nipple", "she twists her nipple with her fingers", "she uses her fingers to rub her nipples on both breasts", "she uses her fingers to pinch her nipple"
If her top has a zipper, you should begin by reinforcing the description of exactly how it looks at the beginning of the scene, like "her shirt has a front zipper" or "her jacket has a zipper down the front". Also describe whether the zipper is fully closed or partly open, "the shirt is zipped closed up to the neckline" for example. Then describe the action, "she grabs the zipper with one hand and pulls down the zipper exposing her white lace bra". Be sure to describe what you will see when the zipper is open: bare breasts, cleavage, a sports bra of whatever color, etc. If you don't mention it, then Wan will invent something for you.
If her top has buttons, then the prompting is similar to the zipper. When she opens her top, the training videos used a phrase like "she uses her hand to pop open the buttons on the front of her blouse one at a time". If you don't use a phrase like that, you may end up with the whole front opening at once which doesn't look as realistic.
I hope all of this text is helpful, and I can't wait to see what everyone else makes with this model, so please remember to tag either the high or low noise model when you post your videos on Civitai.
Wan 2.2 Version 2.0 Notes - This model works well with Wan 2.2, and will work best when used with a non-standard setup that favors the Low Noise stage of the render. If you want to try my workflow, it's embedded in the sample videos above, or you can download the workflow from Civitai.
ComfyUI workflow guidance
For my current favorite ComfyUI configuration, please refer to my article for all the necessary info you'll need to make the most of my Wan 2.1 LoRAs in Wan 2.2 videos.
Based on the Busty Women Flux LoRA, this model will enhance the bustlines of women in Wan video clips. Trained on Wan 2.1 14b text-to-video, it will work best in that setting. In my testing, it appears to play well with other Wan LoRAs like character and motion models. All of my sample clips were rendered with Boob Physics, for example.
Important: This model is 100% fictitious and is not based on images of any real person. The data set for this model originated with images generated from still image checkpoints and has evolved from there using only AI rendering. No images (neither actual nor AI-generated) of any real person was ever involved in the making of this model.
The data set for version 2.0 had all faces clipped out from the images so this LoRA should have very little influence on faces in your videos. The LoRA can be used at any strength from 0.1 to 2.0. At strengths above 2.0, the results were not good so I don't recommend anything higher than 2. The data set models all had thin figures, so you may see some other body shape influence beyond just bust size, especially at high strengths.
No keywords are needed. The effect can be enhanced by including words like "busty", "very busty", "large breasts", "huge breasts", etc.
As always, please post your videos back to this model page or at least tag this model in your video posting so I can see what other people are making with the model. I hope you enjoy.