I am admittedly kind of a noob at this, but if you enjoy my humble work, any donations are greatly appreciated: https://ko-fi.com/insult_to_ninjary

Your generosity helps me stay motivated to keep improving and to keep trying new things!

Description:

There are a handful of nice LoRAs for bouncing breasts that cover vertical movement, but nothing (that I've come across) that easily produces side-to-side shaking/swinging/swaying. So I made this LoRA for just that purpose!

From my initial testing so far, it seems to work pretty decently, with a couple of notable caveats. First, though I tried to use my training captioning to prevent it, some random zooming/panning seems to have gotten burned into the LoRA. If you don't like it, I'll go over a somewhat consistent solution in the Notes section below. Second, the amount of breast movement you get can be inconsistent, from a little to a lot, and sometimes the upper body movement doesn't quite align with the breast movement, so the physics seem kind of wonky and off. If this happens, just try a few different seeds and you should get something useable after not too long (in my experience).

CallMeMaybe has made a Wan version here. Please note that I will not be making Wan versions of any of my LoRAs at the the present time. I can barely run Hunyuan Video via FramePack-Studio on my low-VRAM laptop. Running Wan in any sort of practical way is probably not in the cards for me without getting a new computer. However, if anyone wants to try to make their own Wan version of any of my LoRAs, I am more than happy to share my training datasets. Just ask!

How to Use:

Full disclosure: my experience with the LoRA is entirely based in FramePack-Studio. Your experience may vary depending on which platform you're using.

No specific trigger word, but you'll want to use the following phrases: shaking breasts. A woman is shaking her breasts from side to side. I mostly gen i2v, so I haven't really tested this LoRA with t2v at all. I'm assuming it should work ok for t2v as well, but you may need to have a more descriptive prompt. For i2v, having a well-set-up starting image is very helpful. Luckily, that just means having a woman with exposed breasts. I haven't tested the LoRA with clothed breasts at all, so I can't guarantee that will work properly, but I hope it does (or that someone can figure out the prompting to get it to work)!

My general rule of thumb with all Hunyuan Video LoRAs in FramePack is to start at a strength of 0.8 by default and then move up or down as necessary. This one seems to work pretty well at 0.8, but you can go higher or lower to try to increase or decrease the motion if necessary. Just experiment and play around with different strengths to find one that suits your needs!

Below is a sample prompt that I used for the first showcase video. The second showcase video uses almost exactly the same prompt, just with a different character description. I don't know how rigidly you need to follow this prompting structure for the LoRA to work well, so feel free to experiment by changing, adding, and/or removing sentences! Let me know how it goes!

shaking breasts. A woman is shaking her breasts from side to side. She has medium breasts. Her nude breasts are exposed. She has pale skin. She is wearing a blue shirt that is open, a plaid skirt, thighhighs, hair accessories, a pendant choker, and a black belt. Her arms are down, and her hands are in front of her groin area. She has medium brown hair in twintails, and she has blue eyes. The background appears to be an office or a library. cowboy shot. anime style. The woman is a cute anime girl. She is smiling, and she is blushing slightly. static shot, locked-off camera, constant focal length. The woman's facial appearance generally does not change. The art style does not change. Image texture and color remain consistent always.

Please note that both showcase videos are i2v. The same general prompting structure should hopefully also work ok for t2v, you just might need to be a little more descriptive with some things. The quality seems much better with i2v though, so I recommend going that route if you can. If you're wondering why my showcase videos might look a little crisper than yours, it's because I always run a 2x upscale with RealESRGAN_x2plus and 2x frame interpolation with RIFE after generating. I strongly recommend it!

Some Notes:

• As mentioned earlier, you might get an inconsistent amount of breast motion and/or upper body movement that doesn't quite align with the breast motion. The easiest way to get around this is just to try different seeds. You could also try adding more specific prompting about what kind of body movement you want, but I haven't experimented with that too much yet. If I discover any good prompt additions, I'll edit this description to add them in.

• Also as mentioned earlier, some camera movements seem to have inadvertently burned into the LoRA. If you're getting unwanted zooms, pans, depth shifts, etc., you can try adding phrases like the following to your prompt: static shot, locked-off camera, constant focal length, etc. In my testing so far, those three specific phrases seem to generally do a pretty good job at preventing major camera movement most of the time.

Outro:

If you make anything awesome, please consider posting them to the gallery! That goes for all of my models! I really like seeing what kind of stuff you all are able to make!

If you have any questions, feel free to ask! I will try my best to help! Also, if you have any ideas for future models, please don't hesitate to share them with me!