I am admittedly kind of a noob at this, but if you enjoy my humble work, any donations are greatly appreciated: <a target="_blank" rel="ugc" href="https://ko-fi.com/insult_to_ninjary">https://ko-fi.com/insult_to_ninjary</a>Your generosity helps me stay motivated to keep improving and to keep trying new things!Description:***Scroll down to "Some Notes" section for troubleshooting tips!***A simple movement, but one that I had a lot of trouble getting to work properly without a LoRA. Maybe I prompt like a dummy? I don't know. But I made this LoRA, so here it is!This LoRA was trained on 10 video clips of AI-generated women squeezing their breasts together with their upper arms. As fun as it would have been to scour through hundreds of porn videos for the exact moment a woman squeezes her boobs together with her upper arms for like 2 seconds, ain't no one got time for that. So I made some clips with Grok Imagine - 5 with women upright, 5 with women lying on their backs, and multiple different skin tones, breast sizes, and body types. Seems to have worked well enough, with a few notable quirks that I'll go over in a bit.<a target="_blank" rel="ugc" href="https://civarchive.com/user/wagenta">wagenta </a>has made a Wan version <a target="_blank" rel="ugc" href="https://civarchive.com/models/2558713/arm-breast-squeeze-t2v-i2v-wan22-video-lora">here</a>. 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 <a target="_blank" rel="ugc" href="https://github.com/FP-Studio/framepack-studio">FramePack-Studio</a> 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: upper arm breast squeeze. and A woman squeezes her breasts together with her upper arms. While I mostly gen i2v, this LoRA should work ok for t2v as well, but you may need to have a more descriptive prompt. t2v pretty often likes to start the video out with the breasts already squeezed together for some reason, so you may need to find some different ways to prompt around that to get the full motion. For i2v, having a well-set-up starting image is pretty important. The more the model needs to move the arms to get them in the right position, the lower the success you'll experience. The best is to have the "v arms" position, or at least arms at sides, and nude breasts. I have yet to get a successful video with a fully-clothed woman, so don't go in expecting that, although if you do get it to work, please share how you prompted it!While 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, you can probably get away with using this one at 1 (or 0.9 if you're worried). The motion is pretty small and subtle, so there's a lower chance than usual of getting mutations or other weirdness at a high strength, and you may end up needing that extra strength to get the motion to trigger properly. I have had some success at 0.8, but it's sometimes a little more hit and miss in getting enough motion, and I have had some minor mutations/deformities at 1, but it wasn't too common. Just experiment and play around with different strengths!Below is a sample prompt that is sort of a mixture of what I used for the two showcase videos. 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!<pre><code>upper arm breast squeeze. A woman squeezes her breasts together with her upper arms. The woman is standing in the center of the frame. She has medium breasts. She is partially nude. She has pale skin. The background is an office. cowboy shot. anime style. The woman is an anime girl. She is moving her arms together. She is wearing a skirt. 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.</code></pre>Please note that both showcase videos are i2v. The same general prompting structure should 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:<ul><li>The model can sometimes be a little picky and random. Sometimes you can easily get the motion with a minimal prompt, and sometimes you really need spell everything out super detailed and carefully just to get a shot at the motion.</li><li>If you're not getting great motion, try using at a strength of 1. Also try adding something like the following to your prompt: She is moving her arms together. or Her arms are straight. She is moving her straight arms together. Finally, it can sometimes be helpful to describe the woman's appearance in detail, even for i2v. It can sometimes help give the model some anchor points to figure out what is where, and what should be moving.</li><li>If you're frequently getting the woman squeezing with her hands instead of her arms, try replacing all forms of the word "squeeze" in your prompt with some form of the word move, even in the triggers. Hunyuan seems to strongly associate squeezing with hands, and the LoRA does not always fully override this connection, so changing the wording can help (though it's not perfect). Also include a description of appropriate arm movement to create the squeeze, such as: The woman moves her shoulders and arms forward and inward.</li><li>If you are experiencing breasts magically growing in size, try replacing all instances of the word "breasts" in your prompt with chest, even in the triggers. Hunyuan seems biased towards big boobs, especially when creating cleavage, so not mentioning breasts at all seems to somewhat take its attention away from making big changes in that area (though it's not perfect). I've had mixed success with it, but you can also try additionally adding a sentence like: Her chest size stays constantly and consistently {size} throughout the video.</li><li>Some of the training video clips had a slight and subtle zoom because that's something Grok Imagine loves to do unprompted for some reason. Knowing this, I actually prompted against it when making the videos, but Grok ignored me and still did it anyway in some of them. I didn't think that the zoom would burn into the LoRA at all, let alone strongly, but it seems it's in there, so please be aware!</li><li>If you're getting unwanted zooms, pans, depth shifts, etc., you can try adding phrases like the following to your prompt, but no guarantees: static shot, locked-off camera, constant focal length, fixed camera, fixed perspective, fixed depth, fixed composition, fixed background, etc. You can also try lowering the LoRA strength to 0.8. If you can still get the squeezing motion at 0.8, you'll also be a bit less likely to get the zoom, especially if you also use some of the prompt suggestions above.</li></ul>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!

Upper Arm Breast Squeeze

Maybe the only version, but please let me know if there's interest in a v2!

Files with SHA256: a6a7a4a57deef1577ebbb554932b5d4d8af68c07ddec7e4771cd6116a20eafbe