🌀 Heavy Breasts Physics — V1
My first LoRA for Minimax H3, focused entirely on improving natural breast physics and movement.
After training on 100 video clips and 7,000 steps, this LoRA brings a major improvement to breast motion: bounce, sway, inertia, jiggle, damping, gravity/weight and interaction with the shoulders and torso.
⚙️ Settings
Base model: Minimax H3
Trigger:
bbwbounce, shakes her breastsRecommended weight: 0.6 (natural) to 1.1 (maximum effect)
I2V: Tested
T2V: Not tested yet
Works with all breast sizes and shapes, not only very large breasts.
The goal was simple: make breast movement feel heavier, more natural and physically convincing across different types of motion.
This is currently my most advanced breast-physics LoRA for Minimax H3, and one of the strongest options available for this specific purpose on Civitai.
💬 Feedback
I'd love to see your generations. Feel free to post your videos, workflows, comparisons, and experiments in the gallery. Your feedback will directly help improve future releases and continue refining the physics and realism of the model.
⚠️ Usage Notice
By downloading or using this LoRA, you are responsible for the content you generate with it and for complying with applicable laws, platform rules, and terms of service. Do not use it to depict minors or anyone represented as underage.
Description
V1, send your feedbacks !
FAQ
Comments (21)
thank u! physics simulations are always fun to play around with.
hehe you will like it !
Could you include some of the prompts you used?
will do soon !
Why no prompts included for the examples you posted? It would help us to see how the creator structures their prompts so we can get the most value out of the lora and therefor provide better feedback. Without example prompts or even a guide its kind of a guessing game. Sure you showed us the trigger words but I'm guessing its more involved than just pasting those in and hoping for the best. The examples you showed look good but understanding how to get that with your lora would help us understand how to use it better, which would allow us to give better feedback in my opinion :)
Hey ! You are absolutely right, it's just that I'm lazy to add it while publishing a model ahah, plus theses prompts were very bad haha ! But I will add the prompts now. Just so you know, as long as you prompt breasts, natural breasts, shake, bounce, up and down, left to right, anything like that it will produce an effetc (even without theses tags I think it will). The lora will modifiy the movements accordingly to what actions you prompt. Check the videos it should be captionned now !
Also, all the videos I posted were I2V ! Never tried T2V with it yet, tell me if it works :)
I mean you could also use an LLM or use your imagination. If you just copy prompts word for word it mose likely not work for your setup. It's an H3 based model, H3 understands nuance and context very well. You gotta stop thinking in terms of WAN or SDXL. I've had promps that were 3000 tokens and it got everything from top to bottom in the video.
So the potential to do a lot more is there but it does require more effort. I'd suggest download Qwen-VL Prompt for COmfyUI, and using the qwen 3.5 4b heretic v2 model fromthe drop down ((it'll download for you)), and then use one of the presets for H3. (ALso from a drop down). You may need to adjust the token count and repetition penalty to 1.1 and the top k to 0.2 from 0.6.
I got titties to bounce and flop and drop without any examples. I fucking typed "dem titties be bouncin'" and it made them titties bounce.
@DaddyWolfgang Oh, for the dataset I did that of course ! I used qwen heretic, but thanks for the advices i will try to adjust it. I also use Seedmaster's workflow now, it is much better for prompting. As I said, it is still a V1 I will improve it. I'm just lazy when posting lol. Your caption made me laugh lmao, you obtained great physics with ou without my lora?
@DaddyWolfgang Agreed Minimax is a pretty big step up on the nuance front over the likes of Wan and SD but even still, using an optimized prompt structure in a way the lora expects or was best trained on will be better than pasting a couple words into an llm with no further context. And even an llm either locally hosted like the qwen models or online ones like grok for example, can give better results if it has a few good example prompts with the correct structure to reference instead of trying to go off a few words with no further context in my experience.
@Globullux Ah, fair enough. I mainly do pure text generation with the occasional reference image sprinkled in here and there. I'm generating a bunch now and seems to be working out so far. I will go through them once they all finish and post the best ones back here :)
1GB of SSD are you sure
hell yeah lol
Moar bounce moar better.
Yes, do you want quality or do you want trash? Are you seeing the examples? Show me one other lora that does this at a smaller size and I'll show you an SDXL model that doesn't have same face syndrome.
@DaddyWolfgang lmaooo thanks, that's real you won't regret the 1Gb size fore sure
This and a lora for really stringy-bubbly dripping/dangling strands of saliva or cum mucous are what I am waiting for most. Thanks!
Thanks ! I take notes ;)
This is probably the most gravity-correct LORA yet
Thanks !! And it's only V1 aha
Overal works really well. Been testing T2V, and it does seem to have a pretty prominent face bias.