For more LoRAs and updates, Join my Discord:
https://discord.gg/ZVWVhT43GW
You can deploy MiniMax on RunPod with my template: https://get.runpod.io/minimax-template
Follow me here so you see the next one.
Breasts, areoles and nipples on MiniMax H3, and control over them from the prompt. Base H3 renders a chest as an approximate shape and ignores anything you say about it. This gives you size, areole size and areole color as things you type.
Trigger is HMBreasts. Lead with it.
The four axes
These aren't tags. Write them as plain English inside the sentence about her chest, the same way the model learnt them.
Breast size - tiny · small · medium · large · very large
Shape - natural · perky · round · perfectly shaped · saggy · sagging · hanging · pressed together
Areole size - tiny · small · medium · large · very large Areole colourpale · ghost · brown · dark
Nipples - erect · hard · pierced
Spelling matters here. It learnt areoles, so write it that way.
The dataset has plenty of each, but as you all know, AI is non-deterministic so don't get mad if the results you're expecting doesn't appear right away.
How the clause reads
HMBreasts, large, natural breasts, medium sized brown areoles and erect nipples
HMBreasts, small, perky breasts, tiny sized pale areoles and hard nipples
HMBreasts, very large, round breasts, large sized ghost areoles and erect nipples
Mix the axes freely. They were trained independently, so tiny breasts with large sized dark areoles is a combination it can actually give you.
What it doesn't do
Genitals. That's HMPussy, and the two stack. This one owns the chest and nothing below it.
Faces, bodies or identity. It's a single-concept anatomy adapter, so put it under your character or scene LoRA.
One honest limitation
medium is the weakest of the five sizes. The training material clustered at the two ends, so tiny and small steer hard, large and very large steer hard, and the middle is softer than either. If medium drifts bigger or smaller than you wanted, that's why. A shape word will usually pull it back.
It's trained on stills, so it teaches H3 what a chest looks like. How it behaves through motion still comes from the base model.
Description
FAQ
Comments (34)
Amazing work on all your loras!
is there a way to make lora work nicely with other lora with disfigured body horror?
I've found each of the loras need to have their own turbo-lora value at the least. And you really need to use the language described in both the lora, and the H3 guide. When you get all three right, all of a sudden most of the loras start looking like the example videos.
@makiaeveli >turbo-lora value at the least what does that mean beyond setting strength?
@2legsRises357 that's all it means. Im just saying that some loras perform better with more or less steps. That can roughly be translated into turbo lora strength. Ive had good results from 0.8 to 1.1, but it all depends on the lora it seems.
Great. Can you share the prompts of your examples? Not the workflows, just the prompts.
I put the prompts in the images when I have them.
I generate the prompts using Gemini 3 Flash Preview and the OpenRouter node, so if I forgot to save and click generate again, they're gone.
Works well. in my case values above 0.5 tends to break the overall composition and kill prompt adherence or natural movements, but this is mi isolated case using turbo lora, just a little feedback.
I’m not using Turbo LoRA, but I’m seeing something similar. Values above 0.5 can sometimes add weird lighting effects and make the movement behave strangely. Lowering the LoRA strength seems to give much better results for me.
Great lora, but are my eyes deceiving me or there's no puffy ones on the menu?
There are some, not captioned tho
for some reason this lora seems to glitch out my generations i think with the tubro lora, the HMPussy lora works fine. no idea if it's because I'm attempting to use both this and the pussy lora at same time.
I've had the same issues with the Pussy lora too. Something is wrong with how these are trained. Also heavy quality degradation as well
Use lower strengths.
Mixing a few causes horrors for me as well.
same here, but with 0.6 strength works great. I hope it gets even better in version 2
@HearmemanAI so far I have tested different strengths but it seems to just have issues no matter what sometimes, the pussy lora works fantastic at 1.0 for me for most part
can confirm. I've been using larryvrh tubro + blowjob just fine. After adding this, i'm getting deformed limbs, random ghostly blurry textures or a blue tint all over. I will say that lowering the strength of this lora to something as low as 0.30 got rid of the deformed stuff but the movements are still a bit messed up for some reason. Not saying it's this loras fault. I'm new to this so hoping someone figures it out because this is a great lora!
Nice work!
This feature is great! It's just a shame the texture isn't realistic enough. I previously used a Lora model created by someone else; the texture of her breasts, nipples, and areolas was quite realistic. Also, I'd like to hope they add a longer nipple effect! Thank you!
Thanks! that is the missing pieces! let me try
inverted nipples as well please!
Does this LORA bring realistic bounce and jiggle to I2V?
No, trained on images
no
We need sloptwerk 2.0 up in here asap. Im surprised no one has gotten on it yet. On wan 2.2 it was great to use as an always on lora with no special prompting.
Thank you for putting the trigger in the filename. It's wild to me that I have to keep detailed notes on LORAs instead of there being a better way... unless I'm just being a noob and there's a LORA detail node I'm missing...
I am not sure , triggers etc are needed or work for any of the modern models since last two years
Hello everyone ! Someone has recomendation about the value that works the best for the Lora ?
I'm new to this so don't take me too seriously but from what i've tested, without any lora, using this on 1.0 works just fine. I use larrys turbo lora + this one and it works great. However, adding another lora just makes everything look BAD. Try lowering this lora with each test. For me lowering it as low as 0.30 helped a little but its still not 'right' .
@marmoth5 Thanks for the awnser ! I'm new in video too !
I try at 1.0 strengh with a Turbo Lora and it not follow my prompt (Not follow the audio text and some details). At 0.5 it's seems to upgrade the bodies and the realistic quality but i'm not sure to use the Lora at it's full potential !
@Skeyz are you prompting as per the written guide on hugging face? I found that that really helps
@marmoth5 I can't find the guide on Huggin. Do you have a link ? Thanks a lot :)
@Skeyz https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
they cover a looot of details. since i'm new it was mostly gibberish for me. so what you can do is paste these links into gemini ai or something, ask it to read it thoroughly and ask it to generate prompts for you. obviously it cant generate nsfw prompts so just trick it. for example i told it a woman is licking an icecream lmao.
@marmoth5 =Thanks a lot man !!