Venus Body Slider
Designed specifically for PonyXL, Illustrious, and other compatible bases (anime & realistic), this slider is perfect for creating feminine figures with narrow shoulders, wide hips, and thick thighs. It enhances body proportions naturally while leaving you in full control of your character's appearance.
Works seamlessly in 0.10 increments for fine-tuned adjustments.
Tested extensively between -0.5 and 3.5 for consistent, high-quality results.
Using negatives like fat or chubby will make the character's body more fit, while using positives like fit or skinny will refine the figure further.
Naturally amplifies existing tags related to body shape and proportions, elevating your creations.
No activation prompt required!
Pro Tip: Combine with tags that focus on the figure or lower body for the best results.
Enjoy the slider? Leave a like, share your results, and let me know how it worked for you! Feedback and requests are always welcome—I’d love to hear your ideas!
Description
FAQ
Comments (24)
Been waiting for this since you mentioned it on Reddit.
Good work. Highly appreciate it! <3
how the f is this only 5mb???
I literally just posted asking the same question lmao XD
dim (or dimensions) of the lora is very small. it effects size output when training. to counteract this to make the style more prominent, you up the alpha while keeping dim low. this makes a small sized lora that still protrudes a stylistic effect when applied. if size of lora is more important than finer details you can do something like 2/16 for dim/alpha setting of a style lora to make the lora very small.
the metadata of this lora has been stripped but if you want to verify what i'm saying is true (other than a simple chat gpt query) you can evaluate this lora, a popular flux one, that was trained with 100dim/100alpha and thus the lora is 1gb in size: https://civitai.com/models/721039?modelVersionId=806265
@skyger aw hell nah im not downloading a 1gb lora 😭😭ill try the 2/16 tip out tho 👀
@BMTZZ TLDR: the dim is the dimensions of the lora, which affects its size. the alpha is how loud that dim is being applied. think of it in these laymens terms:
style = learning_rate * alpha/dimensions
at a certain point, the a/d ratio no longer gains much relevant information extracted from the training set because the base model is limited by its vae (or more specifically, how many channels the vae supports). sdxl vae supports 4 channels, while flux supports 16. these all work congruently to output your final image.
with sdxl i typically dont go over 32 for either a or d as i dont see much net gains in quality or details, while file size just keeps going up dramatically from those settings. flux (or other newer models ... think chroma or qwen) i dont know because i dont have much experience training those, but most likely they would value in higher rates for both if quality/fidelity is your goal.
@BMTZZ @skyger @mrgebien my sliders work differently, because they operate as latent-space directional deltas, not as standalone network adapters like conventional LoRAs. They don’t store a huge matrix of learned concepts, they store a vector direction inside the latent space that blends from A to B, hence why the effect is only added to the concept I want to manipulate and rest is untouched.
theres your answer then i suppose.
this model or type sounds as though its a vera type which is under the umbrella term of lycoris and explains the file size to the op, but civit doesn't allow model types of 'vera' when uploading, although ... it does allow lycoris. maybe something to swap the type of model to @klaabu if it is not a lora.
@skyger my sliders aren’t using a VERA type architecture internally.they don’t train or store low-rank matrices at all, so they don’t rely on dim/alpha scaling or the a/d ratio.instead, the training isolates a latent-space vector direction between concept A and concept B. The output is literally just that directional delta, so the file stays tiny and only the target concept moves when you apply the slider. Nothing else in the base model gets touched.That’s why the sliders behave cleanly at any weight and don’t accumulate side-effects like a conventional LoRA/Lycoris would when pushed to 1–3 strength. The method is closer to latent-manipulation tooling than to model adapters.
@klaabu i see, very interensting. Any chance i could take a look a json/toml(or whatever u use for training) of yours to see how its done? Ive got a couple of very specific body types id like to train and really dont wanna clog up my own and other peoples pc with 100+mbs. U can strip the secret sauce id just like to see a general rough setup so i can go from there worst case 👀
@BMTZZ of course. Feel free to reach out on discord, I can send the metadata. Do note that I'm on holiday and might not get back to you straight away.
@klaabu thats great! enjoy your holidays
I'd love to know how you can make such slider by just using a 5Mb Lora, It's really impressive.
It alters existing weights inside models, not add new information. Like all sliders.
Sometimes it also means it might not really work, if model just "cant do it". Mostly doesnt happen tho.
Very nice. Well executed! Doesn't mess with anything but the thiccness!
Question? Will you upload sliders to specifically modify just one body part? For example: hips, lips, thighs, waist?
Since you haven't uploaded any solo sliders yet, that would be cool. your loras are amazing.
Hei @Garatumon , I can make whatever you want.
looks like it doesnt do anything even at strength 5. using cyberpony 14.1
@thumperunit441 , are you generating locally? There's clearly a mistake in your workflow
I just stopped by to say that this is ABSOLUTELY FANTASTIC!
@DownwardPressure ❤️