Higher Weight = bigger butt
Lower Weight = smaller butt
This LoRA is trained using multi-lora inversion with SVD to rerank back to 4, so it is very well targeted to just the butt-size concept. Tt can be set to weights beyond -[1,1], but extremely high weights can still cause issues.
Description
FAQ
Comments (19)
*bakes this into the checkpoint permanently at a weight of 15.0*
You may experience eldritch horrors :D
How did you do this? Many people told me that training sliders just doesn't work on Minimax H3. Is there new AI Toolkit code that allowed for this, or did you figure out a useful approach?
No, this uses a custom python script I put together. This does not use AI Toolkit's code, but instead uses what I call inversion training.
@GlowingGuardianGirl I already https://civitai.red/models/2920106/boobslider-breast-size-slider-minimax-h3?dialog=commentThread&commentId=1321397
@GlowingGuardianGirl Depends on the concept. For butt size, I train at rank 8 and derank down to rank 4 while combining.
if multiple characters ? one small ass other with huge ass
@gambikules858 you could do this with anything. Literally a slider for apples to oranges if you wanted. But it may just temporally explode. The author does mention that, having to attack the A and B loras differently sometimes depending on how well they come together
@gambikules858 This is a concept-style LoRA, so it affects the whole scene. It is not possible to only target singular subjects in a scene. You could use a Ref2V in MiniMax to do this, though.
@ComfyTinker Is that custom python script anything you're willing to share? I think what you've done is amazing.
@Jellai The article I wrote a year ago has the important parts:
https://www.reddit.com/r/StableDiffusion/comments/1ku2ijc/discovery_creating_sliding_loras_for_wan_and/
@ComfyTinker Okay, looking at the way this works, it looks like it can't work like AI Toolkit. You actually need two datasets, instead of training with text like Ostris's slider stuff, and you need to train two loras. So this slider lora takes twice as long to train, right?
But by using datasets, you can train a continuum of a concept where both sides may not even exist in the model, correct? Or at least may exist in small/weak amounts.
@Jellai I have used diffusion_pipe, Ai-toolkit, and musubi tuner to generate these loras, but yes you are correct. This LoRA was not trained in 1 iteration, but instead 2 separate trainings (large butt and small butt), then the inverse SVD merge of the 2 results in a sliding LoRA.
The super useful thing is if your datasets have very similar details (always nude, always in front of a black background, always a white woman), then they will dampen or completely cancel those out, allowing the LoRA to go beyond 1.0 strength without causing drastic issues.
Additionally, there's the emergent behaviour of being able to produce asses that are way bigger and way smaller than my training dataset. My training dataset's asses are all actually around the size of weight:1 asses or weight:-1 asses.
@ComfyTinker Great dataset tip. I'll keep that in mind.
Thanks
Brilliant. Thank you so much. You've made H3 much better. Are there any other sliders planned? Perhaps for hips, head size, waist, or overall thinness?
Awesome work. Will this modify a reference body from i2v or ref2v?
With following minimax's super detailed prompting instructions for retention analysis, you could get ref2V to discard the body and regenerate, but i2v is a different story.
Good lora, i appreciate your effort for the community.