This is a model which produces that slender "v" shaped region between the lower abdomen and the groin. Many models produce a "masculine" style "adonis belt," however, this LORA is trained off ~ 55+ images of females that exhibit that feminine/strong core musculature, without looking too 'manly.' If you are into that sort of thing, this is the SDXL lora for you...
The model is intentionally slightly overbaked, this means you can get visible results using a weight as low as 0.4 (depending on prompt complexity and model), but 0.8 seems to be a good sweet spot. If your images are looking fried, try to reduce the CFG until you like the result. This LORA also plays pretty nice with other LORAs as well, since the weight can be so low and still have good activation, you don't really need to worry about mixing in other more dominant feature rich LORAs (e.g. character LORA).
Each example image is being generated from a popular model, and they were uploaded as 1 try generations (they are not cherry picked), so as you can see, if you REALLY have a look you're going for, with a little persistence, you should be able to get exactly what you want -- no problemo!
PS. If you enjoy the LORA, please upload your favorite image(s) that you have generated, regardless if they are done with the civitai generator or otherwise. I'd like to train an even better model in the future, based on what people are liking. With that said, the more images I can get that the community enjoys for reference, the better the next training should turn out! =D
Thanks for the support, kind words, and helpful feedback for improving this project going forward.
Description
Initial release -- there are various other training runs with different results. I'm including the training set for those interested in tweaking the model. I only ask that you reference this model as a source/inspiration and provide proper accreditation for the images/training.
FAQ
Comments (7)
Hi! Just find this lora and it is AMAZING!
Can you consider also a porting to pony please?
Yeah, I think I can do that. Are you thinking the default Pony V6 model, or should I use one of the newer realistic merges?
@SlothSimulation well this is a good question... For example I think it would be great if the resulting lora can be used also for non realistic purposes, so the pure pony model should be preferred.
But I'm a newbie in lora training so IDK if the result is going to be screwed up since the dataset is made up only with real photo... I'm this case maybe a realistic merge would perform better...
Mmmm well at the end I don't want to trash away your buzz, time or other resources, so I will be certainly happy with your final choice! đ
@neisanawsaib Thank you for the information about the intended use. The output content is the main concern. In this case, it sounds like the original Pony model is the way to go. I can simply train with a low alpha and moderate dimensionality combined with a slower learning rate, and I appreciate your concern, but I don't care for buzz or anything else. I'm a software engineer and I train on my own hardware. I should be able to train in the next few days. :)
Â
Also, in case you were curious to start learning about LoRA training and whatnot, you can think of dimensionality as "how much we know about something" (i.e. unique traits and identifiers) and the "alpha" as "how quickly we converge on these traits."
Â
For example, if you train for a character, person, or hyper specific style, you'd use a small to moderate dimensionality (e.g., facial features, body shape, clothing styles). These features are highly correlated and specific, and you'd use a higher alpha value which is roughly 1/2 to 1/1 of your dimensionality. This is akin to saying "this character will always look like X and it will always be approximately the same," therefore, we can converge on its traits (appearance/identifiers), quickly with the high alpha; these will train quicker and become easily 'over-baked' because we can converge at a rapid pace; this is ideal for a specific person or character.
Â
However, in our case (using photos of real people for a model which is by default non-realistic), we will use a lower alpha rate (no more than 1/4 of the dimensionality, but most likely an alpha of 1). This will train slower, but converge on traits more slowly during inference (This is akin to saying, "look at these traits and imagine how it might be applied to various scenarios.") It can encompass a wide range of variations and subtleties across different images, and therefore, we can apply these traits to a vast range of scenarios, and it will interpolate something reasonable.
Â
TLDR; Think of dimensionality as the amount of data we are substituting in the model, and the alpha as how quickly we try to 'push' the model to converge on those various traits/identifiers.
- Your friendly neighborhood engineer.
@SlothSimulation OMG man this information is pure gold! THANKS
I'm very interested in lora training and I recently published my first one. I've done some train on 1.5 in the past using my HW (never published as they was not so good) but for this one I've used the on site trainer just to give it a try and speed up the process a little.
With these informations maybe I can finally try a facial expression lora that I'm dreaming for ages...
I really need to study more but I've little time so it is not easy, but quick and well explained concept like the one you provided is always very useful! Thanks again!
@neisanawsaib Oh, so, I still plan on doing this for pony so you can have more control, but I just remembered the reason I did this for SDXL to begin with: I'm pretty sure think that the tag, "groin" or "adonis_groin" will trigger this effect easily on most pony models. However, I was having trouble getting it to look right on standard SDXL models. In the meantime, give that a whirl and let me know if it works!
Here is an example I found on a tag site that was tagged "groin" (I tried to find an image that was close to your area of interest) :)
https://img3.gelbooru.com//samples/25/1d/sample_251d71e864d7d6bed2b7e199596b9ff5.jpg
Could I add to the pony request please? This is Exactly what I'm looking for and I can't seem to get things to look right with groin or adonis_groin. This lora is perfect. Thank you for making it.