This LoRA model is similar to my HyperAss model but focuses more on thick thighs and hips. Great for that pear shaped look. You can always increase the LoRA strength for a bigger effect.
If you don't want crazy large sizes, just keep the LoRA strength at 1. It does in-between sizes just fine.
v2 Changelog 2023/06/13:
Damn, its been almost 5 months o.o
Improved the overall quality. Less overcooked results at high strength.
The shape should be even more pear like now, smaller top and larger bottom in general.
Added ass size tags, similar to my other models. See tag data attached with model in the downloads section for details. A few tags are pretty important.
Uses Kohya's LoCon LoRA but does not require any additional extensions to run.
Doubled the dataset size.
Added tag "from front" to assist with front facing shots, and a few others in the tag docs
Notes:
I used this to train my image tagging classifiers for ass sizes
https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification
Description
Training Details:
270 images
LR 8e-5
batch 2
GA 8
dim 4
base model Av3
Av3 VAE
flip aug
clip skip 2
Only Unet training
bucketing at 768
FAQ
Comments (16)
Where did you download <lora:SuperShorty:0.4>?
@throwawayjm ty mate
this works surprisingly well with photo realistic models, it just has a tendency to change clothing a lot is there a way to make it less clothing dependent
Honestly hadn't even thought to try it on realistic models. Usually using an LoRA on a model that is very different from the one it is trained on will cause some additional parts to change unexpectedly. Also the downside to training with a single keyword like I've done here is you don't have as much control over other tags (like clothing). Maye in the future I could try training with clothing related tags included, but at this point I want to wait and see what the next big SD model will be before re-training any of these LoRAs.
how do i put a picture here so u can see what i did wit ur lora
@jkermdawg I think you can only do that when you post a new comment, not when replying to a comment. The new comment menu has a "link" chain for adding pictures.
@samiamrg7 lame af
Maybe it's just my prompting or my inability to properly set lora strength, but whenever I've tried using this lora I don't get anywhere near the sizes you get, I'll get half as large at best and it severely limits poses because it wants to constantly generate from-behind ass-shots even or turn my poses into human spaghetti.
It doesn't matter what base model I use, how strongly or weakly I prompt the Lora, how I change my prompts, I just don't get anything close to the results above.
I want to love this Lora, and while it can look too flabby and realistic at times, it's on the right track for what I want, but I just can't get it to work even half well enough.
Start with one of the prompts in the provided images, and use landscape orientation. Limiting poses might just be a quirk of the model since it was trained on mostly standing images.
@throwawayjm I was afraid of that being the case, where there wasn't a robust enough pose-set for the images used, what's landscape orientation?
@Brushnot like 768x512 instead of 512x768. The wider the image the wider the hips
@throwawayjm I actually try that often, and I still don't get it to blend well or do poses to my prompt. It's aggressive about posing them standing from behind or bent over, and if I set the lora strength too low it doesn't really make the hips or ass nearly large enough.
@Brushnot Another good pose option that recently became available is the ControlNet extension for a1111. I've seen others use my models in conjunction with that to generate some hard to do poses while keeping the LoRA effect (even expanding on the overall size)
where is the download for fantasy race pack? i cant find it
Actually, I don't see it anymore, so I'm not sure what happened to it.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.