About to Burst

This model aims to help generate a variant of bursting ???????: ??????? that are about to completely burst of clothes and have been pressed into an unusual shape because of that. Basically, the before image of exploding clothes.

I intend to improve and update this with more views, clothing, etc. as I learn more about training models.

V2.2
Another small expansion of the dataset and some retraining. The model is a little more responsive to tag weights and can sometimes do a clothed concept of itself (i.e. ??????? bursting out of a bra while wearing a collared shirt over).
My image saving node messed up, so the example images may not be available for a while. The metadata didn't save to the images themselves and I may or may not get around to manually inputting the generation info for each.

V2.1

Slightly expanded the dataset and made an Illustrious model.

V2.0

Retagged the dataset with a newer tagging model and slightly modified how the tags work to help with the way ??????? spill out. Most important is replacing atb with bursting rbeasts since that produced better results in all my trials. In short, cleavage is ??????? spilling from the top, sideboob is from the side, and underboob is from below. In many cases, it's just as important to specify the locations where you don't want spillage as it is to the locations where you do. Tags also seem more stable, so increasing the weight of them even up to 2+ has produced good results.

XL V1.4

Finally got around to making this for SDXL. I'd recommend running the SDXL LoRA at 1 normally. SDXL does a way better job of following prompts, so a higher weight seems to add in a style more than it improves the image. Some tags also introduce a style. I'll be trimming the dataset and re-tagging it shortly to try and eliminated this.

V1.4

Expanded the dataset by 80 or so images. I also re-tagged every image and combed through the tags to make certain everything looked right. Poses, sizes, and prompts are more consistent as a result. Now combinations like cleavage in the prompt and underboob in the negative should help produce results with only cleavage spilling out of clothes. Likewise, including both cleavage and underboob eliminates most sideboob instances. I also nudge the upper limit of the enormous breasts tag and the wide hips and thick thighs tags are slightly larger on average.

I'd recommend running this LoRA at 1.3 normally and around 1.3-1.5 when other LoRAs are involved. This LoRA is heavily impacted by the canvas size and orientation. Wider images allow for larger sizes but less diversity, and a more narrow canvas produces more dynamic poses but may force larger tags to generate smaller ?????? sizes. Don't be afraid to try all sorts of canvas layouts.

The posted images were upscaled with StableDifusion's built in Anime upscaler to x2 and ran through GFPGAN at a visibility of 0.6.

V1.3

Greatly expanded the dataset to help with tag recognition. I also changed some tags around to make the enormous ??????? and gigantic breasts tags more more consistent.

V1.2

The from side tag now consistently generates images and the ??????????? tag will generate most of the time. I recommend running this at 1.3 weight. The biggest change is the addition of the enormous ??????? tag. This was included to allow for larger sizes than gigantic without making the ???????????????? tag inconsistent. This new tag favors wider images, so if it seems to generate a smaller size than anticipated, try a 640x512 or 768x512 ratio. If you can't quite get a pose right with view tags, try some helper tags. For instance, I had to use looking away and backboob to get a good from behind pose.

V1.1

I made the from-side tag more consistent and finally got the from-behind tag to generate something coherent. From-behind is mostly a proof-of-concept right now. Also made the ?????? size distinction more impactful.

V1.0

Further refined the dataset and added a few more images. Also refined the tagging so side view images could be recognized. The image generation for all sizes from the front seems to be consistent now, but images from the side are hit or miss. The from-behind view is still unsupported.

V0.2

Refined the dataset so generation is more consistent and ?????? sized are more defined.

To do:

Fix trigger word
Fix poor hand generation
Fix poor ?????? generation
Fix multiple ?????? generation
Add side view support
Add rear view support
Add top/bottom view support
Be able to specify if ??????? burst from top or bottom
Try and support smaller ?????? sizes
Remove SDXL LoRA adding a style
Support torn clothes better