This is a model for girls using their own breasts as a pillow. Unfortunately, out of the thousands of images I skimmed, there were not many images I could find for this, so the dataset is pretty small. I don't expect this to be too versatile LoRA, but it seems to generate what you want with decent consistency. Works best with a 3:2 aspect ratio.V1.2 Re-tagged all the images and expanded the dataset. I also switched up some training setting so now more framings and positions can be generated. For example, you can now differentiate between the subject resting their head on their breasts or pushing their face into their breasts. Overall, this LoRA is more versatile and consistent.V1.1 Slightly expanded the dataset and re-tagged all images with the newer tagging models.V1.0: 1.0 brings with it support for more position, consistent generation, 512x512, and some 512x768 support. It also allows this LoRA to mingle with other LoRAs consistently. I'd recommend running this at 1.2 weight and keeping pillow in the negative prompt.V0.3:Added a couple more images to the dataset. More importantly, I adjusted my training method to help refine the concept. The LoRA seems to interact with tags and weight better now and can support gigantic breast sizes with decent consistency. I also discovered that a VAE I was using reduced my image quality, so these images were generated with no VAE. If the images you generate seem to match the concept but generate in a low quality or smeared fashion, try a different or no VAE. I plan for the next update to be the 1.0 release, meaning all the essential things like quality, consistency, tags, etc. work how I want when I want. I may add support for more than the 3:2 aspect ratio after v1.0.V0.2:After digging pretty deep, I managed to find a few more images to add to the dataset. I also experimented with some optimizers for smaller datasets. This seems to have improved the LoRA all around, but it's still not complete enough to be considered in 1.0.To do:<s>Give more variety to 512x768 images</s> <s>Incorporate non-sleeping positions</s> <s>Add interaction with breasts (grabbing, breast suppress, etc.)</s> <s>Expand dataset</s> <s>Make all breast sizes generate consistently</s> <s>Make more consistent</s> <s>Incorporate different body positions</s> <s>Add support for 512x512 and 512x768</s>

Solo Breast Pillow

Version 1.0 brings with it better tagging, a larger dataset, more position, and 512x512 support with some 512x768 capabilities. Also seems stable enough to combine with other LoRAs as well.Usage: I'd recommend running this at 1.2 weight to start. If you find the prompt replaces the breasts with pillows or the subjects head isn't fully in contact with their breasts, increase the weight. I'd also keep "pillow" in the negative prompt at all times. When using another LoRA in tandem with this, set the weight higher or the other LoRA lower (assuming this prompt is the more important of the two).

Files with SHA256: 9168357dd3dc50b8115672b8cbde6dc9c99d4659b4cb47efa56b0db3aa017f0d