Trained on about 200 images each consisting of one image before and one after bukkake, next to each others, from pixiv galleries and game cgs. Most of them are very similar other than the cum and maybe the expression, but a few have a different pose. Similarly, the lora usually outputs nearly identical pose and such, but is able to sometimes do something interesting like a comic, timelapse progression, or pose change.
Available for both SD 1.5 and SDXL Pony Diffusion models.
I recommend you leave lora weight at 1, especially for pony version.
This concept is quite unusual, so there are a few things to take note of, and some known issues.
- Image dimension and ratio is very important, and of course horizontal/vertical lora choice for division. Particular types of acts will have a ratio and division that works best. For example, one girl standing will work best with horizontal, but 2 girls cooperative pov will ONLY work with vertical (because if it's horizontal, the 2 girls make the model think that there are already 2 images; the girl on the right will be covered in cum). If the result you're trying for is poor or the division line doesn't appear, try a different version or image ratio. In general, consider what dimension works better for the act or pose normally, and then make sure each half of the image is in that dimension.
- Unwanted (but understandable) changes can occur in the after image; hair changing color (usually white) and clothes coming off or becoming white (legwear especially), are the most common. Specify hair color and attire in the prompt to reduce the frequency of this problem. Happens more on SD 1.5. As for backgrounds, they're rarely identical, but it's much better than I had hopes for.
- Pony version has more trouble outputting a clean division line, but it's much better at everything else.
- I've tried a few dozen loras combined with this one, and most combine well, but a few cause issue with the concept and force a single image output. If you're getting a single image, try removing other loras.
- Generating full body images is hard because outputting 2 images at once further reduces the details for each face and other small details, and Adetailer is difficult to use because it will add cum on both images. Some artist loras can make faces better detailed with full body than others, I suggest trying different ones if the result is poor, or use manual img2img inpainting to improve it.
