⚠Please read the release notes for each version!
⚠Gallery images for v1.0 were generated with various base models (Realvis, Juggernaut, Zavy Chroma, others).
⚠Gallery images for v1.1 were generated with only the base SDXL 1.0 model with 0.9 VAE.
The bizarre gallery photos are intended to showcase model flexibility.
I created this LoRA mostly as an experiment to try out a new way of training introduced to me by micvrotom. It uses color masking and prompting to train. Please check out his write-up located here if you are interested. Please check the files if you are interested in the dataset used to train. On version 1.0, I've noticed is that generating the color associated images look a little better than just images by themselves. It might work better as an inpainting model. Version 1.1 improves on this.
This model can generate both the side-by-side image with the color mask, as well as the individual images, for example:
Color-associated region in two identical side-by-side photographs of a naked woman laying on a bed with cum between her breasts and on her stomach. Trimmed pubic hair. The magenta region is cum.
V1.0 Notes
Initial test of using color guided training. Only trained on side-by-side images. Released the last epoch of training.
V1.1 Notes
Started a new training but added in single images with their own captions to the dataset. Released the last epoch of training.
Description
This is the initial release of Cum XL. I will continue to refine this with more images and improved captioning.
Sample captions used for training:
Color-associated region in two identical side-by-side photographs of a naked woman laying on a bed with cum between her breasts and on her stomach. Trimmed pubic hair. The magenta region is cum.
Color-associated region in two identical side-by-side photographs of a woman sitting on a bed with cum on her chest and the upper part of her breasts. The magenta region is cum.
Color-associated region in two identical side-by-side photographs of an asian woman wearing glasses with cum on her breasts. The magenta region is cum.
Color-associated region in two identical side-by-side photographs of a topless woman with cum on her chin and on her chest. The magenta region is cum.
FAQ
Comments (10)
This looks so promising! Did you find out any drawbacks? You said that the images by itself are worse than the side by side. But it looses quality on the concepts (cum) or the whole image?
It would be nice to compare it with another Lora with the same dataset, but without the side-by-side description and pictures. Also a Lora with mixture of this method and a standard method. A lot of work, but would be interesting.
At this point I can't even tell anymore... some images I generate turn out really good, others not so much. There is a lot more consistency when generating the side-by-side images. I think I am going to add more images to the training data and try to lower the learning rate and increase the epochs. I am also going to improve the captioning.
@MachineMinded looking at your dataset I see they are not square. I know that with buckets it normally doesn't matter, but I wonder if in this case it might be important to not let the trainer cut the image in any way. Did you choose a specific resolution to prevent that?
@diogod Nope, I just used Kohya's bucketing system. I might experiment with that though!
I think I'm going to add some images to the dataset and run another training on this tonight...I like the side by side results but individual results aren't as good.
"messy-pov" + "semen" ... try it together with your model
@stapfschuh Will do!
@MachineMinded better use "cum" instead of "semen". semen looked more like cream and cum looked more realistic
When a caption is given to an image during training, the AI has to try to identify the concepts in the image. If the details are subtle, the AI will have a hard time linking those details to the words from the caption describing them.
Looking at the image I posted, SDXXXL was probably trained on a whole lot more images of cumshot than this lora. As you can see, the cum on the SDXXXL-only generation is much more opaque. That's because without any help, the AI sees most easily the white parts of the cum during training, and that's what it associates cum to as a result.
With the help of the color association, the AI understands much better what is cum in the picture, that it's not only the whiter part that is more visible. So the result is a less opaque cum that looks more realistic.
Folks can view this image here: https://civitai.com/images/6760486