Note: Flux2 Klein9B has come out and I have a checkpoint <a target="_blank" rel="ugc" href="https://civarchive.com/models/2350642?modelVersionId=2643815">https://civarchive.com/models/2350642</a> and Lora <a target="_blank" rel="ugc" href="https://civarchive.com/models/2343427/klein-foreskin">https://civarchive.com/models/2343427</a> I now recommend them as your new image to image models (You use one or the other). I recommend a workflow similar to this: <a target="_blank" rel="ugc" href="https://civarchive.com/models/1870483/simple-kontext-inpaint-cropping-workflow">https://civarchive.com/models/1870483/simple-kontext-inpaint-cropping-workflow</a>. It helps focus only on the penis which helps keep the resolution higher.. It will stitch the penis back to the original image. This way you aren't reducing the resolution. I haven't finalized the workflow, but I think it works better than the standard workflow.v2.0 (long foreskin)- This creates longer foreskins that completely cover the head. This version keeps the original shape of the glans much more consistently and things are more anatomically correct. in regard to the overhang. It is also able to adapt to many more poses and postures. For these reasons it is superior to v1.0, but it still has some issues with some styles.v 1.0 - I finally have consistent results and changing this from a prototype to a working model. It works best with higher resolutions and close up penises.<s>v .6 - this version is much more consistent, though is still at the very early stages of training. I used photoshop to create a wider dataset. It is still a prototype as I find the best settings and dataset for this type of model. </s><s>(don't use, I tried to delete it but didn't take.)</s><s>v .5 - this is my first semi-working version but is not consistent at all. I recommend using the newest version above instead. I trained v .5 for an additional 10000 steps and didn't get any better - more than half of images do not work for v .5. </s><s>(don't use, I tried to delete it but didn't take.)</s>This lora is meant to add foreskin to a penis that already exists in an image. It is not trained to add penises to an image. (See also the foreskin peeler: <a target="_blank" rel="ugc" href="https://civarchive.com/models/1840323">https://civarchive.com/models/1840323</a>)*please have low expectations for this model - v1.0 is a huge improvement, but it is still very rudimentary. it is maybe just a step above inpainting. If your image is high resolution, I recommend using GIMP or another image software to crop a square cut out of your penis, add guides to remember where it was located, use kontext to edit the image, then put the edited image back. You can blend the layers a bit if you need to before flattening the image. This lets you keep your high resolution image without lowering the quality. This will get you the best looking foreskin. Kontext is cool but not as robust as regular Flux. This model may not make the most realistic images, but may be helpful for creating new images for datasets.<ul><li>It does not always work on every image, for some reason it may not find the penis when the body is sitting at some angles. Even though it is trained primarily on close-ups, it sometimes doesn't work on all close up images. Sometimes it thinks the fingers are penises, especially on low resolution images.</li><li>It is not 100% consistent. Some images work better than others and some don't work well at all. You may have to generate images multiple times and sometimes it will never work.</li><li>It is intended to work on both anime AND photos, but it may not be as high quality as desired. Test it out.</li><li>It was more difficult than expected to create a good dataset. If anybody wants to donate images to help grow the dataset, please DM me</li></ul>

Kontext Foreskin

addsuperskin_v3_64rank_000013000.safetensors

Files with SHA256: ec2b72cbaf09e58c65c18056f55f1e99a919659428a44adb45a96018b781d43a