Motivation behind creation: To achieve a deep French kiss in a first-person POV perspective, you traditionally have to bring the character's face extremely close to the camera, which unfortunately ends up destroying the overall composition of the scene. This is a classic dilemma. The concept of a "disembodied tongue" is used as a pseudo-solution to this problem. In other words, while it is anatomically impossible, it serves as a highly convincing manga-style expression for "deep kissing during intercourse" without ruining the frame.I discovered this technique through the works of amazing Japanese artists such as Double Shampoo and 2equal8. Additionally, the dataset contains training images of the disembodied tongue being used for cunnilingus, anilingus, and foot licking.Important Notes / Troubleshooting: Since this model was trained on a limited dataset, unintended elements may occasionally appear as artifacts (for example, onomatopoeia/sound effects tend to pop up frequently). If this happens, please filter them out using your negative prompt or NegPip.

Disembodied tongue

v0.1.0 Changelog<ul><li>Boundary Separation &amp; Artifact Fixing: Improved the boundary distinction between the subject's actual tongue and the disembodied tongue. In the initial version, the two tongues would frequently "melt" or fuse together, resulting in bizarre deformities of the subject's mouth. To fix this, I added 512x512 closely cropped, center-focused close-up images to the dataset. This successfully reinforced the core concept and prevented the blending.</li><li>Art Style Contamination Reduction: In the initial version, the training images contained too much background noise and elements outside the core focus, causing the LoRA to lean heavily into the original dataset's specific art style. To counteract this, I generated "corrected/ideal" images by running the initial LoRA through a popular FT (Fine-Tuned) model to clean up the art style, and added 10 of those refined images back into the dataset. While this has slightly mitigated the contamination, some style bleeding may still remain.</li></ul>

Files with SHA256: edc88ca6060a8eaecde362f2eeb53ad279ff1c406227ad3356acb4f5f406ded4