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.

Concept > Disembodied tongue

The issue of the art style being easily altered due to bias in the training dataset has been somewhat improved.However, reducing the bias in the training images appears to have caused another issue to resurface: the tendency for tongues to visually blend or merge together.Changes from v0.1.0<ul><li>Training image selection: The number of training images that were contributing to the unwanted art-style bias was reduced from 32 to 7. The total number of training images increased from 68 to 75.</li><li>Changed the composition of the training dataset:<ul><li>v0.1.0: <code>french kiss 44, cunnilingus 10, foot 12, anilingus 2, total 68</code></li><li>v0.2.1: <code>french kiss 22, cunnilingus 14, foot 16, anilingus 9, armpit 3, misc 11, total 75</code></li></ul></li><li>Training settings:<ul><li>v0.1.0: <code>dim16/alpha8, AdamW8bit, UNet LR 1e-04, constant 612, 68 images (1024² resolution)</code></li><li>v0.2.1: <code>dim16/alpha16, AdamW8bit, UNet LR 1e-04, constant 1,400, 75 images (1024² resolution)</code></li></ul></li></ul>

Files with SHA256: de8e974f99d9f9376518047d3a0ccdda7c39bcf54542eebe4db47d65df4f9de2