Fellatio afterimages can broadly be divided into two types.The first is <a target="_blank" rel="ugc" href="https://civarchive.com/models/2912244">a single afterimage of the mouth</a> appearing near the base of the penis or close to the mouth itself.The second is this type: a single, complete afterimage of the head appearing with its details clearly preserved.For v0.0.0, I trained the LoRA on 36 training images featuring a single head afterimage. The result is more manageable and easier to use than the <a target="_blank" rel="ugc" href="https://civarchive.com/models/2723087/?modelVersionId=3102478">previous version</a>.However, when using an FT model, afterimages may still be less likely to appear, or the resulting afterimage may have blurry or poorly defined details. I have confirmed that this can be somewhat improved by increasing the LoRA strength to around 1.5 or by increasing the strength of the trigger words.<h3 id="tips">Tips</h3>For vertical afterimages, using a portrait-oriented resolution tends to make them appear slightly more easily. Likewise, for horizontal afterimages, using a landscape-oriented resolution may help. The difference is subtle, but this seems to improve reproducibility somewhat.

Concept > Afterimage (head)

Tweaks.No changes were made to the training images.The total number of training steps has been increased from 620 (constant 540 + cosine 80) to constant 1200. After testing the previous version, I found that the reproducibility of the afterimage was particularly poor, so I changed the training approach to intentionally train somewhat closer to overfitting.As with the previous version, using a landscape-oriented resolution for horizontal afterimages and a portrait-oriented resolution for vertical afterimages tends to improve reproducibility. dim16/alpha16, AdamW8bit, UNet LR 1e-04 cosine 1,200steps, 36 images(1024² reso).

Files with SHA256: db18c6972d2dbe549446cc0aa43d64fa4c0f15e1fc1b07ab54606bf8ec8dc661