<h3 id="concept-and-description:">Concept &amp; Description:</h3>This LoRA was trained on 103 dataset images featuring a 'POV perspective framing a man from the chest down' to handle more flexible character placement and compositions in POV shots. As a rule, images showing 'only the crotch (<a target="_blank" rel="ugc" href="https://danbooru.donmai.us/wiki_pages/pov%20crotch">pov crotch</a>)' or 'the viewer's nose/chin (<a target="_blank" rel="ugc" href="https://danbooru.donmai.us/wiki_pages/pov_nose">pov nose</a>)' were excluded from the dataset.Dataset Breakdown (Total: 103 images):<ul><li>Sitting on lap: 11 images</li><li>Lying on person: 38 images</li><li>Licking nipple: 56 images (of which over 40 involve sexual acts with two or more women)</li></ul>※ The total sum of the breakdown exceeds 103 because certain images overlap multiple conditions.)*<h3 id="known-issues-and-tips:">Known Issues &amp; Tips:</h3>While the dataset is almost entirely composed of heterosexual imagery, only 62 images were captioned with 'hetero' in v0.0.1. Therefore, omitting the 'hetero' tag in this version may yield more flexible generation results.Additionally, when attempting to successfully render multiple characters, it is necessary to use another LoRA such as <a target="_blank" rel="ugc" href="https://civarchive.com/models/2782139/fixed-point-camera">Fixed-Point Camera</a> and apply cues like 'fpcview, head out of frame' to push the subjects away from the center of the frame.

Concept > POV Chest

The training images and captioning have been adjusted to make the intended cowgirl composition easier to reproduce. This is a somewhat destructive change compared to the previous version.Other changes include various captioning adjustments and changes to the training settings, including increasing the LoRA dimension from 8 to 16 and increasing the training image resolution from 512 to 1024.Surprisingly, v0.0.2 contained only a single training image tagged with <code>1girl, pov sex</code>. Here, <code>pov sex</code> specifically refers to a <code>1girl</code> training image where the penis can be clearly seen inserted into the vagina. Images where it is unclear whether penetration is actually occurring—in other words, compositions that could be considered <code>implied sex</code>—are not included.Other Changes<ul><li>Removed <code>2girls</code> images where one of the characters was not participating in the sexual activity, such as stealth sex compositions.</li><li>Cropped <code>multiple views</code> training images down to only the panel showing the intended composition.</li><li>Removed one incorrectly captioned <code>pov crotch</code> image.</li></ul>Depending on the composition, it may work better without including POV-related tags such as <code>pov chest</code> in the prompt. Since all training images place the chest near the bottom of the frame, the LoRA already tends to produce a similar composition even without these tags. In other words, explicitly specifying them may make the composition more rigid and reduce flexibility.Known Issues<ul><li>During sex involving multiple people, nipple licking, or nipple stimulation, there are cases where an extended arm from the person licking or stimulating the nipple becomes positioned over the vaginal penetration area, causing the arm itself to disappear. This can be somewhat mitigated by removing the <code>pov crotch</code> tag or shifting the camera perspective with <code>from ~</code>.</li><li>Due to overfitting to a head-tilting motion, characters may occasionally split into multiple copies. As a workaround, adding <code>multiple views</code> or <code>afterimage</code> to the negative prompt can remove the duplicated character.</li></ul>dim16/alpha16, AdamW8bit, UNet LR 1e-04, constant 1500 + cosine 900(900/1500), 1024² resolution, 168 images.

Files with SHA256: 1ebdf1ff8bb4b3ecb4d30bfa51a554b7f916ceca827da1ab43eea9f65b699938