Concept & Description:
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 (pov crotch)' or 'the viewer's nose/chin (pov nose)' were excluded from the dataset.
Dataset Breakdown (Total: 103 images):
Sitting on lap: 11 images
Lying on person: 38 images
Licking nipple: 56 images (of which over 40 involve sexual acts with two or more women)
※ The total sum of the breakdown exceeds 103 because certain images overlap multiple conditions.)*
Known Issues & Tips:
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 Fixed-Point Camera and apply cues like 'fpcview, head out of frame' to push the subjects away from the center of the frame.
Description
Changes from v0.0.1:
Training dataset adjustments (103 → 143 images)
Removed several images with compositions that deviated significantly from the POV perspective, as well as images where the character's own chest was too small within the frame.
Added more training images covering previously underrepresented situations, including multiple characters and the 69 position.
Training settings
Increased dim from 4 to 8.
Changed the training schedule from cosine 1100 (1100/2000) to constant 900 + cosine 50 (50/100).
dim8/alpha8, AdamW8bit, UNet LR 1e-04, constant 900 + cosine 50(50/100), 512² resolution, 143 images.