This LoRA was trained using 112 reference images featuring extreme foreshortening in order to amplify the effect as much as possible. As a result, v0.0.0 produces strong perspective distortion on nearby objects, similar to the appearance of a fisheye lens.
Description
Tweaks.
Changes from v0.0.0:
Added 58 images to the dataset (112 → 170 images), featuring highly dynamic action scenes, including motion blur.
Training settings
Increased dim from 4 to 8.
Changed the training schedule from cosine 1000 to constant 500 + cosine 200.
dim8/alpha8, UNet LR 1e-04, AdamW8bit, warmup 0, constant 500 + cosine 200, 512² resolution, 170 images.
FAQ
Comments (1)
one of the best