Type
LORA
Publish Time
2024-06-20
Base Model
Pony
Trigger Words
fizel, nlwx
score_9, source_anime
Copy
Version Introduction
Initial release: Version 1.0
Each char was LoRA trained on a dataset of 32 images using the PonyV6 checkpoint, the OneTrainer GUI/scripts, and the Prodigy optimizer. Both chars were trained for the same number of training steps, in batches of 2, for 50&25 epochs respectively, and with 1:1 balanced regularization using the Waifu Research Department's dataset, available here:
Linel's set was repeated once to account for the increased blonde/brunette regularization.
-- Fizel: (32 x 2/2) x 1 x 50 = 1600 steps
-- Linel: (32 x 2/2) x 2 x 25 = 1600 steps
Dimension/Alpha: 32/16
