Caption & Token Quality Comparison

The Lora compares three onsite trainings all using the same dataset and training settings with captions described in the study below.

• Default Pony Settings used other then batch size 4

• 10 repeats per image 10 epochs

• Dataset

• Caption Data

• T5 Online Caption Summary Tool

Study:

Training

• 50 Royalty Free images have been captioned onsite, and trained. LORA Big ASP.

• The Same images and captions trained with the same settings but using the T5 summary tool. LORA T5

• Again the same images but with captions made by hand. LORA Hand-Captioned

• All Training done on base PONY using Captions setting not TAG

Image generation

• Images are seed to seed comparison with CyberRealistic Base model

Royalty Free images from 2008