Skip to content

Lexica2 Dataset

Compare
Choose a tag to compare
@Phhofm Phhofm released this 03 Sep 14:44
· 10 commits to main since this release
775b15d

Lexica2 Dataset

A dataset to train single image super resolution models specifically on/for SD/AI generated images.

When releasing my older datasets I also saw the quality and size of the Lexica dataset and thought I would rework the dataset into a version 2. So this should be the higher quality with less images, consisting now of 6'500 512x512 image tiles - retiled, dejpg'd, deblurred, scored, filtered.

Dataset Name: Lexica2
Num images: 6'500 512x512
Purpose: For training upscalers of SD/AI generated images.

Based on Lexica Aperture V3 Images

Example of HR images from the dataset:
lexica2

hyperiqa score: 0.7842237427784846

The tar files include the hr folder with 6’500 images.