This dataset contains ~2.4 million image pairs intended for improvement of image quality in VQGAN predictions. Each pair consists of:
A 512x512 crop of an image taken from Open Images.
A 256x256 image encoded and decoded using VQGAN, corresponding to the same image crop as the original.
This is the VQGAN implementation that was used for encoding and decoding:
https://github.com/patil-suraj/vqgan-jaxThis dataset is created using Open Images… See the full description on the dataset page:
https://huggingface.co/datasets/dalle-mini/vqgan-pairs.
