VQGAN Pairs

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-jax

License

This dataset is created using Open Images, which has the following license: The annotations are licensed by Google LLC under CC BY 4.0 license. The images are listed as having a CC BY 2.0 license. Note: while we tried to identify images that are licensed under a Creative Commons Attribution license, we make no representations or warranties regarding the license status of each image and you should verify the license for each image yourself.