VQGAN Image Pairs for Super-Resolution Enhancement

Encompassing approximately 2.4 million image pairs for improving VQGAN prediction quality. Each pair consists of a 512x512 image crop from Open Images and a corresponding 256x256 VQGAN-encoded-and-decoded version.

Use Cases

• Train a super-resolution model to reconstruct 512x512 Open Images crops from their 256x256 VQGAN-decoded counterparts.
• Analyze the visual artifacts introduced by the specific VQGAN implementation on a large scale of 2.4 million image pairs.
• Benchmark image enhancement algorithms using the paired high-resolution and VQGAN-processed low-resolution images.

Strengths

• Approximately 2.4 million image pairs provide a substantial volume for model training.
• Image pairs are consistently structured with 512x512 and 256x256 resolutions.
• Derived from the Open Images dataset, a known and sizable source.

Limitations

• The specific VQGAN encoding/decoding process may introduce a fixed, non-random type of artifact, limiting generalizability.
• Dataset is derived from a single source (Open Images), which may carry its own inherent biases.
• No information is provided on the diversity of image categories or subjects within the 2.4 million pairs.

Provenance

Source

Open Images dataset, processed via a specific VQGAN implementation.

Collection Method

Images from Open Images were cropped to 512x512, then encoded and decoded to 256x256 using a VQGAN model.

Freshness

Last updated in October 2022.