VQGAN Image Reconstruction Pairs for Artifact Correction

A collection of over 100,000 examples of images for training models to correct VQGAN reconstruction artifacts. Each example includes a 512px source image, a 256px version, and a VQGAN-reconstructed version from the 256px image. It was created by johnowhitaker and last updated in April 2022.

Use Cases

• Train a generative model to map from the VQGAN-reconstructed version to the 512px source image to reduce artifacts.
• Develop a perceptual loss function by comparing the 256px image and its VQGAN-reconstructed version to quantify reconstruction error.
• Benchmark artifact correction algorithms using the paired 512px source and VQGAN-reconstructed image examples.

Strengths

• Contains over 100,000 examples, providing substantial data for model training.
• Provides paired data with three specific image versions (512px, 256px, VQGAN-reconstructed) for direct comparison.
• Specifically targets a known issue (VQGAN artifacts), making it a focused resource for a defined research problem.

Limitations

• The specific VQGAN model used (f16, 16384 imagenet version) may limit applicability to artifacts from other model variants.
• The dataset's composition and potential biases (e.g., subject matter, image sources) are not described.
• The dataset has not been updated since April 2022, which may affect relevance to current model architectures.

Provenance

Source

huggingface

Collection Method

Created by encoding 256px images with a specific VQGAN model and decoding them to generate reconstructed versions, paired with original 512px images.

Freshness

Last updated 2022-04-24.