VQGAN Image Reconstruction Pairs for Artifact Correction

Aggregating 100,000 examples for training models to correct artifacts in VQGAN reconstructions. Each example includes a 512px original image, a 256px downscaled 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 original image for artifact removal.
• Develop a super-resolution model using the 256px and 512px image pairs to enhance VQGAN outputs.
• Analyze the specific artifact patterns in the VQGAN-reconstructed version compared to the original 512px image.
• Benchmark image restoration algorithms on the 100,000 provided reconstruction pairs.

Strengths

• Contains 100,000 examples, providing a substantial base for model training.
• Each example provides three aligned image versions (512px, 256px, VQGAN-reconstructed) for direct comparison and learning.
• Specifically targets a known issue (VQGAN artifacts) with a curated dataset for correction.

Limitations

• The dataset's temporal freshness is limited, with a last update in April 2022.
• The specific VQGAN model used for reconstruction (f16, 1024 version) may not represent newer architectures.
• Potential bias exists towards the types of images and artifacts present in the original 100,000 source images.

Provenance

Source

huggingface, created by johnowhitaker.

Collection Method

Created by encoding 256px images with a specific VQGAN model (f16, 1024 version) to generate reconstructed versions.

Freshness

Last updated 2022-04-23.