super-resolution

This is the LAION repository for creating open super-resolution models with the help of LAION-5B subsets.

1. Create LAION-5B high-resolution subset
2. Train models

Project outline

• Collecting datasets
• Establish training pipeline and architecture
• Create big model

Datasets

• 170M subset of LAION-5B, of high-resolution (>= 1024x1024) images (metadata 26GB, data 80TB, aws s3 ls s3://s-laion/laion-high-resolution/)
• VQGAN 16k reconstructions

Methods

• Swin transformer theory
• Swin transformer github
• BasicSR github repository
• Image Super-Resolution via Iterative Refinement
• Latent-diffusion