metadata

license: cc-by-4.0
tags:
  - super-resolution

BHI Mini SISR Dataset

This is a mini version of my BHI SISR Dataset, it consist of 4.5k 512x512 png training tiles.
The focus of this dataset was to extract the high information density and good quality images from the dataset, together with deduplication.

This is for training sisr upscaling networks, can be used for small, medium, or even big network because of its information density.
Though if you are doing paired training with pre-applied randomized degradation, you could use the BHI Small version for better degradation distribution together with a medium or big network.
For complex pre-applied degradation (many different types and strengths) or diffusion network, probably the standard BHI might give better results.