---
license: cc-by-4.0
tags:
- super-resolution
---
# BHI Small SISR Dataset

This is a small version of my [BHI SISR Dataset](https://huggingface.co/datasets/Phips/BHI), it consist of 12k 512x512 png training tiles.    
The same way as [BHI_mini](https://huggingface.co/datasets/Phips/BHI_mini), the focus of this dataset was to extract the high information density and good quality images from the dataset, together with deduplication.  
This dataset contains more training tiles than BHI_mini, which is beneficial for medium and big networks, especially when pre-applying degradations for paired training for better degradation distribution.     
For complex pre-applied degradation (many different types and strengths) or diffusion network, probably the standard BHI might give better results.  

 <figure>
  <img src="https://cdn-uploads.huggingface.co/production/uploads/634e9aa407e669188d3912f9/4971HNTaLFqw_Zbb583hR.webp" alt="BHI Small SISR Dataset Example Tiles">
  <figcaption>BHI Small SISR Dataset Example Tiles"</figcaption>
</figure> 

I provide two LR sets with this, x2 and x4, which have been downsampled with pillow bicubic. In the validation folder I also provide an Urban100 HR, x2 and x4 set which have been downsampled the same way.