hamadichihaoui / BIRD

This is the official implementation of "Blind Image Restoration via Fast Diffusion Inversion"
244 stars 19 forks source link

How to improve a low resolution face without the high resolution version? #5

Open loboere opened 5 months ago

loboere commented 5 months ago

In the example you show that you improve the low resolution version based on the high resolution version but that is useless in practice, the idea is to use it without having a high resolution version. Am I missing something?

hamadichihaoui commented 5 months ago

@loboere Our method is fully unsupervised and we don't use the high resolution version. In the example we show we synthetically generate the low resolution from a high resolution image (the high resolution image is never used then) just so that we have a reference and we can calculate PSNR, SSIM.. One can easily modify the code to directly load a low resolution image instead. I will try to add this option.