deepak112 / Keras-SRGAN

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network implemented in Keras

276 stars 146 forks source link

How to train to no-squared images #19

Open JeffSaa opened 5 years ago

JeffSaa commented 5 years ago

Hi... I want to use a single model to take any image with any shape and reduce it x4 (for example) and recreate it to original resolution. Is this posible with this code? How can I train this model?

JeffSaa commented 5 years ago

Also... how can I edit the resolution of output image? Is always 640h 480w

mrciolino commented 5 years ago

i had the same problem with the output image size, if you just want the SR img....

18

as for the image of any shape - this code is for a single resolution SR. I think there are other 'multi-scale' super-resolution algos that chip an image into smaller pieces and then reassemble them at the larger res

JeffSaa commented 5 years ago

I already write a version of code where output only the sr image but the problem is that always this output image is 640h 480w....

There anyone know a open source multi scale sr gan project?