New Feature: RealSR (Real-World Super-Resolution)

[Selur]

If you have the time VSGAN (https://forum.selur.net/Thread-Why-VSGAN...-supported) with some models from https://upscale.wiki/wiki/Model_Database might be worth a try.

Cu Selur

[Dan64]

Hello Selur,

  as suggested I tested VSGAN. The encoding speed is better, now a 30s movie is encoded in about 30m. This imply that a 1h movie is encoded in about 2.5 days.
  But the faster speed is at cost of lower resizing quality.

  For the comparison I performed a 4x resize of the following picture



Just for comparison the image encoded with Lanczos+sharpen is the following



The image resized with VSGAN using the model 4xESRGAN.pth is the following



The image resized with VSGAN using the model 4xPSNR.pth is the following



Finally the image resized with RealSR using the model DF2K_JPEG is the following



In my opinion the image obtained using RealSR is still the best among the tested resizer.

[Selur]

Speed wise you might also want to try RealSR_JPEG.pth (see: https://github.com/cszn/BSRGAN/tree/main/model_zoo).
Problem is to find the right model when using VSGAN.

For amusement I converted the suzie clip using the BSRGAN model a few days ago and it looked nice if you consider it's low start resolution.

(see: here for the video)
Sadly BSGRAN is mainly meant for faces, so it's really bad with everything else.

-> hoping there will be ports of tecogan, iSeeBetter and similar methods which are meant for video and not 'just' image upscaling.

Cu Selur