search

idiosyncracies / Sebica

Official code for Sebica
MIT License
8 stars 0 forks source link

Request: Please update pretrain file to match infer.py #2

Open FNsi opened 1 week ago

FNsi commented 1 week ago

Nice work!

FNsi commented 1 week ago

Meanwhile, seems mini perform better in tiny size degrade pic. The pretrain standard model official and others trained, leads to significant artefacts in high jpg degrade photo. Especially the dark part.

idiosyncracies commented 1 week ago

Meanwhile, seems mini perform better in tiny size degrade pic. The pretrain standard model official and others trained, leads to significant artefacts in high jpg degrade photo. Especially the dark part.

Thank you for being interested with my work. To address the issue you raised up:

  1. In the infer.py file, ensure that the line img = np.clip(img, 0, 1) is included within the tensor2img(x) function.
  2. It's recommended to split and train your dataset, which are darker images you mentioned. When I trained using the DIV2K and Flickr2K datasets, there might not have been enough dark images for the model to learn effectively from.
FNsi commented 1 week ago

I can sure I didn't change the img part yet

No much clue about that😅

As an example, image

(Pth from this repo)