It seemed tiny differences before and after correction using my own retrained model

[hermosayhl]

I trained the model with the original paired images of ExpertC in MIT-Adobe-5K, only underexposed and well-exposed images are provided.
When I run the demo, I got almost the same image as input, what happend ?

Left is the input and the right one is corresponding output using model_512.mat

[mahmoudnafifi]

That is odd! Are you sure that the code uses the proper ground-truth directory? Did you try the provided pre-trained model on the same image?

[hermosayhl]

Thanks for your reply!

Listed below are infered results of this picture using your provided pretrained model and it works well

In the course of training from scratch, only underexposed and well-exposed images are provided, which is not same as yours. In fact, G_loss, one of the training loss, is decreased to about 0.02 for training and validation.

I'll check more and figure out where I make mistakes.

[hermosayhl]

That is odd! Are you sure that the code uses the proper ground-truth directory? Did you try the provided pre-trained model on the same image?

I checked this. I truly use the proper ground-truth directory. Perhaps, I have changed the code some where. I noticed that those images you used for training are all with a max length of 1024px, which differs from my 512px.