Thanks for sharing your insights regarding the issue with desaturated and faded images in GANs. I’m encountering a similar problem but with an autoencoder model I’m training using TensorFlow 1.15.8 (DirectML). Problem Description:

My model outputs blurry and low-contrast images compared to the expected results. Here’s what I’m working with:

Python Version: 3.7
TensorFlow Version: TensorFlow 1.15.8 (DirectML)
GPU: AMD Radeon RX 6700XT
Model Type: Convolutional Autoencoder for image reconstruction.

Despite data normalization and implementing data augmentation (rotation, brightness adjustment, horizontal flipping), the model struggles to generate high-quality reconstructions. I suspect it might be related to the convolutional layers or loss function settings. What I’ve Tried:

Reducing the learning rate.
Normalizing the dataset ([0,1] range).
Adjusting the number of filters in the encoder and decoder.
Using MSE as the loss function.

Images:

I’ve included comparisons of the input, expected output (target), and the model’s predictions below:

Example 1:

Example 2:

Questions:

1-) Could doubling the filters in the encoder/decoder layers help address the blurriness as it did for the critic in your GAN?

2-) Is there a way to combine MAE loss with MSE during training to prevent this desaturation?

3-) Are there any specific adjustments I can make to the learning process or network architecture to avoid the blurry and faded outputs?

I appreciate your advice and any suggestions you can provide to tackle this issue.

Thanks in advance!