Closed Feynman1999 closed 1 year ago
Thank you for your interest. If $g$ is removed and $g^{-1}$ is replaced with a regular neural network, I think the entire model will still work properly. But the performance may reduces if the compression restoration network is not specificly designed.
In fact, $g^{-1}$ in our work can also be seen as a bi-directionally optimized compression restoration network that we also do its inverse $(g^{-1})^{-1}=g$ to simulator the compression distoration and apply corresponding loss functions. Since bi-directionally optimized INN is better than mono-directionally optimized SR networks, I believe $g$ will work better than usual networks.
Besides, in this work we don't force the distribution of the HF (high-frequency) split output of $f$ (and also the HF split input of $f^{-1}$), this components is generated by $g^{-1}$ with the sampled GMM signal and compressed image. If $g^{-1}$ is replaced with regular network, the HF component may need some further guidance.
I understand better, thank you. Additionally, the paper states that $\lambda5$ * $L{rel}$ can use other compression methods, such as WebP. But during training, differentiable jpeg compression was used, wouldn't this cause conflicts between g and g-1?
Yes it will cause conflict. Using differentiable jpeg for training WebP model is sub-optimal, but it's hard to implement each compression algorithm differentiable. Some reasons for what our model achieve satisfying performance on WebP:
thank you for your reply ! it help me a lot,
Hello, very interesting job. I have a question: for compression simulator g, if we remove it and replace g^{-1} with a regular neural network to de-compress, will the entire model still work properly? (lambda_3 and lambda_5 loss removed correspondingly)