Greetings and big thanks for your contribution to the AI-gen detection research.
The idea of using semantic information in addition to the visual noise for classification is very cool.
Could you please clarify something to me?
Let's consider the case of training the whole AIDE network from scratch.
After reading the paper, I was under impression that according to architecture diagram in Figure 2., the convnext Projection Layer is one of the layers being trained. Which made perfect sense, however when I looked into the code here, that does not seem to be the case.
As it sits indented within the torch.no_grad(): block, the weights will stay unchanged in their random initialization state forever.
Please do correct me if I am wrong, but It looks like that might not be behaving as intended.
Greetings and big thanks for your contribution to the AI-gen detection research.
The idea of using semantic information in addition to the visual noise for classification is very cool.
Could you please clarify something to me?
Let's consider the case of training the whole AIDE network from scratch.
After reading the paper, I was under impression that according to architecture diagram in Figure 2., the convnext Projection Layer is one of the layers being trained. Which made perfect sense, however when I looked into the code here, that does not seem to be the case.
As it sits indented within the
torch.no_grad():block, the weights will stay unchanged in their random initialization state forever.Please do correct me if I am wrong, but It looks like that might not be behaving as intended.