Would training and testing with DIRE images in the same format(jpg or png) be effective?

[xqy853174787]

Previously, when testing with pretrained model weights, I found that PNG-format DIRE images were all classified as AI-generated images, while JPG-format DIRE images were all classified as real images.

Now, I have generated PNG-format DIRE images from both ADM and real images in the DiffusionForensics/images/train（val）/lsun_bedroom directory provided by the authors for training. Similarly, the test set has also been converted into DIRE images for testing.（It is worth noting that my training dataset contains only 15,000 AI-generated images and 15,000 real images.）

I tested the generator data in the DiffusionForensics/images/test/lsun_bedroom directory and found the results to be quite good. However, I am not sure if this proves that the DIRE method is reliable, because testing with the author's code (where the DIRE image format depends on the input image format) and model weights also produced very good results.

I then tested the ADM and BigGAN test sets from Huawei's GenImage dataset, but the performance was not ideal. (Previously, when testing with the author's code and model weights, the accuracy should have been around 99%.)