Odd results in demo.py with your data

[yaseryacoob]

I am running simple tests with your data, something like

python demo.py -f test/celebahq/real/000009.jpg  -m checkpoints/celebahq_sdv2.pth
Testing on image 'test/celebahq/real/000009.jpg'
**************************************************
Prob of being synthetic: 1.0000

Every single run is returning 'Prob of being synthetic: 1.0000'

Can you explain?

[ZhendongWang6]

You may use the wrong testing data. Our model input is DIRE but not image.

[yaseryacoob]

Thanks, I was indeed using the real images and not DIRE images. How do I inference on real images? I would have expected the reconstruction of the real image and computation of the DIRE image to be part of general inference? Thanks for the clarifications.

[Konerusudhir]

The code is hard to understand that its a 2 step process. Testing the code on a set of images is a 2 step process. I mentioned the steps in this issue.

https://github.com/ZhendongWang6/DIRE/issues/7

[yaseryacoob]

Thanks for the guidance, I looked at #7 and the repo again. In https://github.com/ZhendongWang6/DIRE/tree/main/guided-diffusion and https://github.com/ZhendongWang6/DIRE/blob/main/guided-diffusion/model-card.md there is no face model, did you use imagenet model, if so which one? Otherwise is there a face specific model you have? Can you provide feedback on the first step computational requirements for a single image? (is a single GPU enough and how long is the computation?).

thanks for sharing your work, it is much appreciated.

[SilverRAN]

You may use the wrong testing data. Our model input is DIRE but not image.

I used outputs in dire_test for tesing but the predictions are still all of 1.0000. The testing data is DiffusionForensics/images/test/imagenet/real.tar.gz.

If I use the dire results provided in DiffusionForensics/dire/test/imagenet/imagenet/real.tar.gz directly, the predictions are correct.

Seems like compute_dire.py did not give the correct dire results.