search

grip-unina / DMimageDetection

On the detection of synthetic images generated by diffusion models
Apache License 2.0
186 stars 18 forks source link

fingerprints #8

Open legion-s opened 10 months ago

legion-s commented 10 months ago

Can you share the training code for fingerprints

RCorvi commented 10 months ago

Hi thank you for your interest. You can find the code to train the denoiser at the following link. In the paper we use a sigma equal to 1.

legion-s commented 10 months ago

@RCorvi Thanks for your answer, but I still want to know some training details:

  1. The dataset used for training (dataset description) and the resolution of the training image (original or cropped)
  2. When training, whether the convolutional layer of the neural network uses padding, and if so, what is the padding parameter?
  3. Loss of use.
  4. I see that the network you use when generating noise residuals does not use padding. Is padding not used during network training?
RCorvi commented 4 months ago

Hi sorry for the late reply, here there is the information you have asked for:

  1. We used a large dataset of pristine images collected from two popular photo-sharing websites: Flickr (www.flickr.com) and DPReview (www.dpreview.com). The whole dataset contains 24,757 images acquired from 1375 different camera models. We are not sure whether we can make it public, we are evaluating. The training is done on random patches of size 64*64
  2. Padding is not used.
  3. The loss used is MSE