Matthew Tancik

The detector model provided is the one we used in the paper and video. It works best when the StegaStamp has a clear border around it (the edges of the...

The decoder will produce a random output. If the BCH error correcting codes are used, the BCH decoder will report that the code is invalid.

The total message budget is 100 bits but some of the bits need to be used for error correction. We use BCH codes which you can learn more about here...

YUV is a color space (https://en.wikipedia.org/wiki/YUV). The scaling varies the loss for color versus luminosity (the u,v channels are color).

By setting u and v larger you increase the loss on the color channels. Weighing them more helped avoid weird colors from emerging during training.

The released model was trained using `bash scripts/base.sh EXP_NAME`. What differences are you finding in performance?

Can you confirm the following, _ No modification of the code were makes _ The parameters in `scripts/base.sh` are being used _ The MIRFLICKR data is being used for training...

Hmm, interesting. Can you also provide a screenshot of the remaining scalar logs and screenshot of the image logs.

It is odd that the input images are black. It doesn't seem to be using the dataset images.

Make sure that you clone the submodules, `git clone --recurse-submodules https://github.com/tancik/StegaStamp.git` If you have already cloned the repo without the submodules, run `git submodule update --init`