[Proposal] Place invisible watermark on generated images to distinguish from originals

[jotaf98]

Hi, first of all great work!

Just wondering if it wouldn't be possible to place an invisible watermark on images, to ensure that their generation is traceable, or at least identifiable as being generated by StableDiffusion.

This would be mostly to take a step towards more responsible use of the images, addressing potential misuse of fake photos and so on. The cat is out of the bag, but at least you can put a GPS on it.

One example library that could be used for this: https://github.com/ShieldMnt/invisible-watermark

Instead of applying it post-hoc, it would be nice if the model itself was trained with automatically-watermarked images, so that models would potentially only generate images that are identifiable.

[dagelf]

I'm pretty sure the model's creations already has a tell. Why don't we train a net that can show us what it is?

[Quasimondo]

The txt2img script is already using the WatermarkEncoder from the imwatermark libary to do exactly that.