Do any of these watermark removal systems support simple "training" on multiple images with identical watermarks? Having multiple example images with consistent watermarks should make removing watermarks much easier than trying to remove one with no context.
The legitimate uses feel kind of rare. Maybe there's some stock photo abandonware out there (questionable "legitimacy", but it's not so out there)? Maybe someone bought stock photos from a company that went bankrupt and never downloaded the non-watermarked version, and somehow that company's IP isn't accessible now? Feels like a stretch.
Upscaling old purchased images feels like a more common need.
Incorrect. Removing the watermark constitutes a derivative work. To distribute this work you need permission from the copyright owner to be legal. This you will almost certainly not get since the point of watermarks is to keep people from stealing copyrighted material.
You have a picture of yourself or a friend, with a watermark on it. You remove the watermark. Now the picture looks nicer when you look at it. Why's this difficult?
Their first example at https://www.clear.photo/en is absolutely terrible. I assume a showcase would show "good" results, but they display a complete failure.
- Incorrectly identifies areas for inpainting. You can see this with the figure, a lot of detail, not obscured by the watermark, is erased and then redrawn. This leads to a totally distorted look. The belt just disappears into nothing, the cloth just becomes a gradient, where a crisp line used to be.
- Low quality inpainting. Even the inpainting is done terribly. This looks like something done with some very simple diffusion based inpainting. Absolutely not state of the art.
Yeah their approach of using two different models to detect and then inpaint is very subliminal given that many watermarks are semi-transparent. They could have just trained a UNet with adversarial loss + LPIPS to do all the work and it would have worked much better already.
State of the art is obviously a deep neural network trained for image generation/inpainting. Their inpainting mostly looks like a gradient smeared over the image. Current models can even create fine details and their problem, if anything, is being too detailed.
This is technicaly impressive, but I wonder if this could be put to a use which is generally more constructive. Like maybe removing stains from scans or red eye from pictures.
LOL. They switched out the image on their page. FYI before there was an image of a miniature Baker figurine with chocolate poured down, the Baker figure was totally mangled by the removal process.
Now they replaced it with an image where the inpainting just needs to fill in a gradient. Which is of course trivial.
Why do you not make your product better instead? Obviously the first one was what the customer should expect from your product. Also look at the top left tree! The segmentation still fails to correctly identify the watermark.
I haven't found a tool that implements the techniques described in this Google paper from 8 years ago: https://watermark-cvpr17.github.io/