Alright, "breaking" punches a little bit above the weight class in this context: It turns out that Pixxi.io puts the watermark on the pictures as they are being fetched from their backend servers. Also, the quality is given by an additional url parameter. I have no clue why one should ever do this, as this increases server load and makes the whole concept of watermarking completly useless, but here we are.
Caution
Do not do stuff that you are not supposed to do :)
Usage:
- Get all the backend urls by opening the networking tab, then export the HAR Footprint of the network activity as a file ending in .har
- Run pixxibuster.py to extract the image urls to a new file