Discover and read the best of Twitter Threads about #protectart
Most recents (1)
A.i. watermark removers are popping up like mushrooms. These are trained a.i. models that recognize and remove watermarks, even those never seen before.
After some testing Francesco realized how almost everything seems to depend simply on the contrast between pixels.
After some testing Francesco realized how almost everything seems to depend simply on the contrast between pixels.
His knowledge in rendering generation (and about the quality settings f the final rendering) allowed him to understand that primarily the whole process of these watermark remover, was nothing more than a pixel-per-pixel contrast analysis and patching.
Having understood this, the a.i. process can be easily broken.
It’s just needed to prevent A.i. from finding "valid contrasts” as a clue as where to act.
Therefore: reduce the contrast of shapes, reduce the contrast of colors, and alter the repetitiveness of the pattern used.
It’s just needed to prevent A.i. from finding "valid contrasts” as a clue as where to act.
Therefore: reduce the contrast of shapes, reduce the contrast of colors, and alter the repetitiveness of the pattern used.
