Technically the project evolved too. At first it used crude frame differencing: identify a static rectangle, blend surrounding pixels, and hope. That worked for DVDs and ancient camcorder logos, but failed spectacularly on modern, animated marks. So Mina added intelligent inpainting models—lightweight, privacy-conscious neural networks trained on synthetic watermarks and non-copyrighted footage. The models ran locally, and the CLI offered presets: “restore home video,” “educational reuse,” and “archive cleanup.” A careful mode preserved subtle artifacts when requested, so restorers could keep historical fidelity rather than producing a glossy, untraceable fake.