When AI watermarks meet AI erasers

What it does

WatermarkRemover-AI is a desktop application for scrubbing watermarks from images and videos. It spots marks using Microsoft’s Florence-2 vision model, then fills the holes with LaMA (Large Mask Inpainting). The whole thing is wrapped in a PyWebview GUI with Alpine.js, plus a remwm.py CLI for scripting.

The interesting bit

The project is explicitly aimed at AI-generated content — Sora, Runway, and friends — where watermarks are increasingly baked into synthetic media. The two-pass video mode with frame skipping and fade-in/fade-out mask extension suggests someone actually thought about how video watermarks behave over time, not just slapping a static detector on each frame.

Key highlights

• Florence-2 for detection, LaMA for inpainting — two distinct models, not a single end-to-end network
• Video support with audio preservation via FFmpeg, batch folder processing, and preview-before-you-wreck mode
• CLI with tunable detection prompt (default: “watermark”), max bounding-box percentage, and forced output formats
• GPU acceleration via CUDA, multi-language UI including “Brainrot” as a listed option
• Settings auto-save between sessions

Caveats

• The README calls it “cutting-edge” and “seamless” — the usual breathless adjectives, though the actual stack is reasonable off-the-shelf tooling
• FFmpeg is “optional” but required for audio preservation; without it, video output is silent
• No mention of quality metrics, benchmark comparisons, or how LaMA handles complex backgrounds behind watermarks

Verdict

Useful if you’re sitting on folders of watermarked AI video and want a local, scriptable cleanup tool. Skip it if you need provenance-preserving workflows or fine-grained control over inpainting quality — this is glue code with a GUI, competently assembled but not research-grade.