Watermark Remover AI is an open-source desktop tool with 2,000+ stars on GitHub that removes watermarks from images and videos using two modern AI models — Florence-2 to find the watermark and LaMA to inpaint what was underneath.
Watermark removal is really two problems: locating the mark precisely, and hallucinating the occluded pixels convincingly. This project splits the job between two specialist models. Microsoft’s Florence-2 vision model detects and segments the watermark region, and LaMA — a well-known inpainting network — reconstructs the area so the result looks continuous rather than smeared. The pipeline explicitly targets the visible marks stamped onto AI-generated content from services such as Sora and Runway, alongside ordinary stock-style watermarks, and it handles video as well as still images. Everything runs locally behind a modern PyWebview graphical interface, so there is no upload to a third-party remover site and no per-image fee — useful both for privacy and for batch work like dataset creation, which the project lists among its use cases. It is MIT licensed and actively maintained, with updates in October 2026.
Watermark Remover AI is free and open source under the MIT license. It runs on your own hardware; a GPU makes the models faster but there is no subscription or credit system.
Creators, dataset builders and researchers who need to clean watermarks from images and videos locally — especially AI-generated content — and prefer an open tool they control over uploading files to an online remover.