Unwatermark — AI-Powered Watermark Detection & Removal

Unwatermark solves a problem no single AI model can handle alone: baked-in watermarks where the pixels replace original content — no layer to delete, no metadata to strip. The solution is a layered detection pipeline that starts cheap and escalates: EasyOCR catches text-based watermarks for free, Florence-2 handles visual detection, Grounded SAM produces pixel-perfect binary masks, and Claude Vision serves as a fallback for non-standard cases. Removal is handled by LaMa neural inpainting, which reconstructs what was beneath the watermark rather than cloning or blurring. Multi-pass processing catches residual marks across up to 3 detect-remove cycles. The tool runs as a production web app and CLI, processing 14-slide PPTX exports in under 60 seconds.

Technical Highlights

• Layered detection: EasyOCR → Florence-2 → Grounded SAM → Claude Vision → heuristic — each tier adds cost only when cheaper methods fail
• SAM pixel-perfect masking isolates exactly the watermark pixels, preventing collateral damage to adjacent content
• LaMa neural inpainting reconstructs texture and content rather than cloning or blurring
• Multi-pass pipeline with up to 3 detect-remove cycles catches marks only visible after first pass
• PPTX baseline reuse — first successful detection on any slide is cached for subsequent slides
• Provider-agnostic ML inference via Replicate API with swappable local/Modal backends
• NDJSON streaming progress for real-time UI updates during processing
• Stateless Docker deployment on Hetzner VPS behind Caddy with HTTPS