Unwatermark

Overview

Unwatermark is an AI-powered tool that detects and removes baked-in watermarks from images, PDF documents, and PowerPoint presentations. It was built to solve a specific, recurring problem: exported slide decks from tools like NotebookLM embed watermarks directly into slide images, making them impossible to remove through normal editing.

Rather than relying on a single detection method, Unwatermark uses a layered pipeline — EasyOCR for text-based watermarks, Florence-2 and Grounded SAM for visual detection with pixel-perfect masks, and Claude Vision as a fallback for non-standard watermarks. Removal is handled by LaMa neural inpainting, which reconstructs the area beneath the watermark rather than cloning or blurring over it.

The tool runs as a production web application at unwatermark.cushlabs.ai and is also available as a CLI for batch processing.

The Challenge

• Baked-in watermarks — the watermark pixels replace the original content, so there's no layer to delete and no metadata to strip
• Detection diversity — watermarks range from small text labels (NotebookLM, DRAFT) to semi-transparent logos (Shutterstock, Getty) to rotated diagonal overlays, each requiring different detection approaches
• Content preservation — aggressive removal destroys the surrounding content; conservative removal leaves visible remnants
• Scale — manually cleaning 14+ slides in Photoshop is tedious and error-prone, especially when each slide needs the same treatment
• Format diversity — watermarks appear in PNGs, PDFs, and PPTX files, each requiring different extraction and reassembly logic

The Solution

Layered AI detection pipeline: Detection starts with EasyOCR (free, deterministic, local) to catch text-based watermarks. If nothing is found, Florence-2 runs via Replicate for visual and text detection. Results are refined through Grounded SAM to produce pixel-perfect binary masks. Claude Vision serves as a fallback for non-standard watermarks that escape both OCR and visual models. A heuristic fallback covers the remaining edge cases.

Neural inpainting with LaMa: Instead of cloning or blurring, LaMa (Large Mask Inpainting) reconstructs the area beneath the watermark by predicting what the original content looked like. Combined with SAM's pixel-precise masks, this limits removal to actual watermark pixels — surrounding content stays untouched.

Multi-pass removal: Each image goes through up to 3 detect-remove cycles. Pass 1 uses the full detection stack. Subsequent passes use only cheap, deterministic methods (OCR + Florence-2) to catch any residual watermarks that the first pass exposed.

Format-specific handlers: Images are processed directly. PPTX files have their image blobs replaced in-place inside the PowerPoint package. PDFs are rendered to high-resolution images, cleaned, and reassembled. Each handler shares the same multi-pass core pipeline.

Technical Highlights

• Layered detection architecture — OCR → Florence-2 → Grounded SAM → Claude Vision → heuristic, with each tier adding cost only when cheaper methods fail
• SAM pixel-perfect masking — binary masks isolate exactly the watermark pixels, preventing collateral damage to adjacent content
• LaMa neural inpainting — state-of-the-art inpainting model that reconstructs texture and content rather than cloning or blurring
• Multi-pass pipeline — up to 3 detect-remove cycles to catch watermarks that are only visible after the first pass
• PPTX baseline reuse — first successful detection on any slide is cached and applied to subsequent slides with the same watermark position
• Provider-agnostic ML inference — LaMa, Florence-2, and SAM all run via Replicate API in production, but can swap to local or Modal backends via config
• NDJSON streaming progress — real-time progress updates to the web UI as each detection/removal pass completes
• Stateless Docker deployment — runs on a Hetzner VPS behind Caddy with HTTPS, no database required

Results

For the End User:

• Clean PPTX, PDF, and image exports from NotebookLM and other watermarked sources in under a minute
• Professional-quality output — neural inpainting produces results that are visually indistinguishable from unwatermarked originals
• Simple drag-and-drop web interface with real-time progress — zero image editing knowledge required

Technical Demonstration:

• Production AI pipeline orchestrating 4+ ML models (EasyOCR, Florence-2, Grounded SAM, LaMa) with cost-aware fallback routing
• Practical exploration of AI precision limits — pixel-perfect removal is achievable for corner watermarks but breaks down for large overlays
• Clean architecture with pluggable detection tiers, technique strategies, and format-specific handlers
• Full deployment pipeline: Docker → Hetzner VPS → Caddy reverse proxy → HTTPS