A fast command-line tool that removes visible and invisible watermarks from images and videos generated by Google Gemini, Veo, and NotebookLM. Each download is self-contained and runs on a clean system with nothing to install. macOS builds are Developer ID signed and Apple-notarized, so they pass Gatekeeper with no extra steps.

detect can also locate watermarks (including SynthID) without modifying the file.

Grab a prebuilt binary from the Releases page (no build step). Every package is self-contained (bundles the AI models and any runtime libraries it needs).

• macOS ships native CoreML MI-GAN (Neural Engine, ~28 ms/frame). macOS 14+ required.
• Linux/Windows ship ONNX Runtime MI-GAN (~225 ms/frame, CPU).
• AI denoise (FDnCNN) uses the GPU via Vulkan/MoltenVK when available, else the CPU.
• macOS builds are Developer ID signed and notarized, so Gatekeeper allows them on first launch (a one-time online check). If Gatekeeper still blocks a build (for example on an offline machine), run xattr -dr com.apple.quarantine <extracted-dir>.
• Third-party licenses: LICENSE-THIRD-PARTY.md.

# Gemini image (auto-detects + removes the sparkle logo)
wmr remove image.png -o clean.png

# Gemini / Veo video (auto-detects the watermark position + size)
wmr video video.mp4 -o clean.mp4

# Veo legacy text watermark
wmr video veo.mp4 --legacy -o clean.mp4

# NotebookLM video (auto-detects the logo + wordmark)
wmr video notebooklm.mp4 --notebooklm -o clean.mp4

# SynthID invisible watermark (no codebook needed)
wmr synthid image.png --codebook-free -o clean.png

Batch a folder: wmr remove folder/ -o cleaned/ --recursive.

Which command? Image → remove · Gemini/Veo video → video · NotebookLM video → video --notebooklm · Invisible (SynthID) → synthid · Just locate → detect

Supported inputs: PNG / JPEG / WebP images; MP4 and other FFmpeg-supported video.

Build a codebook from clean + watermarked reference pairs: wmr build-codebook refs/ -o codebook.wcb.

Measuring a NotebookLM --rect: if auto-detection misses, grab a full-resolution frame and measure the mark's x,y,width,height in any image editor:

ffmpeg -ss 30 -i input.mp4 -frames:v 1 frame.png   # then measure the mark in frame.png

Leave ~1px border around the mark, and pick a frame where it's clearly visible (it can be faint or absent on some scenes).

Release binaries ship an FDnCNN denoiser (NCNN/Vulkan, CPU fallback) that cleans residual artifacts after reverse-blending. It's the default cleanup when built.

wmr remove in.png --denoise ai  -o out.png    # AI (default in releases)
wmr remove in.png --denoise soft -o out.png   # Gaussian
wmr remove in.png --denoise off -o out.png    # reverse-blend only, no cleanup
wmr remove in.png --sigma 75 --strength 150 -o out.png   # tune

Source/dev builds default to AI-OFF (a lean fast build), exposing only --inpaint-strength. Use WMR_AI_DENOISE=1 to build with AI (see Build from source).

Visible watermarks (Gemini, Veo) are alpha-blended overlays: watermarked = α·logo + (1−α)·original. Since Gemini's logo and alpha map are known, removal inverts the blend exactly: original = (watermarked − α·logo)/(1−α). An optional denoise pass then cleans compression artifacts. Two logo sizes: 48×48 (images ≤1024px) and 96×96 (larger), bottom-right corner. Video applies the same reverse-blend per frame, with shot-level detection and anchor-based fallback so no frame is skipped; audio is passed through untouched.

NotebookLM marks are semi-transparent and color-adaptive (not a reversible alpha overlay), so they're removed by AI inpainting rather than reverse-blending. MI-GAN (MIT, ICCV 2023) synthesizes the missing region; on Apple Silicon it runs on the Neural Engine (~28 ms/frame), elsewhere on ONNX Runtime CPU, falling back to Navier-Stokes. The mark is auto-detected per video via template matching (polarity-invariant, stable across scene cuts). On Apple Silicon every scene uses MI-GAN by default; elsewhere a complexity gate picks MI-GAN for textured backgrounds and NS for uniform ones. --notebooklm-method overrides.

SynthID invisible watermarks live in the frequency domain. Removal estimates the carrier signal and subtracts it, either from a prebuilt spectral codebook (--codebook) or from the image's own noise residual (--codebook-free). Limitation: effective on uniform/dark images where the carrier dominates; on content-rich images the carrier is <0.1% of spectral energy, so reliable removal isn't currently possible.

Requires CMake 3.21+, a C++20 compiler, Ninja.

vcpkg (all platforms):

cmake -B build -S . -GNinja
cmake --build build

macOS (Homebrew): scripts/build.sh self-heals stale caches and verifies formulas:

brew install cmake ninja opencv fftw ffmpeg fmt spdlog cli11 catch2
scripts/build.sh                 # release + tests

With AI denoise + MI-GAN (matches the release binaries; macOS uses CoreML, no ORT):

brew install vulkan-volk vulkan-loader vulkan-headers molten-vk
WMR_AI_MIGAN=1 WMR_AI_DENOISE=1 scripts/build.sh

Tests: ctest --test-dir build --output-on-failure.

Single-pass C++20 tool. Everything compiles into one wmr binary. Pipeline: detect → remove → inpaint, orchestrated by WatermarkEngine.

src/
├── core/        engine, alpha blend, FFT, inpaint, ai_denoise, migan_inpainter
├── detection/   NCC visible detector, SynthID Bayesian detector
├── synthid/     spectral codebook, codebook/noise-residual subtractors, builder
├── video/       FFmpeg reader/writer, processor, scene + NotebookLM detectors
├── cli/         CLI11 subcommands, batch processor
└── main.cpp
assets/          embedded alpha-map PNGs + AI models (.mlpackage / .onnx, Git LFS)
tests/           Catch2 unit + integration

See CLAUDE.md for the full architecture deep-dive, and CHANGELOG.md for version history.

PRs welcome. Tests use Catch2; integration tests need the project root as CWD (they look for test-images/). Build with tests via scripts/build.sh, then ctest --test-dir build --output-on-failure. See CLAUDE.md for conventions, platform quirks, and the design notes behind each watermark path.

Built on research and code from:

• GeminiWatermarkTool & VeoWatermarkRemover: reverse alpha-blend, NCC detection, inpainting, FDnCNN conversion (Allen Kuo)
• reverse-SynthID: SynthID spectral analysis
• Picsart-AI-Research/MI-GAN: MI-GAN inpainting (MIT, ICCV 2023)
• KAIR/FDnCNN · Tencent/ncnn · zeux/volk · microsoft/onnxruntime: AI inference stack

MIT. Bundled third-party licenses (NCNN, volk, KAIR/FDnCNN, MI-GAN, ONNX Runtime, OpenCV) are in LICENSE-THIRD-PARTY.md.