reverse-SynthID

Properties

category: watermark-remover
maturity: growing
approach: Reverse-engineers Google DeepMind's SynthID image watermark using signal processing rather than machine learning: profiles the watermark's fixed carrier frequencies and phase values in the spectral (Fourier) domain across many generated images, then runs a multi-stage pipeline (VAE round-trip, elastic deformation, geometric transforms, resize-squeeze, color/contrast micro-shifts, FFT subtraction against a pre-built codebook, and JPEG recompression) to strip the signal while preserving visual quality.
access: free
owns: A research codebase and GUI for discovering, detecting, and surgically removing SynthID watermarks from Gemini-family generated images (including gemini-3.1-flash-image-preview and nano-banana-pro-preview), published for security research.
distinctFrom: Unlike the already-tracked noai-watermark and remove-ai-watermarks (which lean on diffusion-model regeneration passes to disrupt pixel watermarks), reverse-SynthID's core technique is mathematical/spectral: it extracts and subtracts the watermark's known carrier-frequency signature via a codebook rather than regenerating the image through a diffusion model.
source: reverse-SynthID
notes: 4.7k stars. Custom 'reverse-SynthID Research License v1.0' (quoted directly from the repo's LICENSE file) permits use 'solely for Non-Commercial Use' and states 'Commercial Use of this Software is strictly prohibited without a separate written Commercial License Agreement' requiring 'no less than 7.5% of gross revenue'; marked 'free' here (non-commercial only) rather than 'open-source' since it is not an OSI-style permissive/copyleft license. Covered by byteiota.com (April 18, 2026) reporting 91% removal at 43+ dB PSNR in the author's own tests; Google disputed the claims per that coverage, and those figures are not independently reverified here. Origin predates the 2026-07-27 window (first covered April 2026, roughly 1,600 stars then) but was not previously tracked in this catalogue and has grown substantially since, to 4.7k stars; flagged as an untracked gap-fill, not a new launch.

Provenance

Source: https://github.com/aloshdenny/reverse-SynthID

Properties

category

watermark-remover

maturity

growing

approach

Reverse-engineers Google DeepMind's SynthID image watermark using signal processing rather than machine learning: profiles the watermark's fixed carrier frequencies and phase values in the spectral (Fourier) domain across many generated images, then runs a multi-stage pipeline (VAE round-trip, elastic deformation, geometric transforms, resize-squeeze, color/contrast micro-shifts, FFT subtraction against a pre-built codebook, and JPEG recompression) to strip the signal while preserving visual quality.

access

free

owns

A research codebase and GUI for discovering, detecting, and surgically removing SynthID watermarks from Gemini-family generated images (including gemini-3.1-flash-image-preview and nano-banana-pro-preview), published for security research.

distinctFrom

Unlike the already-tracked noai-watermark and remove-ai-watermarks (which lean on diffusion-model regeneration passes to disrupt pixel watermarks), reverse-SynthID's core technique is mathematical/spectral: it extracts and subtracts the watermark's known carrier-frequency signature via a codebook rather than regenerating the image through a diffusion model.

source

reverse-SynthID

notes

4.7k stars. Custom 'reverse-SynthID Research License v1.0' (quoted directly from the repo's LICENSE file) permits use 'solely for Non-Commercial Use' and states 'Commercial Use of this Software is strictly prohibited without a separate written Commercial License Agreement' requiring 'no less than 7.5% of gross revenue'; marked 'free' here (non-commercial only) rather than 'open-source' since it is not an OSI-style permissive/copyleft license. Covered by byteiota.com (April 18, 2026) reporting 91% removal at 43+ dB PSNR in the author's own tests; Google disputed the claims per that coverage, and those figures are not independently reverified here. Origin predates the 2026-07-27 window (first covered April 2026, roughly 1,600 stars then) but was not previously tracked in this catalogue and has grown substantially since, to 4.7k stars; flagged as an untracked gap-fill, not a new launch.