Print message if verbose.

Detect SynthID watermark in a numpy array image. Uses empirically-verified carrier frequencies with per-set reference phases. The wa

Save in compact format (~20 MB for typical dual-resolution codebook).

(H, W//2+1, C) → (H, W, C) via conjugate symmetry.

Load codebook (auto-detects format version).

Best-matching profile for ``(H, W)`` and optional ``model``. Returns ``(profile, (model, H, W), exact_match)``. Prefers an exact ``(m

Load pre-extracted codebook. Supports .pkl files and auto-discovers .pkl when given a .npz path (common user mistake).

Extract noise using multiple methods and fuse results. Uses weighted median fusion for robustness against outliers.

Return (magnitude[3], phase_unit[3]) stacks for one image.

Orthogonal pixel-domain stages applied after the FFT subtraction. Each stage is independently gated: if a stage drops PSNR below ``cfg['psnr_

Collect image files, excluding unwatermarked baselines.

Load image, optionally resizing to target_shape=(H, W).

Lightweight block-SSIM on luminance (mirrors v3).

Extract noise using a single denoising method.

Wavelet-based denoising.

Apply JPEG compression/decompression.

Union a sequence of signed-offset lists; deduplicate at native res.

Proportionally project centred signed bin offsets between shapes. Filters out bins that fall outside the destination Nyquist envelope.

Residual-phase bin subtraction — the blog's published recipe. The correct phase of the watermark at bin ``k`` on *this* image is not

Best-matching profile for target (H, W). Returns (profile_dict, (prof_H, prof_W), exact_match: bool). Prefers exact resolution match,

Extract the top-K consensus bins from ``profile`` as signed offsets. Returns offsets in ``(-H/2, H/2)`` × ``(-W/2, W/2)`` centred coordinates,

Project reference-shape bin offsets onto an arbitrary ``(H, W)`` shape. Returns only bins that stay within the image's Nyquist envelope and are

Encode-decode ``image_uint8`` through the SD-VAE; return a uint8 image. Args: image_uint8: HxWx3 RGB uint8. strength: Scales the

Per-colour FFT accumulation at a fixed resolution. Returns ``(avg_magnitude, mean_phase_angle, phase_consistency, n)``.

CombinationWorst-style geometric attack in a single resample. Applies rotation (``deg``), centre zoom-in (fraction of linear dimension croppe

Average ``F(k)`` with ``conj(F(-k))`` so ``ifft2(F).imag ≈ 0``. Needed after modifying a handful of bins by hand — the ifft of an asymmetric

Compute the complex watermark FFT to subtract from ``image_fft``. Same shape as ``image_fft``. Zero where ``consensus_coherence < tau``.

Per-channel Gaussian blur at the given ``sigma`` using cv2.

Read image dimensions without loading full array.

Inject bandpass noise in a single channel.

Anti-symmetric variant (for phase).

Scramble phases in a single channel.

Analyze entropy of each bit plane - watermarks often reduce entropy in certain planes.

Analyze DCT coefficients - JPEG-based watermarks often modify DCT.

Quick LSB analysis.

Quick LSB analysis.

Analyze Least Significant Bit patterns - common watermark hiding technique.

Analyze pixel value distribution for anomalies.

V2 Combined Worst-Case Bypass Pipeline. Stacks transforms from 6 DIFFERENT categories to maximize the attack surface

Check EXIF and other metadata for watermark signatures.

Chi-square test for LSB.

Chi-square test for LSB steganography detection. High chi-square values suggest embedded data.

Compute Peak Signal-to-Noise Ratio.

Compute SSIM using vectorized block-based approach (no loop).

Detect SynthID watermark in an image file.

Detect SynthID watermark in an image using the extracted codebook. Returns: dict with detection results

Estimate watermark FFT for subtraction. If *profile* / *ref_shape* are not supplied, auto-selects the best-matching profile

Extract and save LSB plane visualization.

Apply JPEG compression/decompression cycle.

Load a V4 codebook ``.npz``. Handles both format v1 (float16/uint8, uncompressed) and the current format v2 (quantized int8/uint8, zl

Load image safely.

Load image safely.

Load image safely.

Remove SynthID watermark from image. Args: image: Input image (RGB, uint8) mode: 'light', 'balanced'

Remove watermark from image file and save result.

RS (Regular-Singular) Analysis for LSB steganography detection. Compares regular and singular groups after flipping operations.

Save all profiles to a compressed ``.npz``. Format v2 changes vs v1: * ``np.savez_compressed`` (zlib-9) instead of ``np.savez``.

Wavelet-based denoising using soft thresholding.

Wiener filter for optimal noise estimation.

Return list of ``(color, (H, W))`` buckets that meet ``min_refs``.

Color channel manipulation to destroy per-channel watermark coherence. SynthID embeds differently per channel (G most, then

Small but non-trivial color/contrast nudge in HSV. Gemini's own guidance lists "extreme changes to the color balance / brightness / heavy fil

Mean pairwise ``|cos(phase_a - phase_b)|`` across all colour pairs. If black and white are both present, their pair is weighted double

Spatially-varying sub-pixel warp — the "collage fragmentation" attack. The Gemini app itself confirms (in its own user-facing error response)

Resolve the disk path for a ``(model, color, shape)`` bucket. ``shape`` is ``(H, W)``; the on-disk folder name is ``WxH`` (Gemini's conventio

Get parameters for each mode.

Get parameters for each removal mode.

Load format_version >= 2 (rfft + mixed precision + optional sparse).

Load original single-resolution format.

Load v1 multi-resolution format (full-precision arrays).

Noise injection + denoising: replaces watermark noise with random noise. Uses both Gaussian and Poisson noise (different dis

Overlay-type disruptions: artifact injection and dithering. These add structured patterns that interfere with watermark dete

Reflect-pad an HxWxC image so H, W are multiples of ``mult``. Returns the padded array and the pad amounts ``(top, bottom, left, right)`` so

Parse ``'model_name/HxW'`` → ``(model, H, W)``.

Parse ``'1440x720'`` → ``(H=720, W=1440)``. Dataset directory names use Gemini's ``WIDTHxHEIGHT`` convention (e.g. ``1440x720`` for a 1440-wi

``y/n`` → ``True/False``; empty/unknown → ``None``.

Quality transforms: JPEG/WebP cycling + resize cycling. Forces requantization across different codec bases (DCT vs wavelet)

Combined worst-case removal using bypass_v2 pipeline. This is the v2 approach that stacks transforms from multiple c

Resize signature to match target image dimensions.

Downsample to ``squeeze × squeeze`` area ratio, then upsample back. The Gemini app's own failure-mode list calls out "shrinking to very small

Pick MPS → CUDA → CPU, honoring an explicit ``prefer``.

Spatial transforms — SynthID's weakest category (52% TPR worst-case). Applies random affine + crop-resize + perspective warp