Main investigation function.

Create a pure-color PNG in memory to attach as input.

Non-local means denoising.

Single pass of noise replacement. This mimics the diffusion model's low-denoise regeneration: 1. Add noise to disrup

Randomize least significant bits. LSBs often carry watermark info; randomizing them disrupts the watermark while hav

Blend processed image with original guided by structure. Preserves edges and structure while keeping watermark disruption.

Run complete benchmark suite. Args: image_dir: Directory containing watermarked images sample_size:

Sample Pairs Analysis (SPA) - another stego detection method.

Save extracted codebook.

Save visualizations of discovered patterns.

Save visualization of findings.

Randomize phases at carrier frequencies to break watermark coherence.

Multiply a profile's ``carrier_weights`` by ``delta`` in place. Used by ``calibrate_from_feedback.py``. ``delta`` must broadcast to `

Estimated watermark in spatial domain at a profile's native resolution. The caller can ``cv2.resize`` the result to any target resol

Manual wavelet denoising.

Manual wavelet denoising using pywt.

Initialize the bypass. Args: iterations: Number of bypass passes extractor: Optional RobustSynthIDEx

Initialize the robust extractor. Args: scales: Image scales for multi-scale analysis wavelets: Wavel

Args: signature_dir: Path to directory containing signature .npy files extractor: RobustSynthIDExtractor instance for

Initialize benchmark suite. Args: codebook_path: Path to pre-extracted codebook verbose: Print progr

Set the process-wide dataset root used by :func:`_find_res_dir`.

Final reconstruction step: restore detail while keeping disruption. Light edge-aware smoothing + selective sharpening.

Check for systematic color modifications that might indicate watermarking.

Simple, effective bypass using just JPEG compression. Testing showed JPEG Q50 is the most effective single technique,

Check for common steganography tool signatures.

Compare original vs robust extractor performance.

Compare perceptual hashes to detect modifications.

Non-local means denoising.

Detect a SynthID watermark using a loaded V4 codebook. Unlike :py:meth:`detect_array`, this runs at the image's **native resolution**

Detect repeated patterns that might indicate watermarks.

Reduce and expand bit depth. Quantizes to fewer bits then expands, effectively removing fine-grained watermark patte

Primary resolution (first added). Backward-compat helper.

VAE roundtrip with low-frequency restoration from the original. Splits both the original and the VAE reconstruction into a low-freq band (Gau