Resolution mismatch: build the watermark at the profile's native shape, resize spatially, subtract. Lower quality than the exact path

Per-channel bin-wise complex subtraction at native resolution. For each ``(fy, fx)`` we look up ``phase`` and a scale reference from

Average across models at a fixed resolution (phase-aware).

Add calibrated Gaussian noise to disrupt watermark. The noise level is carefully chosen to be strong enough to disru

Add an image to the analysis.

Synthesise a ``union`` pseudo-model that averages per-resolution profiles across all real models. Useful when encoders are believed id

Analyze bit-level patterns for watermark embedding.

Detect color banding artifacts common in AI-generated images.

Check if there's a consistent color shift that might indicate watermarking.

Analyze color shifts.

Analyze color shifts.

Analyze correlation between images to find shared patterns.

Analyze DCT coefficients for quantization-based watermarking.

Quick frequency domain analysis.

Quick frequency domain analysis.

Analyze frequency domain differences - watermarks often hide in high frequencies.

Analyze frequency domain for carrier signals.

Compare histograms to show systematic modifications.

Perform full analysis on a single image.

Analyze JPEG APP markers for hidden data.

Analyze LSB spatial patterns for watermark location.

Analyze noise patterns across all images.

Analyze a single pair.

Apply multiple noise replacement passes. Similar to multiple KSampler passes in the diffusion workflow.

Remove watermark from all images in a directory.

Benchmark detection accuracy. Returns: Dict with detection metrics

Benchmark removal quality. Args: images: List of (path, image) tuples output_dir: Directory to save

Bilateral filter denoising (edge-preserving).

Walk ``root/{model}/{color}/{HxW}/*.png`` and build profiles. Args: root: Dataset root directory. models: Model subdi

Build a profile from diverse watermarked images at native resolution. Content averages out across images; the fixed watermark signal

Main bypass pipeline - remove SynthID watermark. Args: image: Input image (RGB, uint8 or float) mode

Bypass watermark in image file and save result.

Bypass watermark in image file using v2 pipeline.

V3 Spectral Bypass — multi-resolution codebook subtraction. Automatically selects the best-matching profile from the codebook

Bypass watermark using v3 spectral pipeline and save result.

Dissolve the SynthID watermark in ``image``. Args: image: RGB image, HxWx3, uint8 or float in [0, 255] / [0, 1]. code

Read an image, dissolve, write result. Returns the BypassResult.

Unified all-in-one attack targeting every Gemini-confirmed failure mode. Pipeline (each stage independently PSNR-gated via a soft floor):

Round-05 attack: diffusion-VAE re-generation + CombinationWorst geo. Pipeline (gated by PSNR floor at every step): 1. VAE roundt

Native-resolution absolute-bin subtraction. Targets a fixed set of ``(fy, fx)`` bin offsets (blog universals + codebook-harvested top

Update ``carrier_weights`` in-place. Returns per-profile summary stats. The update rule, per profile ``P``: Let ``F`` = number of feedback r

Check for C2PA (Content Authenticity Initiative) manifest.

Analyze JPEG compression artifacts and quantization tables.

Subsample and upsample chroma channels. This is similar to what JPEG does but more aggressive, disrupting watermark

Apply small random color adjustments. Subtle color variations break watermark coherence.

Analyze spatial differences between original and edited.

Compute perceptual hash difference to detect invisible modifications.

Compute perceptual hash distance.

Compute perceptual hash distance.

Compute Peak Signal-to-Noise Ratio.

Compute simplified SSIM.

Analyze corners for visible watermarks.

Create difference visualization between original and edited.

Create a summary report of all evidence.

Edge-preserving bilateral filter denoising.

Detect carriers using multi-scale analysis with voting. Carriers that appear consistently across scales are more reliable.

Detect carriers at a single scale.

Detect JPEG compression artifacts that might hide watermarks.

Detect neural network-specific artifacts that might indicate AI generation.

Check for visible watermarks in corners (common placement).

Dissolve one image at one strength; return a manifest row.

Download image folders from HF dataset repo.

Extract all 8 bit planes from an image.

Extract all 8 bit planes from grayscale image.

Extract SynthID codebook from a collection of watermarked images.

Extract comprehensive codebook from watermarked images. Uses multi-scale analysis and ensemble methods for robustness.

Main method: Extract the SynthID codebook from analyzed images.

Extract DCT coefficients pattern.

Extract Fourier magnitude and phase.

Build a profile from black / white / random reference directories. The profile is stored at the native resolution of the black image

Extract LSB (Least Significant Bit) pattern from all channels.

Extract noise residual using wavelet denoising.

Extract watermark signature directly from pure black/white Gemini images. On a pure black image, every pixel > 0 IS the wate

Extract structural information (edges and texture). This is similar to how Canny edges are used in ControlNet to pre

Use ICA to separate watermark pattern from image content. The watermark should appear as a consistent component across image

Extract the watermark by computing the difference.

Extract the average watermark signal from all images.

Extract wavelet coefficient patterns.

Analyze bit planes for consistent patterns.

Find carrier frequency peaks using combined magnitude and coherence.

Find LSB bits that are consistent across all images.

Analyze DCT coefficients for embedding patterns.

Find consistent frequency components (carrier frequencies).

Find consistent noise pattern (embedded signal).

Generate one watermarked reference image via Gemini. Args: client: google.genai.Client instance color: "black" or "white"

Apply guided filter for edge-preserving smoothing.

Inject noise in specific frequency bands where watermark lives.

Resolve ``--input`` (file, directory, or glob) to a sorted list.

Apply multiple JPEG compression cycles at varying qualities. This disrupts the watermark through quantization artifacts

Join manifest + tally on ``(source, variant)``; return labelled rows. Only rows whose tally has a parseable ``still_watermarked`` are returned.

Load and preprocess image.

Load an original and AI-edited image pair.

Load all images.

Load images from directory.

Load image pair.

Load watermark signature from pre-extracted files.

Lazy-load and cache the SD-VAE. Returns (vae, device_str). ``dtype`` is 'float32' on MPS (fp16 is broken for MPS conv on older torches) and c