Image FFT Frequency Analyzer

C2PA (Coalition for Content Provenance and Authenticity) is a Container-Level standard (Layer 1) that stores digitally signed cryptographic assertions in external file headers (such as JUMBF blocks in JPEG/PNG/WebP). It can be completely stripped by taking a screenshot, converting file formats, or uploading to platforms like WhatsApp. Google SynthID is a Signal-Level watermark (Layer 2) that embeds perturbations directly into the mathematical frequency coefficients (Discrete Wavelet Transform sub-bands) and latent representations of the pixels or audio waveforms themselves during neural synthesis. SynthID survives screenshots, resizing, lossy compression, and format transcoding.

Generative diffusion models, GANs, and neural upscalers rely on convolutional layers, transposed convolutions, and pixel-shuffle decoders. These operations introduce periodic spatial grid repetitions that are imperceptible to human eyes in the spatial pixel domain. When converted to the frequency domain via 2D FFT, these periodicities collapse into concentrated harmonic peaks, forming prominent starburst crosshairs, grid points, or high-frequency checkerboard constellations that never occur in natural optical sensor captures.

Concentric Fourier watermark rings are circular halos centered around the zero-frequency (DC) component in a 2D FFT log-magnitude spectrum, typically located at radial distances between 0.60 and 0.75 of the Nyquist frequency. They are caused by latent-space frequency watermarking algorithms (such as Google DeepMind SynthID and Tree-Ring Watermarking) that modulate energy across narrow radial frequency bands to embed multi-bit cryptographic identification signatures.

Physical camera sensors (CMOS and CCD) capture genuine optical photons subject to quantum shot noise, thermal Johnson-Nyquist noise, and analog-to-digital converter dithering. This produces a chaotic, uniform binary distribution in the least significant bit (Bit 0) with a Shannon bit-entropy close to 1.0 (typically 0.980 to 0.999). In contrast, generative AI architectures produce pixels calculated from floating-point tensors that are subsequently quantized, resulting in unnaturally uniform clusters, zeroed bit regions, or low bit-entropy (often dropping below 0.920).

Error Level Analysis (ELA) works by recompressing an image at a known JPEG quantization level (e.g. 90%) and computing the absolute pixel difference between the original and recompressed versions. Because JPEG compression operates in 8×8 DCT blocks that reach an error equilibrium after multiple saves, any newly inserted element—such as an Adobe Photoshop Generative Fill patch or inpainting modification—will exhibit a dramatically higher or lower error level than the surrounding authentic background.

Generative AI music engines frequently utilize internal autoencoder sample rates (such as 32 kHz or 44.1 kHz) that enforce steep brickwall anti-aliasing low-pass filters at 16 kHz or 20 kHz, leaving an unnaturally empty high-frequency band. In addition, proprietary watermarking algorithms (like AudioSeal) inject continuous, inaudible psychoacoustic carrier tones or phase modulations between 18 kHz and 22 kHz. Our client-side Web Audio STFT spectrogram waterfall visualizes these carrier tones and brickwall cutoffs in real time.

No. While taking a screenshot or re-saving an image as a PNG/WebP strips all Layer 1 container metadata (including C2PA manifests, EXIF, and IPTC), it does NOT remove Layer 2 frequency watermarks. Because SynthID and DWT watermarks are embedded across low, mid, and high spatial frequencies within the pixel values themselves, the watermark persists through screen rendering, cropping, downsampling, and social media compression.

Meta platforms scan images upon upload for C2PA provenance manifests and IPTC DigitalSourceType tags. If you edited your photo using Samsung Object Eraser, Google Photos Magic Eraser, or Adobe Lightroom/Photoshop, the software automatically embeds an assertion tag signaling algorithmic modification. Even if 99.9% of the photo is authentic optical capture, Meta automated crawler detects the metadata tag and applies the "AI Info" badge. You can inspect your file with our AI Metadata Checker or strip the tag using our Content Credentials Remover.

Removing a signal watermark like SynthID without degrading visual quality is extremely difficult because the perturbation is distributed throughout the image frequency spectrum. Attempting to filter the frequency bands using Gaussian blurs, notch filters, or aggressive downscaling damages image sharpness and fine textures. In contrast, container metadata (C2PA/IPTC) can be cleanly removed with zero loss of pixel quality using our free Content Credentials Remover.

No. The scanner features a strict Zero-Upload Guarantee. The Cooley-Tukey 2D FFT, Web Worker threading, HTML5 Canvas bit-plane extraction, Error Level Analysis, and Web Audio STFT spectrogram processing execute entirely within your client-side browser memory on your device CPU/GPU. No files, pixels, or audio buffers are ever transmitted across the network.