Frequency Analysis for AI Images: What It Shows and Where It Fails

Different image sources can leave different statistical patterns. Research and detector pipelines often inspect frequency-domain features, noise residuals, compression behavior, or hybrid image features to identify patterns that may be associated with generated or heavily processed images. These patterns are implementation-dependent and should not outrank verified provenance.

A frequency score is usually relative to a reference dataset or threshold. If it is above a threshold, that means it is elevated under that reference, not that the image has a specific probability of being AI-generated.

Detector research repeatedly separates threshold-independent metrics from deployed threshold decisions. If no calibrated threshold is available for the current image source and processing pipeline, the responsible label is uncalibrated. The result can still be shown, but it should not drive a final conclusion.

Screenshots, social-media compression, resizing, noise reduction, sharpening, and format conversion can all change frequency patterns. New generators, camera pipelines, and editing tools can also shift the distribution. This can create both false positives and false negatives.

copy-01 (screenshot-style metadata loss) and copy-04 (resized and converted copy) both report evidenceLevel frequency_suspicious with c2paStatus absent. camera-03 is a recompressed camera fixture that remains camera_like. Frequency moved on processed copies; it did not turn the camera recompress into an AI provenance label.

This is the kind of result where a low-value detector page would usually stop too early. Img2det keeps the finding tied to the evidence category that produced it, then separates strong provenance, marker-only clues, camera-like support, and uncalibrated forensic context. That separation gives a reviewer something they can cite or challenge instead of a single unexplained score.

Read frequency as uncalibrated forensic context. If the report says Above Reference Threshold, record the score and the processing history. If C2PA and markers are empty, keep the conclusion inconclusive rather than 'AI because frequency was high'.

Read the Final assessment first, then open the detailed evidence matrix. The matrix shows whether the strongest available signal came from verified C2PA provenance, an AI-related marker, EXIF or camera-like metadata, byte-level context, or frequency analysis. If the top signal is weak, the right conclusion is usually uncertainty, not a stronger accusation.

Prefer the original file. If you only have a screenshot or resized copy, run it, then compare with copy-01 and copy-04. Note that frequency_suspicious can appear after screenshot-like processing even when no AI marker or C2PA manifest is present.

When the result will be used for editorial, moderation, or public claims, save the report language exactly as evidence language. Use phrases such as marker found, no verified manifest, camera-like support, or inconclusive. Avoid rewriting those into definitive claims like fake, real, generated, or authentic unless you also have source context outside the file.

The false positive is calling copy-01 AI-generated because frequency_suspicious appeared after screenshot-style processing. The false negative is ignoring other evidence because a camera recompress like camera-03 stayed below drama in the frequency panel.

The opposite error is also common: treating a quiet report as proof that the image is camera-original. A quiet report may simply mean that useful metadata was removed, that the generator did not add supported credentials, or that the file was exported through a workflow that stripped the strongest signals.

People return after compression, resize, or meme re-encodes change the score. The published copy fixtures exist so those score changes can be compared without inventing an accuracy percentage.

That repeatable review flow is the practical value of the site. Users can run an image, compare the result with the guide language, and return when they receive a better original file or a new version from another platform. The content supports the tool instead of acting as a doorway page for one keyword.