Forensics Toolset

ELA re-saves the image as JPEG and measures per-pixel differences between the original and the re-saved version. Every JPEG save degrades quality uniformly — regions that were edited or composited from a different source have a different error baseline and appear as brighter patches in the heatmap.

What to look for: Brighter or more saturated patches on smooth surfaces (faces, walls, clear sky); uneven brightness across a large uniform area; sharp-edged bright rectangles marking a paste boundary; JPEG ghosting — faint double-edges around pasted objects.

False positives and limitations: Text, sharp edges, fabric, foliage, and cartoon art produce high ELA naturally. Images saved multiple times at the same quality "even out" and show falsely low ELA. Screenshots and social-media re-encodes produce uniform high noise across the whole image, making them nearly unreadable. Never use ELA as the sole basis for a conclusion.

Further reading: FotoForensics ELA tutorial · 29a.ch Forensically (third-party) · Bellingcat: Forensically guide (third-party — Bellingcat caution: never rely on ELA alone)

Identifies copy-move forgery by finding visually similar regions at different positions. The image is divided into overlapping 16×16 pixel blocks; feature vectors are built and sorted, then nearest-neighbour pairs are flagged as potential clones and highlighted with colour-coded overlays.

What to look for: Matching colour-coded boxes far apart in the image — same colour means the same matched pair. Long connecting lines between distant regions are the most suspicious. Watch for repeated textures repositioned to fill gaps: sky patched in, crowd backgrounds duplicated, objects removed by cloning a nearby region over them.

False positives and limitations: Sky, walls, solid-colour fills, and smooth gradients are legitimately similar and generate noise — especially at Medium/High sensitivity. Does not detect AI inpainting, diffusion-based content-aware fill, or frequency-domain edits. Start with Low sensitivity and increase only if specific suspicious regions warrant closer inspection.

Further reading: 29a.ch Forensically (third-party) · Bellingcat: Forensically guide (third-party)

Re-encodes the image through the browser canvas API, which discards all embedded metadata bytes. The output is a clean JPEG with no EXIF, IPTC, or XMP.

Use cases: Remove GPS location before sharing; prepare a clean file for publication; preserve only copyright fields for attribution.

Limitations: Canvas re-encoding introduces a small quality reduction, controlled by the quality slider. "Keep date" re-injects DateTimeOriginal only (via piexifjs) — all other fields including GPS are removed. "Keep copyright" preserves Artist, Copyright, and IPTC byline fields only; everything else is stripped.

Detects noise patterns and compression artifacts by comparing each pixel to its immediate neighbours. A 3×3 median filter is applied to the greyscale image; the difference between each pixel and its local median is its noise residual. High residuals appear red in the heatmap, low residuals appear blue.

What to look for: A patch with visibly different noise texture from its surroundings — unusually smooth (denoised, airbrushed) or unusually noisy (composited from a different camera or ISO setting). Rectangular or edge-aligned noise boundaries suggest a region was copied from a different source. A face smoother than the surrounding hair and clothing suggests skin retouching.

False positives and limitations: Clear sky, studio backdrops, and solid-colour areas always show low residuals (blue). Fabric, foliage, and hair always show high residuals (red) — this is natural. JPEG compressed at high quality produces very low residuals throughout. Two cameras with similar sensor noise composited together may not produce a visible noise boundary.

Draw directional arrows on the image to map observed shadow and light directions. All shadows in a physically consistent outdoor scene must point away from a single light source — disagreement between vectors suggests compositing or manipulation.

How to use: Load an image, then click and drag on the canvas to draw arrows tracing shadow edges or light ray directions. Each vector's angle (0–360°) and pixel length are logged in the list below the canvas. Use ↶ Undo to remove the last vector, or Clear all to start over.

What to look for: Vectors pointing in significantly different directions indicate different light sources — a strong indicator of compositing. Vectors that nearly agree but differ by 10–15° may indicate elements photographed at different times of day. Mismatched shadow lengths relative to object heights suggest elements from different times or locations.

False positives and limitations: Multiple artificial light sources (studio lights, streetlights, windows) legitimately produce multiple shadow directions. Curved or tilted surfaces create shadows that don't follow a simple directional model. Perspective distortion can make parallel shadows appear to converge. Accuracy depends entirely on correctly identifying and tracing shadow edges.

Further reading: SunCalc.org — verify sun angle for a given location and time · Bellingcat's Online Investigation Toolkit (third-party — some tools require upload or account)

Checks for three classes of provenance signal that can be verified entirely in your browser — no data is uploaded anywhere.

C2PA / Content Credentials: The Coalition for Content Provenance and Authenticity open standard embeds a cryptographically signed manifest into image files. It records which tools touched the image and whether AI was involved. Adobe Photoshop, Lightroom, Firefly, Nikon Z-series cameras, Sony α-series, and Microsoft tools all embed C2PA. Detection here confirms presence only — cryptographic verification of the signature chain requires an external validator.

IPTC Digital Source Type: The IPTC Photo Metadata Standard includes a DigitalSourceType field that declares whether an image is an original photo, AI-generated, or a synthetic composite. Adobe tools write this field automatically for AI-generated content. Like all metadata, it can be edited or removed.

AI generator fingerprints: Known AI tool names and signatures in EXIF/XMP software, creator, and history fields.

False positives and limitations: All three checks rely on declared or embedded metadata that can be stripped, fabricated, or simply absent. A clean result does not confirm authenticity. Authentic news photographs rarely carry C2PA. Most AI-generated images in circulation have had metadata stripped before distribution.

What cannot be checked locally: SynthID (Google DeepMind) and other invisible pixel-level watermarks require their proprietary detector models — they cannot be verified in a browser. See the SynthID ↗ tab for more information and links to Google's resources. Their absence tells you nothing.

Further reading: Content Authenticity Initiative · IPTC Photo Metadata Standard · Bellingcat toolkit (third-party — some tools require upload)