How AI Image Detectors Work: Metadata, Noise and Artifacts

AI image detectors look for evidence that generators and cameras leave inside files. This article explains the four signal families our AI picture detector combines into one score — and the honest limits of the technology.

Signal 1: Metadata fingerprints

Image files carry more than pixels. JPEGs embed EXIF blocks; PNGs carry text chunks; newer formats may include C2PA provenance manifests.

• Generators sign their work. A1111 and ComfyUI write the full generation prompt, seed and sampler into PNG text. Midjourney, DALL·E, Firefly and others embed creator strings. Finding one is near-certain evidence.
• Cameras sign their work too. EXIF Make/Model fields (Canon EOS R5, iPhone 15 Pro) are strong evidence of a real capture.
• Absence is a weak signal. Social networks strip metadata on upload, so a clean file is not proof of AI — just a hint.

Signal 2: Sensor noise

A camera sensor is a physical device. Photons hitting silicon produce noise that rises with brightness and varies across the frame. Diffusion models have no sensor: their noise is either absent (over-smooth regions) or statistically uniform.

Detectors measure the high-frequency residual left after blurring the image, then test how that residual correlates with local brightness. A positive correlation is characteristic of real sensors; flat correlation is characteristic of synthesis.

Signal 3: Texture and edge statistics

AI generators optimize for plausibility at a glance, not for physical consistency at the pixel level. That produces a distinctive mix:

• hyper-detailed focal areas (faces, eyes, jewelry),
• unnaturally flat regions (backgrounds, skin, skies) with near-zero gradient,
• edge statistics that differ from lens blur — diffusion sharpness is painted, not optical.

Measuring local contrast distributions and the share of perfectly flat pixels exposes this pattern.

Signal 4: Color statistics

Camera pipelines (demosaicing, tone mapping) produce rich, slightly noisy palettes. Diffusion models produce smoother gradients and narrower effective palettes in flat regions. Counting unique quantized colors and measuring histogram smoothness quantifies the difference.

Why no detector is 100% accurate

Every signal can be fooled:

• Compression and resizing destroy noise patterns, making real photos look synthetic.
• Post-processing (denoisers, upscalers, filters) erases generator traces.
• Newer models learn to mimic sensor noise, narrowing the gap.

This is why responsible detectors report a probability with an evidence breakdown instead of a binary verdict. When the metadata signal fires, confidence is high; when only pixel statistics are available, treat the score as a strong hint and check the source.

Try it yourself

The best way to build intuition is to experiment: scan a camera photo, then a Midjourney render, then a screenshot that passed through social media compression. Our free AI picture detector shows the full signal breakdown for each — and because it runs entirely in your browser, your images never leave your device.