Free AI Image Detector - Which Model Made It?

How the detector works

Six checks run at once, and none of them look at the picture the way you do.

GAN and diffusion fingerprinting looks for the mathematical residue of the model that produced the image, in the frequency domain rather than the visible one. Error Level Analysis re-saves the file at a known quality and measures where the compression response is inconsistent, which catches edits and composites. PRNU asks whether a camera sensor was involved at all, because every real sensor leaves a noise fingerprint across the whole frame and no generator reproduces one.

Frequency-domain analysis exploits a physical fact: a lens is a low-pass filter. So genuine photographs roll off smoothly at high frequencies. Diffusion models leave energy where a lens would have removed it. Metadata forensics reads what the file claims about itself, and notes when a photograph carries no record of any hardware ever having taken it. Neural classification supplies a learned second opinion trained on labelled output from current generators.

Running them together matters more than any one of them. Each has a failure mode, and the failure modes are mostly independent. Heavy compression weakens ELA but leaves frequency signatures intact. A screenshot destroys metadata but not decoder artifacts. Adversarial noise added deliberately to defeat a neural classifier does nothing to PRNU.

You get all six back separately. When five agree and one dissents, you can see that, which is more useful than an average that hides it.

Why looking harder stopped working

The gap between what people can spot and what models can produce closed some time ago.

Tested properly, humans identify AI-generated images about 62% of the time, which is barely better than guessing and considerably worse than most people assume about themselves. Confidence does not track accuracy here. The visual tells everyone learned - hands, text, over-smooth skin - are exactly the failures each new model version targets first, so relying on them means relying on a list that goes out of date with every release.

Meanwhile the volume moved. AI images are now routine in marketplace listings, dating profiles, rental ads, insurance claims, review photos and news feeds. Most of them are harmless. The problem is that harmless and fraudulent look identical, and the only thing separating them is evidence you cannot see by looking.

What each model leaves behind

Different architectures fail in different ways, which is what makes attribution possible at all.

Midjourney, across all versions, produces the most photographic output of the major generators and the most distinctive frequency signature. Its aesthetic tuning shows up as a characteristic smoothness in mid-frequency detail: skin, fabric and foliage carry less stochastic variation than a lens would produce. Hands improved dramatically from v5 to v6, so hand-counting is no longer the tell people still repeat.

DALL-E 3 renders text better than almost anything else, which removed the old giveaway of garbled signage. What it did not remove is the decoder signature underneath, and its tight coupling to prompt adherence tends to produce compositions that are slightly too centred and too obliging.

Stable Diffusion is the awkward one, because it is open. SD 1.5, SDXL, SD3, LoRAs, community checkpoints and heavily fine-tuned derivatives all exist in the wild, and a custom checkpoint drifts from the base model's signature. What does not drift is the VAE decoder, which introduces a consistent grid-like artifact during the eight-times upscale from latent space to pixels. That survives fine-tuning because it is downstream of it.

Flux.1 came from the original Stable Diffusion authors and uses a diffusion transformer rather than a U-Net. So classifiers trained purely on older diffusion output perform poorly on it. Flux renders readable in-image text, which defeats another traditional visual tell. Dev, Pro and Schnell differ measurably from each other: Schnell is distilled and takes fewer sampling steps, which leaves more compression-like artifacts than Dev.

Also covered: Google Imagen, Adobe Firefly, and the video generators on the video pages.

The honest caveat is that this list is a snapshot. New model versions ship faster than any detector retrains, so a brand-new generator always has a window where accuracy against its output is lower. Detection generalises better than attribution, so treat the verdict as the stronger claim and the model name as the weaker one. The evidence chain marks which is which.

How to read your report

The FauxLens report has three sections: the overall verdict, the confidence score, and the per-layer evidence chain. The verdict is one of three values - "AI-Generated," "Likely AI-Generated," or "No AI Detected." A result of "AI-Generated" means multiple forensic signals fired with high agreement and confidence is above 85%. "Likely AI-Generated" means signals are present but the image has characteristics that reduce certainty. Often because it was heavily post-processed, re-compressed from a social media platform, or is an unusual crop of a larger synthetic image. "No AI Detected" means none of the six forensic layers identified reliable AI generation signatures, but it does not rule out the possibility of an extremely well-masked AI image.

The confidence percentage reflects the Bayesian fusion of all six signals. A confidence of 92% does not mean the image is definitely AI-generated. It means the evidence strongly points in that direction. The per-layer breakdown shows which signals fired. If only the frequency domain and GAN layers fired but PRNU passed, the image may be a real photo that was AI-upscaled rather than fully generated. Use this breakdown to understand the nature of the AI involvement, not just its presence. If you receive a high-confidence result that contradicts other evidence you have, escalate to the raw per-layer data and consider whether the image was significantly processed before you received it.

Where this actually gets used

AI image detection is not an academic exercise. It addresses active, costly fraud across multiple domains.

Romance scams are the most financially damaging use case. Scammers generate consistent AI personas using Midjourney or Flux (a military officer, a successful engineer, an overseas contractor) and sustain multi-month relationships that cost victims tens of thousands of dollars each. The FTC reported $1.14 billion in US romance scam losses in 2024. Because these AI personas have never been photographed, reverse image search cannot catch them. Forensic pixel analysis is the only reliable detection method.

In hiring fraud, North Korea-linked threat actors have been documented by the FBI using real-time deepfake video filters to fraudulently obtain employment at US technology companies, gaining access to internal systems and source code. Security researchers and hiring teams across multiple industries are now reporting a significant rise in suspected deepfake candidates in video interviews.

In insurance fraud, AI-generated damage photos (flooded basements, crumpled fenders, fire-damaged roofs) are submitted to claims departments. Industry analysts and insurers warn that AI-fabricated claim documentation is a growing source of fraud losses that manual review alone cannot address.

In journalism, at least 12 major outlets published AI-generated photographs in 2025 that were later retracted - all authenticated by eye rather than by forensic tools.

When the answer is not clear-cut

No AI image detector achieves perfect accuracy on all inputs, and FauxLens is no exception. Understanding where accuracy degrades helps you use the tool correctly.

Heavily re-compressed images (those downloaded from social media platforms like WhatsApp, Instagram, or Twitter) have been through multiple JPEG compression cycles. Each cycle degrades the ELA and PRNU signals. The GAN fingerprint and frequency domain signals are more resilient to re-compression but are not immune. Heavy re-compression is the single biggest reason a scan comes back inconclusive rather than wrong, which is why the report tells you which findings survived the compression and which were degraded by it.

Adversarially processed images are the hardest case. Techniques like adversarial perturbation specifically designed to fool AI detectors exist in research contexts and are beginning to appear in sophisticated fraud operations. Our multi-signal approach is more resilient than single-signal detectors. But a determined adversary with access to our detection pipeline could potentially reduce our confidence.

Hybrid images, those that start as real photographs with AI inpainting applied to specific regions, show mixed signals. The non-AI portions may pass PRNU and ELA checks while the AI-inpainted regions show GAN artifacts. The overall verdict may be "Likely AI-Generated" with moderate confidence rather than a high-confidence determination.

When confidence is low (below 70%) and the stakes are high, do not rely on FauxLens alone. Cross-reference with reverse image search. Request the original uncompressed file. Ask for additional context about the image source. For legal or high-value financial decisions, engage a certified digital forensic examiner who can apply chain-of-custody procedures and provide documented expert analysis.