Model card

TrueEye 2.1

The detector behind every Perspicax scan. This card covers what it does, what it is for, and what we observe.

Version: TrueEye 2.1
Released: 24 October 2026
Made by: The avancer.ai research lab
Output: Three bands, plus a score for API users

What it does

TrueEye detects fully AI-generated, photo-like still images. It reads the whole image and answers in three bands: likely AI-generated, inconclusive or likely real.

Intended use

Newsrooms and fact-checkers verifying submitted photos
Trust and safety teams screening uploads
Marketplaces checking listing photos
Anyone who wants to check a photo before believing it

Scope

Built for photo-like still images that were fully generated by AI.
Screenshots, documents and graphics return inconclusive, with the reason "outside scope".
Not built for edited photos or for video.

Observed results

From our own result files, at a decision point of p = 0.5.

Up to 99%of AI images caught, depending on the model
99.4%of original camera photos called real: fewer than 1 in 100 flagged
91.8%of AI images caught from generators it never saw in training

The newest models

of ChatGPT images caught in testing (GPT Image)92.1–93.9%
of Gemini images caught in testing (Nano Banana and Imagen)93.6–98.3%
of Seedream images caught in testing91.0–98.9%
of Grok Imagine 2.0 images caught in testing96.2%
of FLUX 3 images caught in testing96.6%
Other generatorsup to 99.9%

Public test sets

Camera photosOriginal camera photos correctly called real.99.4%
AIGenImages2026A 2026 benchmark of real and AI images. AUC 99.7.98.0%
ImageTimeAI images caught.97.8%
OpenFake, RedditReal and AI images posted on Reddit.97.1%
So-Fake-OODReal and AI images from social media.95.6%
OpenFakeIts core set of real and AI still images.95.1%
DailyBenchSeven current generators.93.7%
RealChainImages passed along real sharing chains.93.5%
Never-seen generators17 generators the model never saw in training, 31,334 images.91.8%

See every number on the accuracy page

How results are reported

Every scan answers in one of three bands: likely AI-generated, inconclusive or likely real.
Consumers see a coarse signal (low, middle or high). API users also get a score from 0 to 1. The score is not a probability.
An inconclusive answer always says why: the evidence fell between the thresholds, or the image is outside scope.

Provenance first

When a file carries Content Credentials (C2PA) or an IPTC digital source type, Perspicax reads and shows them before TrueEye's answer.

Content Credentials (C2PA)Live
SynthID watermark checkSupported

Privacy

Images you scan are not used to train TrueEye unless you opt in.