Our Benchmarks

Built to Detect What It Hasn't Seen Before.

AI generation models are evolving constantly. A detector shouldn't only recognize the generators it was trained to detect. We tested our model on independent public and external datasets that were never used to train it. This shows how well the model can detect AI-generated media from new generators and unfamiliar data.

How To Read These Numbers

AUC (Area Under The Curve)

a single score for how well the model separates real images from fake ones.

1.00 = perfect separation
0.50 = chance

Zero-Shot

tested on generators and datasets the model never saw during training.

Generator

the specific AI tool that produced the image (e.g. Midjourney, StyleGAN). Different generators leave different signals, which is why results are broken out below instead of blended into one score.

96%

Image AUC

Image Detection

92.3%

Voice Accuracy

Voice Detection

10+

Languages

Multilingual coverage

How Cyberette Compares To Other AI-
Generated Image Detectors

To make the comparison consistent, we evaluated each detector on the same Cyberette benchmark using the same images and the same AUC metric.

Cyberette

0.0%

Community Forensics

0.0%

AIDE ICLR’25

0.0%

NPR CVPR’24

0.0%

The comparison highlights one of the central challenges of AI-generated media detection: performance on a known benchmark does not necessarily translate to new data.

Tested Across Generators, Not Just Datasets

Different AI generators leave different signals behind.

We evaluate our model across multiple generation architectures to understand how detection performance changes as the underlying technology changes.

GenImage

Different AI generators leave different signals behind.
On GenImage, one of the standard public benchmarks for AI-generated image detection, we reach a mean AUC of 96% across seven generators.

Diffusion #1

SD1.5

0.000 AUC

Diffusion #2

MidJourney

0.000 AUC

Diffusion #3

Wukong

0.000 AUC

Diffusion #4

VQDM

0.000 AUC

Diffusion #5

GLIDE

0.000 AUC

GAN #6

BigGAN

0.000 AUC

Diffusion #7

ADM

0.000 AUC

* 1.00 = perfect separation 0.50 = chance

High Performance Beyond Cyberette's Own Datasets.

Evaluated on independent third-party datasets using modern SDXL and FLUX generations, across both facial imagery and general scenes. At the calibrated threshold, false-positive rates stayed low: 0.4% on facial images, 2.6% on general scenes.

GENERATOR TYPE AUC

SDXL General scenes (MS-COCO) 0.000

SDXL Faces (FFHQ) 0.000

FLUX General scenes (MS-COCO) 0.000

FLUX Faces (FFHQ) 0.000

Performance Across Generations Of AI

ForenSynths:

Evaluated zero-shot against GauGAN, BigGAN, ProGAN, StyleGAN, CycleGAN, StyleGAN2, and StarGAN. Mean AUC of approximately 97% across the eight generators, without training on GAN data. Real-world detection can't depend on knowing which generator produced an image before analysis begins, which is why this cross-architecture is important.

Voice Deepfake Detection
Across Languages

Audio model was evaluated on a held-out multilingual test set. The test included speakers, recording sources, and text-to-speech systems that were kept separate from training.

0.0% ACCURACY

0.0 AUC

0.0% FALSE-POSITIVE
RATE

0.0% FALSE-NEGATIVE
RATE