Detection Metrics

Pixel-model smoke test on 23 internal images (8 AI, 15 human). These figures are a regression check, not a public-generator benchmark. C2PA and generator metadata are a separate first pass and are not included here.

87%

Accuracy

72.7%

Precision

100%

Recall

0.842

F1 Score

Ensemble vs. Individual Models

The Cacuda Score outperforms every single model in isolation.

Detector	Accuracy	Precision	Recall	F1
Cacuda Score	87.0%	72.7%	100%	0.842
Ensemble (SigLIP2 + DINOv2)	73.9%	57.1%	100%	0.727
NPR (Frequency Domain)	65.2%	50.0%	87.5%	0.636
SPAI (Spectral Patch)	78.3%	100%	37.5%	0.545

Model Architecture

Five complementary detection techniques working together.

Ensemble

Weight: 0.45

SigLIP2 + DINOv2 with LoRA. Dual-encoder combining semantic and self-supervised features.

45%

NPR

Weight: 0.30

ResNet-50 analysing Neighbouring Pixel Relationships in the frequency domain.

30%

SPAI

Weight: 0.25

Vision Transformer with Masked Feature Modelling for spectral patch-based learning.

25%

DIRE

Bonus: +0.05

ResNet-50 measuring Diffusion Reconstruction Error. Secondary confirmation signal.

+5%

PatchCraft

Bonus: +0.05

SRM-CNN for texture patch analysis using the Smash & Reconstruct technique.

+5%

Score Interpretation

How to read the Cacuda Score.

> 0.7

Likely AI-Generated

0.5 – 0.7

Possibly AI-Generated

0.3 – 0.5

Probably Real

< 0.3

Likely Real

Try It Yourself