nonescape-v0

Nonescape AI-generated image classifier

This model predicts whether an input image was created by an artificial‑image generator (e.g. Stable Diffusion, Midjourney, GAN, deepfake pipeline) or captured with a conventional camera. It ships with Python and JavaScript/TypeScript SDKs and open‑weights in both Safetensors and ONNX formats.

Model Details

Model Description

Nonescape is a binary image‑classification model trained to distinguish synthetic (AI‑generated) from authentic (photographic/hand‑crafted) imagery.

• Model type: Image Classifier
• License: Apache 2.0

Model Sources

• Repository: https://github.com/aediliclabs/nonescape
• Demo: https://www.nonescape.com

Uses

How to Get Started with the Model

Python library

from nonescape import NonescapeClassifier

# Load model
classifier = NonescapeClassifier("nonescape-v0.safetensors")

# Single image
result = classifier.predict("image.jpg")
print(f"AI Generated: {result.is_synthetic}, Confidence: {result.confidence}")

# Multiple images
results = classifier.predict(["img1.jpg", "img2.jpg"])

Javascript/NodeJS library

import { LocalClassifier } from '@aedilic/nonescape';

const classifier = new LocalClassifier({
  onProgress: (progress) => console.log(`Loading: ${Math.round(progress.current / progress.total * 100)}%`)
});

await classifier.initialize();

// Single image
const result = await classifier.predict(imageFile);
console.log(`AI Generated: ${result.isSynthetic}, Confidence: ${result.confidence}`);

// Multiple images
const results = await classifier.predict([image1, image2, image3]);

classifier.dispose();

Recommendations

• Use a high threshold (e.g. 0.9) for tasks like filtering images, that require a low false-positive rate
• Use the default threshold of 0.5 for tasks like content moderation, that require a low false-negative rate but can live with some false positives

Training Details

The model was trained on 1M+ images scraped from across the internet.

Evaluation

The model was evaluated on the Adversarial AI-Art dataset to make it comparable with other SOTA models. It achieves the best overall accuracy, outperforming open-source and commercial models.

The GitHub repository contains all code to replicate the model evaluation.