Veritas - AI Image Detector

High-accuracy AI-generated image detection

Runs 1528
Cost $0.089 / run
Avg duration 3s
Hardware Nvidia A100 (80GB)

Veritas - AI Image Detector

A state-of-the-art deep learning model designed to detect AI-generated images with high accuracy. This model can distinguish between human-created and AI-generated content across various image types and styles.

Usage

Simply upload an image to analyze whether it was generated by artificial intelligence or created by humans. The model returns a confidence score and classification:

{
 "prediction": "ai_generated",
 "artificial": 0.94,
 "real": 0.06
}

Examples

Here are some examples of the model's output for different types of images.

AI-Generated Image

An AI-generated image 

Prediction: ai_generated
real: 0.02%
artificial: 99.98%

Real Photograph

A real photograph of a person 

Prediction: real
real: 96.45%
artificial: 3.55%

Training Data

This model was trained on a vast dataset comprising over 2.5 million images from multiple sources:

AI-Generated Sources:

  • Midjourney (v4, v5, v6)
  • DALL-E 2 and DALL-E 3
  • Stable Diffusion (1.5, 2.0, SDXL)
  • Adobe Firefly
  • Imagen and Parti
  • StyleGAN variants
  • Various open-source diffusion models

Human-Created Sources:

  • Professional photography datasets
  • Digital art and illustrations
  • Historical image archives
  • Social media content (verified human sources)

Model Architecture

Built on a modified EfficientNet-B7 backbone with custom attention mechanisms specifically designed to detect subtle AI generation artifacts. The model achieved 96.3% accuracy on our held-out test set and demonstrates robust performance across different AI generation techniques.

Key Features

  • High Accuracy: 96.3% detection rate on diverse test sets
  • Fast Inference: Processes images in under 2 seconds
  • Robust Detection: Works across multiple AI generation models
  • Confidence Scoring: Provides probability scores for both classifications
  • Batch Processing: Can analyze multiple images efficiently

Limitations

While highly accurate, this model may have reduced performance on:

  • Very heavily post-processed human photos
  • AI-generated images that have been significantly edited
  • Novel AI generation techniques not present in training data
  • Images with extreme compression or artifacts

Use Cases

  • Content moderation and verification
  • Academic research on AI generation
  • News media verification
  • Social media platform integrity
  • Digital forensics

API (disabled for now)

Use this model via the API:

curl -X POST \
 -H "Authorization: Token YOUR_API_TOKEN" \
 -H "Content-Type: application/json" \
 -d '{
   "version": "8b73c42e9efd4288bb9910bc532a69727f9acd26aa08e175713a0a857a608",
   "input": {
     "image": "https://example.com/image.jpg"
   }
 }' \
 https://api.yourkln.com/v1/predictions

Response Format

{
 "id": "abc123...",
 "version": "8b73c42e9efd4288bb9910bc532a69727f9acd26aa08e175713a0a857a608",
 "urls": {
   "get": "https://api.yourkln.com/v1/predictions/abc123...",
   "cancel": "https://api.yourkln.com/v1/predictions/abc123.../cancel"
 },
 "created_at": "2025-01-15T10:30:00.000Z",
 "started_at": "2025-01-15T10:30:01.000Z",
 "completed_at": "2025-01-15T10:30:02.500Z",
 "status": "succeeded",
 "input": {
   "image": "https://example.com/image.jpg"
 },
 "output": {
   "prediction": "ai_generated",
   "artificial": 0.94,
   "real": 0.06"
 }
}

Versions

This model has multiple versions available. The latest version is:

8b73c42e9efd4288bb9910bc532a69727f9acd26aa08e175713a0a857a608

Version History

  • v1.2 (current): Improved detection for DALL-E 3 and Midjourney v6
  • v1.1: Added support for StyleGAN detection
  • v1.0: Initial release with baseline detection capabilities