The Observatory
Detect whether an image was generated via AI or Human.
Inference Backend:
Detecting...
All processing runs locally in your browser
image.png
AI Detection Engine
Loading...
AI Detection Ensemble
Ateeqq ViT (99.23%) + MLEP Entropy (97.1%)
Confidence-weighted ensemble combining a SigLIP vision transformer trained on 2024 generators with an entropy-based ResNet detector that generalizes across 32+ generator architectures.
WebGPU Accelerated
Detecting...
100% Private
No server upload
Model Size
~62 MB ONNX
Analyzing image...
Running detection models
Analysis Complete
Results from browser-based detection
--
Verdict
Ensemble Prediction
Confidence-weighted ensemble with adaptive forensic blending
Human
AI
Total Time
--
Avg/Model
--
Individual Model Analysis
Each model's prediction, weight, and training dataMethodology & Interpretation
Understanding ensemble weights and how to read resultsEnsemble Formula
NN = Σ(Pᵢ × Wᵢ × Cᵢ) / Σ(Wᵢ × Cᵢ)
P = AI probability
W = model weight
C = confidence
Forensic weight adapts from 15% to 35% based on model disagreement.
Model Info
- Ateeqq (W=1.5) – 99.23% accuracy, SigLIP ViT-B/16 (86M params)
- Trained on MJ 6.1, Flux 1.1, SD 3.5, DALL-E 3, GPT-4o, Ideogram 2.0
- MLEP Entropy (W=1.2) – 97.1% accuracy, Truncated ResNet-50 (1.4M params)
- Entropy-based detection (NeurIPS 2025), generalizes to 32+ generators
Interpretation Guide
- ≥70%: High confidence AI-generated
- 30-70%: Uncertain - may be edited or hybrid
- ≤30%: Likely human-created
- Low confidence: Image may be heavily post-processed
Known Limitations
- Generators released after model training may evade detection
- Heavy post-processing can fool detectors
- Hybrid content (real + AI elements) challenges classification