Check if an image is AI generated
Three-stage forensic inspection analyzing cryptographic C2PA Content Credentials, camera hardware EXIF, and diffusion synthesis artifacts across geometry, texture, and light.
How the three-stage forensic pipeline works
Single-pass machine learning classifiers frequently generate false positives on compressed authentic photos while failing against modern diffusion models like Flux.1 and Midjourney v6. Lumoz uses a three-tier inspection architecture:
Deterministic provenance & C2PA
Files generated by tools like Stable Diffusion, ComfyUI, or Adobe Firefly often carry embedded generation prompts, sampler seeds, or cryptographic C2PA Content Credentials. Stage 1 inspects binary file headers in under 50ms, confirming origin deterministically without guessing.
Lumoz Vision perceptual scan
When metadata has been stripped by social media platforms, Stage 2 inspects physical micro-anomalies: mismatched corneal reflections, plastic skin texture lacking pores, deformed fingernails, non-euclidean teeth spacing, and depth-of-field anomalies.
Lumoz Decision cross-modal synthesis
Lumoz Decision synthesizes findings across all forensic vectors. Instead of guessing a crude binary label, it balances hardware camera signals against visual artifacts to produce a calibrated probability rating, suspected architecture, and evidence log.
Comparison: Lumoz vs traditional AI detectors
| Capability | Lumoz AI detector | Standard 1-pass classifiers | Hive / Sightengine |
|---|---|---|---|
| Forensic pipeline | 3-stage (Metadata + Vision + Decision) | Single black-box classifier | Closed commercial API |
| C2PA & PNG chunk provenance | ✓ Deterministic extraction | ✕ Ignored | ~ Enterprise tier only |
| Modern diffusion accuracy (Flux / Midjourney v6) | ✓ High (Vision + Decision calibrated) | ✕ High false-positive rate | ✓ Moderate |
| Evidence & vector breakdown | ✓ 4-vector diagnostic matrix + log | ✕ Single ambiguous percentage | ~ Categorical tags only |
| Price & access | 100% free for everyone | Freemium / Paywalled | $19 – $99 / month subscription |
Supported AI image architectures
Detects subtle skin pore smoothing, high-frequency text embedding artifacts, and ComfyUI generation parameters.
Identifies cinematic color palette bias, specular highlights, signature iris reflections, and soft hair-melt artifacts.
Extracts embedded generation prompts from PNG chunks and identifies characteristic denoising scheduler artifacts.
Inspects C2PA digital credentials, characteristic painterly texture, and optical perspective geometry.
Scans for SynthID watermark signals and photorealistic synthetic lighting patterns.
Validates embedded Content Credentials and Adobe provenance certificates.
Frequently asked questions
How do I know if an image is AI generated?
AI images typically reveal themselves through three distinct vectors: embedded metadata (such as ComfyUI node graphs or C2PA algorithmic tags), anatomical inconsistencies (distorted pupils, melted teeth, or misaligned earlobes), and micro-texture anomalies (such as plastic skin lacking authentic pores, or background details blending unnaturally into one another). Lumoz checks all three vectors automatically.
What is C2PA Content Credentials?
C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard developed by Adobe, Microsoft, Google, and camera manufacturers like Leica and Sony. It cryptographically binds digital provenance metadata to an image, recording whether it was captured by a physical lens or generated by an algorithm (such as DALL-E or Firefly). Stage 1 of our tool directly validates C2PA manifests.
Does Lumoz store or train on my uploaded images?
No. Lumoz does not use your uploaded images to train AI models. Images uploaded for forensic detection are processed in memory and temporary cache, and are automatically purged after analysis.
Can AI detectors be fooled by recompressing or screenshotting an image?
Screenshotting or re-encoding an image strips EXIF and C2PA metadata, which fools basic metadata-only detectors. However, our Stage 2 (Lumoz Vision) and Stage 3 (Lumoz Decision) examine physical lighting, specular reflection geometry, and facial anatomy, which survive re-compression and screenshots.