How It Works
Real or AI uses a multi-pass AI vision pipeline to examine images across 26 forensic dimensions — from pixel-level noise patterns to internet provenance — giving you a transparent, evidence-based verdict on whether an image was captured by a real camera or generated by AI.
Key Features
Visual Forensics
Examines texture consistency, lighting plausibility, edge sharpness, shadow direction, and depth-of-field to detect AI manipulation patterns invisible to the human eye.
26-Point Scoring
Each image is scored across 26 forensic dimensions — noise patterns, color distribution, depth of field, shadow analysis, artifact detection, watermark integrity, AI fingerprinting, internet presence, and more.
Multi-Pass Pipeline
Analysis runs in six sequential passes: pre-analysis scans (reverse search + watermark), image type classification, scene understanding, description verdict, forensic scoring, and final post-processing.
Image Type Classification
Pass 0 classifies the image into one of 12 categories (Photography, Logo Design, AI-Generated Art, Illustration, Painting, etc.) and adjusts the forensic scoring accordingly.
Internet Provenance
Before forensic scoring, a reverse image search checks whether the image appears on stock photo sites, AI art platforms, news outlets, or official sources — informing the internet presence score.
Watermark & AI Fingerprint Scan
A dedicated pre-analysis pass scans for C2PA metadata, Google SynthID, Stable Diffusion invisible watermarks, frequency-domain anomalies, and other steganographic signals from AI generators.
Multi-Source Input
Analyze images by uploading files directly or pasting URLs. Both methods receive the same thorough multi-pass analysis pipeline.
History Tracking
Every analysis is saved locally to your device, allowing you to review past results, compare findings, and export reports as CSV, JSON, or PDF.
Pipeline Logs
Every prompt, LLM input, schema, and response from each pass is recorded in the Pipeline Logs, giving you full transparency into how each verdict was reached.
Local Model Support
Connect a local LMS-compatible vision model (e.g. LLaVA, Gemma, Phi-3 Vision) to run the full pipeline entirely on-device without sending data to the cloud.
Batch Queue
Drop multiple images at once — each one is queued and processed sequentially with configurable inter-image sleep to avoid rate limiting.
Educational Insights
Each result includes a scene description, detected objects, description-based verdict, reasoning summary, and flagged region overlays explaining exactly what triggered the verdict.
The Analysis Pipeline
Each image runs through 6 sequential passes plus pre- and post-processing steps. You can inspect every prompt, input, and LLM response in the Pipeline Logs.
Pre-Analysis
Reverse image search (internet provenance) and dedicated watermark/AI-fingerprint scan run in parallel before any vision pass — results are injected as context into the forensic pass.
Pass 0 · Image Type
Classifies the image into one of 12 categories (Photography, Logo Design, AI-Generated Art, Illustration, Painting, Graphic Design, etc.) and determines whether it is computer-generated or a real photograph. This classification gates the entire forensic verdict — logos, designs, and digital illustrations are always marked AI-generated.
Pass 1 · Scene Understanding
A lightweight vision call produces a factual scene description and a list of detected objects with positions and face flags. This context anchors the deeper forensic analysis.
Pass 1B · Description Verdict
A text-only call evaluates whether the Pass 1 scene description sounds like a real-world photograph or an AI-typical composition — a cheap cross-check that adds a second independent signal.
Pass 1C · Type from Description
A second text-only call re-classifies the image type based solely on the scene description. If this classification is more confident than the visual Pass 0 result, it overrides it.
Pass 2 · Forensic Scoring
The main forensic pass scores all 26 criteria, produces the verdict, analysis summary, reasoning sentence, and flags suspicious regions with bounding boxes. It receives full context from all prior passes.
Post-Processing
EXIF metadata from uploaded files adjusts the metadata_integrity score. A reverse-search penalty can optionally lower scores when no web presence is found. The final confidence score is computed as the average of all 26 breakdown values.
The 26-Point Scoring System
Each image is evaluated on these 26 criteria, scored individually from 0 to 100. Higher scores indicate stronger signs of authenticity. The final confidence score is the average across all 26.
Texture Consistency
Evaluates whether textures across the image are natural and consistent, or show signs of AI smoothing/repetition.
Lighting Analysis
Checks if light sources and reflections are physically plausible and consistent throughout the scene.
Edge Detection
Analyzes edge quality for unnatural softness or sharpness that AI generators often produce.
Noise Pattern
Examines sensor noise patterns — real cameras produce characteristic noise that AI models struggle to replicate.
Facial Analysis
When faces are present, checks for anomalies like asymmetry, earring mismatches, or uncanny features. Scores 50 if no faces present.
Metadata Integrity
Weighted blend of LLM assessment and actual EXIF data (camera make, model, date). Authentic camera EXIF raises the score significantly.
Color Distribution
Analyzes color histograms and gradients for unnatural uniformity or impossible color combinations.
Artifact Detection
Scans for telltale AI artifacts: malformed text, extra fingers, seamlines, or impossible geometry.
Shading Analysis
Checks whether shading is consistent with the lighting conditions and the form of objects in the scene.
Depth of Field
Evaluates whether depth-of-field blur is realistic and physically consistent with the implied focal length.
Motion Blur
Checks if motion blur aligns with the implied movement direction and speed of subjects in the frame.
Background Consistency
Assesses whether the background matches the foreground in lighting, perspective, and style.
Shadow Analysis
Verifies that shadows are cast in the correct direction and at the right intensity relative to light sources.
Perspective Correctness
Checks for realistic perspective with consistent vanishing points and proportions throughout the image.
Compression Artifacts
Detects suspicious JPEG blockiness or compression patterns inconsistent with natural camera output.
Image Text Indication
Scores 100 if no text present; penalizes heavily for illogical, garbled, or AI-generated text artifacts.
Contents Reality
Evaluates whether the overall image content and scenario is physically plausible and reflects reality.
AI Fingerprint
Detects known AI model signatures, GAN/diffusion frequency anomalies, spectral peaks, or embedded watermarks typical of generative models.
Clone / Repeated Elements
Identifies copy-pasted or near-identical patches in unexpected locations — a common artifact of inpainting and diffusion models.
Scale & Proportion
Checks that all objects are correctly sized relative to each other and the scene; unrealistic proportions are a strong AI indicator.
Optical Anomalies & Reflections
Verifies that reflections, lens flares, and optical distortions are physically plausible and geometrically consistent.
EXIF Manipulation
Checks EXIF/metadata fields for signs of tampering: inconsistent timestamps, missing camera model, or GPS mismatches.
Semantic Consistency
Evaluates whether all scene elements make logical sense together — weather, environment, object placement, and context.
Internet Presence
Derived from the Pre-Analysis reverse search. High score if found on credible news/official sources; low if found only on AI art platforms or nowhere.
Watermark Integrity
Derived from the Pre-Analysis watermark scan. Detects C2PA metadata, SynthID signals, steganographic watermarks, and frequency-domain anomalies.
Image Description Analysis
Scores based solely on the Pass 1 scene description: plausible real-world scenes score high; surreal, hyper-perfect, or AI-typical compositions score low.
Tips for Best Results
Use the highest resolution image available — more pixels means more forensic data.
Upload original files when possible — screenshots and re-saves lose EXIF metadata.
AI artifacts are most visible in hands, text, eyes, backgrounds, and fine edges.
If a URL-based image returns no web presence, try uploading the file directly.
Check the Pipeline Logs to understand exactly which signals drove the verdict.
Use the local model mode for privacy-sensitive images — no data leaves your device.
Disclaimer
Real or AI provides AI-powered analysis as a helpful investigative tool, but results should not be considered definitive proof. AI detection technology is continuously evolving, and no system achieves 100% accuracy. Always use critical thinking and consider multiple factors when evaluating image authenticity.
Browser Extension
Analyze on the Go
Install the Real or AI browser plugin to analyze any image you encounter on the web. Right-click on images and get instant forensic analysis without leaving your browser.
Available for:
Plugin Preview