SPRIND Challenge Winner · Nov 2024

A multi-layer forensic engine built on science, not heuristics.

itsreal.media combines classical image forensics, Vision Transformer models, C2PA provenance verification, metadata analysis, and semantic reasoning into a single fused decision pipeline.

No single signal determines the verdict. A trained meta-model learns how to weight each layer based on real-world performance, and updates as generative models evolve.

FRAME:ORIGIN v1.2 · result
VERDICT
Likely AI-generated
CONFIDENCE
0.92
Spectral
0.88
Spatial
0.91
Statistical
0.85
Lightning
0.94
Diffusion trace
0.96

Explainability: Each signal is individually scored and available via API for downstream reasoning.

0%
Internal validation accuracy

Across full test suite under controlled conditions

0%
CMB Bench accuracy

Real-world conditions: compressed, resized, re-uploaded images

0%
Accuracy on real images

SPRIND Challenge evaluation, November 2024

0
SPRIND Challenge ranking

Winner of the Deepfake Detection Challenge. Bundesagentur für Sprunginnovation

Performance depends on image source, platform transforms, and decision thresholds.

Detection pipeline

Every image is passed through all detection layers simultaneously. Outputs are fused by a meta-model that learns inter-signal correlations.

IMAGE INPUT FRAME: DETECT C2PA + METADATA SEMANTIC ANALYSIS META-MODEL FUSION RISK SCORE
Image Input
Frame:Detect
C2PA + Metadata
Semantic Analysis
Meta-model Fusion
Risk Score
Layer 01
Frame Detection Models

Several specialized Vision Transformer models fused into one meta-model for AI-generated and AI-edited image detection.

Layer 02
Technical Analysis Signals

Dozens of low-level forensic signals. spectral, spatial, statistical, and compression analysis.

Layer 03
Semantic & Visual Interpretations

Object coherence, lighting consistency, anatomical plausibility, and visual reasoning.

Layer 04
Metadata & C2PA Provenance

EXIF integrity, encoding artifacts, software fingerprints, and C2PA Content Credentials verification.


Frame:Detect

SPRIND Deepfake Detection Challenge Winner

Ultra-high accuracy multi-model ensemble. We fuse several specialized detection models and dozens of forensic signals into one meta-model, delivering state-of-the-art detection validated by independent benchmarks.

v1.2 Production

Detects fully AI-generated and AI-edited images with ultra-high accuracy. We fuse several specialized detection models and dozens of forensic signals into one meta-model, a multi-layer ensemble validated as the best-performing system in the SPRIND Deepfake Detection Challenge.

Frame Detection Models
Multiple Vision Transformer models trained across all major generator architectures. Stable Diffusion, DALL-E, Midjourney, Firefly, Flux, and more
Technical Analysis Signals
Spectral, spatial, statistical, and compression forensics: dozens of low-level signals fused into the ensemble
Semantic & Visual Interpretations
Object coherence, lighting consistency, anatomical plausibility, and visual reasoning to catch what pixel-level analysis misses
Metadata & C2PA Provenance
EXIF integrity, encoding artifacts, software fingerprints, and C2PA Content Credentials verification
ItsReal.media is a member of the C2PA Coalition. actively contributing to the standard for content provenance and authenticity.
SPRIND BENCHMARKS Challenge results · Nov 2024
1st Place. Deepfake Detection Challenge
Real images
99.8%
Internal validation
98%
CMB Bench
96%
SPRIND ranking
1st

Bundesagentur für Sprunginnovation, independent evaluation on real-world conditions.

v0.8 Beta

Detects AI-modified regions within otherwise authentic images: inpainting, outpainting, object removal, face swaps, and generative fill. Outputs a pixel-level heatmap showing the probability of manipulation per region.

Boundary detection
Detects seams between edited and unedited regions
Noise inconsistency
Identifies regions with non-matching noise profiles
Compression mismatch
Finds re-compressed blocks within a single image
Lighting direction
Detects inconsistent light sources across the frame
Texture synthesis
Spots machine-generated texture infill patterns
Edge coherence
Measures unnatural edge transitions around objects
FRAME:EDIT v0.8 · heatmap
VERDICT
Region manipulation detected
CONFIDENCE
0.87
Low
Medium
High

Central region shows high manipulation probability. Consistent with generative inpainting.


C2PA Verification

We verify Content Credentials embedded in images using the C2PA (Coalition for Content Provenance and Authenticity) standard. Our system validates the full credential chain, from camera or software origin to the final published asset.

C2PA credentials are cryptographically signed manifests that travel with the image, recording every edit and export step. When present, they provide the strongest possible provenance signal.

ItsReal.media is a C2PA contributing member. actively participating in the standard's development and implementation.
Credential chain

Validates the complete chain of trust from origin to current state

Signer identity

Verifies the signing certificate and issuing organization

Edit assertion

Reads declared editing actions recorded in the manifest

Provenance survival

Detects whether credentials survived platform re-encoding

C2PA verification result
Valid credential chain
SIGNER
Adobe Photoshop 25.4
ISSUER
Adobe Inc.
ACTIONS DECLARED
c2pa.cropped, c2pa.color_adjustments
CHAIN STATUS
Intact. 3 manifests verified

Additional detection layers

Metadata Analysis

Examines the technical metadata embedded in every image file for signs of synthetic origin or post-production tampering.

EXIF integrity
Encoding artifacts
Software fingerprints
Metadata tampering
Semantic Analysis

Uses visual reasoning to detect logical inconsistencies that betray AI-generated or manipulated content.

Object coherence
Shadow / lighting
Text / typography
Anatomical plausibility
Reverse Image Search

Searches the web and internal databases to trace an image back to its original source and identify coordinated reuse.

Source origin
Prior appearance
Variant clustering
Coordinated reuse

Meta-model fusion

No single detection layer is reliable enough on its own. Each signal has blind spots: frequency analysis misses certain diffusion models, metadata can be stripped, semantic checks depend on scene complexity.

Our meta-model is a trained classifier that takes the raw outputs of every detection layer as input features. It learns the inter-signal correlations, compensates for individual weaknesses, and produces a single calibrated confidence score.

The fusion weights are retrained continuously as new generative models appear. When a new generator changes the signal landscape, the meta-model adapts without requiring changes to any individual detection layer.

Frame:Origin signal
Frame:Edit signal
C2PA verification
Metadata analysis
Semantic analysis
Reverse image search
Verdict
AI-generated
Confidence
0.94

Integration

API-first. Modular.

REST API
REST API

Simple HTTPS endpoints. Upload an image, receive a structured JSON response with verdict, confidence, and per-layer signal breakdown. Integrates in minutes.

Docker
Containerized

Deploy our detection engine as a Docker container inside your own infrastructure. Full control over data residency, latency, and scaling. Air-gapped networks supported.

Enterprise
On-premise

Full on-premise deployment with dedicated support, custom model training, and SLA guarantees. Designed for organizations with strict compliance requirements.

Test it on your images

Request API access or a live demo. We will walk you through the pipeline with your own images.

Request API access View API docs →