API Integration for
Mimir

Real-time AI & manipulation detection for Axel Springer's editorial photo workflow. Right-click. Verify. Publish with confidence.

42 checks  ·  one verdict  ·  plain-language evidence

42
Tests per Image
14
GPU Models
~20s
Median Analysis
200,000+
Images Analyzed
What the API Returns

Real analyses. Real photos.

Every card below shows an actual API response. Beyond scores and verdicts, ImageWhisperer tells the story behind each image — what it shows, why it's suspicious, and what an editor should know.

Flood rescue fake
AI-Generated
A young girl cradles a dog in murky floodwater — emotionally compelling, but fabricated. No camera metadata. Forensic analysis reveals inconsistent water reflections and unnatural skin texture. Multiple models confirm: AI-generated.
88% confidence · 18.4s
SE:79% BF:84% SP:68% TF:14% CF:55% DR:12%
No EXIF 7/9 models flagged
Real photo
Authentic
Trump did wear a Biden hat. Canon EOS R5 metadata intact, GPS places the shot in the United States. All 9 models agree: authentic photograph, no signs of manipulation.
93% confidence · 19.6s
SE:6% BF:9% SP:4% TF:8% IV:7% DR:5%
EXIF: 35 tags 9/9 models agree
Paris with EXIF
Authentic
A street scene near the Eiffel Tower, shot on a professional camera. 47 EXIF tags confirm the location matches Paris, France. All models agree: genuine photograph with full provenance.
96% confidence · 21.3s
SE:3% BF:5% SP:2% TF:4% IV:3% Me:2%
EXIF: 47 tags GPS: Paris 9/9 models agree
Maduro fake celebration
Known Fake
Shows Venezuelan crowds supposedly celebrating a Maduro victory — but this image is already in our Known Fakes Database. Previously debunked AI-generated political misinformation. Flagged instantly in 0.005 seconds.
99% confidence · 0.005s
Known Fake DB: Match
Skipped full pipeline Instant verdict
Hollywood fire fake
AI-Generated
A dramatic Hollywood hills wildfire scene — but entirely fabricated. No camera metadata. AI watermark detected. Fire and smoke patterns show generative inconsistencies that real fire footage never produces.
90% confidence · 22.7s
SE:85% BF:88% SP:74% TF:16% CF:61% Me:22%
No EXIF AI Watermark: Detected 8/9 models flagged
AI-generated crime scene
AI-Generated
A crime scene with a chalk body outline, evidence markers, and police tape — looks like a press photo but is entirely AI-generated. No camera metadata. The lighting, shadows, and surface textures all show telltale AI rendering patterns. High scores across all models.
87% confidence · 20.3s
SE:80% BF:83% SP:69% TF:15% CF:54% Me:38%
No EXIF 7/9 models flagged
Food boxes
Further Investigation Needed
Boxes of fried snacks on a desk — a casual phone photo. All 14 GPU models confirm it's real, but there's no camera metadata. The image lacks context for editorial use: no location, no source, no story. An informational flag for the editor, not a forensic finding.
68% confidence · 26.8s
SE:15% BF:10% SP:8% TF:12% IV:6% DR:5%
No EXIF LLM concern only
Belarus president edited
AI Detects Edit(s) in Photo
Lukashenko in a formal meeting — the original photo is real, but someone edited a bag of fries into his hand. Forensic models detect manipulation artifacts around the added object. The kind of edit that turns a press photo into misinformation.
74% confidence · 17.5s
SE:35% BF:16% SP:68% IV:42% SV:29% Me:15%
EXIF: 12 tags SPAI: Editing signal
How It Works

Analysis Pipeline

The verdict isn't a single call — it's an orchestrated pipeline with dependencies. Here's what runs, in what order.

Dependency Graph

PHASE 0 Instant Gate Check ~5ms
Known Fakes DB — updated twice daily — if match, skip everything
PHASE 1 All Independent Detectors — Parallel ~0.6–1.2s
These all run simultaneously — no dependencies between them:
GPU Models (14 detectors)
B-Free SPAI TruFor ClipDet RIGID Flux Probe Gemini Probe GPTImage Probe Camera Matcher IML-ViT CommFor Sparse-ViT Mesorch HiFi-Net++
Gemini AI Sessions (4 parallel calls)
Flash — Scene + forensic Flash-Lite — Narrative Flash-Lite — ELA / Photoshop Gemini — watermark confirmation
External APIs + Local
Third-party detection APIs Google Vision API EXIF Metadata AI watermark patterns Composite detection
PHASE 1.5 Dependent — Needs Phase 1 Results ~1–2s
Web Search — waits for Vision API entities Claim Verification — needs scene + entities
PHASE 2 Verdict Engine — Sequential ~10ms
Requires ALL Phase 1 results. Priority: Known Fake → AI Watermark → External API signals → B-Free override → Gemini → Model concordance
The verdict cannot exist without the models. Removing a model changes the verdict.
PHASE 3 On-Demand Add-ons separate calls
Genuinely independent — add or remove freely:
Location Verification Claim Verification Ask About Image PDF Report Shareable Link
What you CAN separate
  • Add/remove Location Verification
  • Add/remove Claim Verification
  • Add/remove Web Search
  • Add/remove Ask About Image
  • Add/remove PDF export
  • Choose response detail: full JSON vs. summary
What you CAN'T separate
  • Verdict without model scores (computed FROM them)
  • B-Free without EXIF (override rules need tag count)
  • Composite detection without face + background
  • AI watermark without Gemini confirmation
  • Pick 3 GPU models and skip the rest (verdict changes)
Honest API Tiers

Tier 1 — Full Verdict

10–30 seconds

The complete pipeline: EXIF + 14 GPU models + external detection APIs + Gemini + ELA + AI watermark detection + Known Fakes + Composite. Returns verdict, confidence, all scores, scene description, narrative.

{ "verdict": "Authentic",
  "color": "green",
  "confidence": 0.94,
  "models": { "bfree": 0.12, "se": 0.08, ... },
  "scene": "Crowd at Brandenburg Gate..." }
This is the core product. Everything else is built on top of this.

Tier 1 Slim — Verdict Only

10–30 seconds

Same full pipeline runs internally, but the response is trimmed to verdict, color, confidence, and a one-line explanation. Same compute cost, smaller payload.

{ "verdict": "Authentic",
  "color": "green",
  "confidence": 0.94,
  "explanation": "Canon EOS R5, 47 EXIF tags..." }
Ideal for Mimir's "factchecked" metadata field.
Optional Add-ons (separate API calls, after Tier 1)

Claim Verification

+2s

Fact-checks editorial claims via Google Custom Search. Pass claim text + analysis ID.

{ "claim": "Photo from Berlin protest",
  "verified": true, "sources": 4 }

Location Verification

+1.5s

Cross-checks EXIF GPS against visual landmarks using Gemini.

{ "city": "Berlin", "country": "Germany",
  "confidence": "high", "gps_match": true }

Web Detection

+1.5s

Google Vision reverse image search: similar images, news articles, web entities.

{ "similar_images": 12,
  "news_articles": 3 }

Ask About Image

+1.5s

Follow-up Q&A: "Is this person real?" or "What software edited this?"

{ "question": "Is the crowd real?",
  "answer": "Yes, consistent with..." }

PDF Report

+1s

Professional forensic report for archives, legal docs, compliance.

{ "pdf_url": "/reports/abc123.pdf" }

Shareable Link

instant

Public URL for completed verification. Share with colleagues or fact-checkers.

{ "url": "imagewhisperer.org/v/abc123" }
Your package: Tier 1 + 0 add-ons
Mimir Integration

How it looks inside Mimir

Four variations for the right-click verification result. Pick the density level that fits your editors' workflow.

Traffic Light

Minimal
Zero cognitive load. One glance, instant decision.
Authentic
93% confidence
35 EXIF tags present. All 14 GPU models + Gemini confirm real photograph. No forensic anomalies.

Forensic Dashboard

Detailed
Model scores, EXIF, expandable. For 250 automated checks/day.
Authentic 96%
B-Free
0.05
SPAI
0.02
TruFor
0.04
IML-ViT
0.03
CommFor
0.02
Canon EOS R5 · EF 24-70mm GPS Paris 47 EXIF tags

Compliance Strip

Inline
Thin strip for MAM grid view. Click to expand model scores.
Authentic 92% 22.7s
B-Free
0.09
SE
0.06
SPAI
0.11
TruFor
0.07
IML-ViT
0.05
DRCT
0.04
CommFor
0.03
Mesorch
0.04
9/9 models agree: Authentic · Nikon Z9 · 38 EXIF tags

Newsroom Card

Editorial
Claims, sources, insights. Built for breaking news desks.
Authentic · 90%
Source Tracking
12 similar images found · Matches Reuters wire · 3 news articles reference this image
Educational Insight
Camera EXIF shows professional equipment. GPS consistent with claimed location. All 9 models agree.
When something is wrong — the same variations with a detected fake:

Traffic Light — AI Detected

Alert
AI-Generated
94% confidence
No EXIF. B-Free 0.94, SPAI 0.82, external APIs confirm. Generative facial artifacts detected. 8/9 models flag AI generation.

Compliance Strip — AI Detected

Alert
AI-Generated 93% 19.2s
No EXIF SE:89% BF:91% AI Watermark
Mimir "factchecked" metadata field — what gets written
Option 1 — One-Liner
AUTHENTIC (94%) — Canon EOS R5, GPS Berlin, 47 EXIF tags, no AI markers
Option 2 — Structured
VERDICT: Authentic | Confidence: 94%
Camera: Canon EOS R5 | GPS: Berlin, Germany | EXIF: 47 tags
AI Scores: SE 0.08, B-Free 0.12, SPAI 0.05 | AI Watermark: No
Red Alert Example
VERDICT: AI-Generated | Confidence: 91%
Camera: None | EXIF: 0 tags | AI Watermark: Midjourney
AI Scores: SE 0.87, B-Free 0.92, SPAI 0.78 | 8/9 models flagged
More interface options — beyond traffic lights

These mockups show ImageWhisperer features that go deeper than a simple verdict. Each can be returned as structured JSON for Mimir to render, or linked as an interactive panel.

Visual Location Comparison

Geo-Intel
Live screenshot from imagewhisperer.org — AI guesses location from visual clues, then shows satellite map, Google Street View, and the uploaded photo side-by-side for human comparison.
Visual Location Comparison — satellite map, Street View, and uploaded photo side by side
Satellite Map — pins the AI-estimated coordinates on Google Maps
Street View — ground-level panorama at estimated location
Uploaded Photo — side-by-side for visual comparison
The editor can drag and rotate Street View to look for matching buildings, road layouts, vegetation, and signage. GPS coordinates link to Google Maps, Google Earth satellite, and similar image searches. Works even when EXIF GPS is missing — AI estimates from landmarks, signs, and architecture.

Known Event Matcher

Database
Matches against 717+ debunked fakes from fact-checkers worldwide. Instant hit in <5ms.
Known Fake — Matched in Database
AI-generated image of Venezuelans supposedly celebrating Maduro victory
Event: Venezuela Election 2024
Date: August 5, 2024
Source: AFP Fact Check
Match: 98.7% similarity
Response: 0.005s · Phase 0 (pre-analysis)

Investigation Toolkit

Research
Reverse image search, fact-check lookup, keyword context search — one-click tools for journalists.
Verify & Cross-Check
1
Reverse Image Search
Google Lens, TinEye, Yandex, Bing
2
Fact Check Database
Fact-check article detection
3
Keyword Context Search
Web, News, Lexica, Shutterstock
Auto-extracted keywords:
food boxes aid distribution

Full Analysis Dashboard

Premium
The complete ImageWhisperer experience — every section collapsible. Link from Mimir to open as a full-page report.
Edited · 78%
Detection Analysis
SPAI: 72% editing
First Observations
Who, What, Where, When
Location Analysis
Street View + Map
Web Detection
12 matches found
The Full Picture
AI narrative + key points
Under the Hood
14 GPU models + Gemini
How it works: Mimir links to imagewhisperer.org/result/{id} — editor sees full interactive dashboard in browser. Every section expandable, with "Ask About This Image" AI chat built in.

Ask About This Image

AI Chat
After verification, editors can ask unlimited follow-up questions about any image. Powered by Gemini.
Is this hat photoshopped onto Biden or is he really wearing it?
He is genuinely wearing it. The EXIF data confirms a real camera, the lighting on the hat matches the scene perfectly, and no manipulation artifacts were detected by any of the 9 models. This was widely documented by AP, Reuters, and CNN.
Unlimited questions · Context-aware · Uses verdict + EXIF + location data

Web Detection & Source Tracing

Source Trace
Live screenshots — finds where the image appears online. Pages linking to it, exact matches, visually similar images, and partial matches. Each with "Trace Source" via Google Lens.
Pages & Exact Matches
Web detection: 15 pages, fact-checked badge, Reuters link
Similar & Partial Matches
Web detection: exact, similar, and partial image matches with thumbnails
15 pages linking
1 exact match
10 visually similar
10 partial matches
Fact-checked Reuters flagged
Each image has View Full Size, Visit Source, and Trace Source (Google Lens) buttons. Pages are sorted by date with source priority. Timeline shows earliest appearance. Unrelated visual similarity matches are automatically hidden.
Direct URL Integration — zero middleware required

No iframes, no middleware, no connectors. Mimir opens ImageWhisperer in a new tab with an origin parameter. The editor uploads the image, gets the full verification result, done. All verifications are tracked against Axel Springer's package.

Integration Flow

Right-Click
"Verify Image"
→
NEW TAB
Open ImageWhisperer
?origin=axelspringer
→
Upload & Verify
14 GPU models + Gemini
→
Verdict
Full analysis report
Mimir opens this URL in a new tab 
// Editor clicks "Verify Image" → Mimir opens a new browser tab
https://imagewhisperer.org/?origin=axelspringer

// The origin parameter tells ImageWhisperer:
// 1. Show a clean, editor-focused UI (no pricing, no marketing)
// 2. Track all verifications against Axel Springer's package
// 3. Log usage for billing & analytics

Why This Approach Works

No Middleware Required
No connectors, no workflow engines, no iframe embedding. Just a URL with a parameter. Mimir opens a new tab — that's the entire integration.
Usage Tracking Built In
Every verification is counted against the purchased package. The origin parameter ties usage to Axel Springer's account for billing and analytics.
100% Reusable
Generic pattern. Any Mimir customer can use it — just change the origin parameter. No custom integration work per client.
Infrastructure

Dedicated Hardware

ImageWhisperer runs on dedicated server hardware in a German datacenter. No shared cloud instances — your data stays in the EU.

Processor
Intel Core i5-13500
13th Gen, 20 threads. Runs all 14 detection models in parallel without bottleneck. Enough headroom for 1,400+ analyses per hour.
GPU
NVIDIA RTX 4000 SFF Ada
20 GB VRAM. Professional workstation GPU that accelerates the deep learning models (B-Free, SPAI, TruFor, IML-ViT, etc.) for sub-second inference.
Memory
64 GB DDR5 RAM
All 14 GPU models stay loaded in memory — zero cold-start latency. No model swapping, no waiting. Every analysis starts immediately.
Storage
2× 1.7 TB Samsung NVMe
Enterprise-grade Samsung SSDs in RAID. Fast image I/O, analysis cache, and Known Fakes Database with sub-millisecond lookups.
Dedicated datacenter, Falkenstein, Germany · GDPR-compliant · Daily encrypted backups · No data leaves the EU
Pricing

Recommended Packages

Two packages tailored to Axel Springer's editorial workflow, from manual spot-checks to full automated verification.

Recommended Start

Editorial Quick Check

~50 checks/day · ~1,500/month
Custom
per month · based on volume
  • Full pipeline (Tier 1) as described above
  • JSON response for Mimir integration
  • Editor triggers check manually
Response: 10–30 seconds

Photo Desk Standard

~250 checks/day · ~7,500/month
Custom
per month · based on volume
  • Full pipeline (Tier 1) as described above
  • JSON response for Mimir integration
  • Higher volume for photo desk workflows
  • Add-ons included:
  • Location verification & GPS cross-check
  • Web detection (sources, exact/similar/partial matches via image search)
  • PDF reports + shareable links
Response: 10–30 seconds
Both parties can cancel with 90 days notice. No long-term lock-in.

What Happens Beyond the Package Limit?

We never block editors. If a newsroom hits its monthly limit, verifications keep working — no interruption, no lockout.

Soft Limit — Overage Rate
Verifications beyond the package are billed at +20% of the normal per-check price for that package. Settled at the end of the month.
Example: Editorial Quick Check includes 1,500 checks at €0.33.
Check #1,501 onwards: €0.40 per verification (0.33 × 1.20).
Upgrading — On Request
Want a higher package? Request an upgrade and it takes effect at the start of the next billing month. The current month finishes on the existing package, with any overage at the +20% rate.
No automatic upgrades — you decide when to move up.
No surprises. We send a notification when usage reaches 80% and 100% of the package limit, so editors and managers can plan ahead.

Volume Estimates from Eva Bauer

Scenario Per Day Per Month Package
Manual check by editor (minimum) 50 ~1,500 Editorial Quick Check
All manual uploads, automated (maximum) 250 ~7,500 Photo Desk Standard
All agency + manual (not recommended) 40,000 ~1.2M Not practical

Why not a simple AI score?

A single AI-detection API gives you one number. That works — until it doesn't. Every model has blind spots: one misses Flux-generated images, another flags real photos with bokeh as "manipulated," a third can't tell a photoshopped composite from a clean original.

ImageWhisperer runs 42 independent tests on every image — neural forensics, metadata analysis, web provenance, fact-check lookup, camera fingerprinting, and more — then cross-validates them against each other. No single model can override the verdict alone; every flag requires corroboration. That's what turns a score into a decision your editors can trust under deadline pressure.

42
Independent tests per image
10+
GPU models cross-validated
0
Single-model overrides allowed

Building this in-house means licensing multiple detection APIs, running GPU servers, writing the orchestration and cross-validation logic, and keeping up as new generators emerge monthly. Our per-image price reflects that full pipeline — not a single API call.

Technical

Integration Guide

How the Mimir right-click action connects to ImageWhisperer's API.

1

Right-Click in Mimir

Editor selects photo, clicks "Verify Image". Mimir opens ImageWhisperer in a new browser tab.

2

Upload & Analyze

Editor uploads the image in ImageWhisperer. 14 GPU models + Gemini analyze in parallel.

3

Read the Verdict

Full verification report with traffic light verdict, confidence score, and forensic details.

4

Back to Mimir

Editor returns to Mimir with the answer. All verifications tracked against Axel Springer's package.

API Request 
# Upload image and get verification result
curl -X POST "https://api.imagewhisperer.org/v1/verify" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: multipart/form-data" \
  -F "image=@berlin_protest_2024.jpg" \
  -F "claim=Photo from Berlin protest March 2026" \
  -F "mode=quick"
API Response 
{
  "verdict": "Authentic",
  "color": "green",
  "confidence": 0.94,
  "explanation": "Consistent Canon EOS R5 EXIF with 47 tags, GPS verified Berlin",
  "models": {
    "bfree": 0.12, "spai": 0.05, "drct": 0.08
  },
  "metadata": {
    "camera": "Canon EOS R5",
    "gps": { "city": "Berlin", "country": "Germany" },
    "exif_tags": 47
  },
  "ai_watermark": false,
  "known_fake": false,
  "processing_time_ms": 18420
}

What We Build — Zero Effort for Mimir IT

ImageWhisperer builds and delivers the complete integration wrapper. Mimir's team only needs to add a right-click action that opens a URL — no connectors, no middleware, no server-to-server communication.

We Build (ImageWhisperer)

  • Origin-aware UI wrapper (clean editor experience)
  • Usage tracking per origin for billing & analytics
  • Package-based verification limits
  • Full forensic analysis (14 GPU models + Gemini)
  • REST API endpoint (optional, for automated workflows)
  • Integration documentation & support

Mimir IT Only Needs To

  • Add right-click action: open URL in new tab
  • Set URL to imagewhisperer.org/?origin=axelspringer
~5 minutes of Mimir admin work

Ready to move forward?

Reply to this document or reach out directly to start the pilot.

Start the Conversation