Looking for AI-edit detection and tamper heatmaps? See Car Damage Photo Integrity v9 for image forgery localization in car-damage photos.
This repository retains its original checkpoint and inference instructions.
Claim Photo Integrity Detector
Detects whether a photo is an authentic camera photograph or an AI-generated synthetic image. Built for insurance claim workflows: verify photo integrity before damage severity assessment or total-loss prediction (trust-first FNOL triage).
Usage
from transformers import pipeline
pipe = pipeline("image-classification",
model="obuladinnesai/claim-photo-integrity-detector")
pipe("path/to/claim_photo.jpg")
# -> [{'label': 'real', 'score': 0.97}, {'label': 'ai_generated', 'score': 0.03}]
Training
- Base model:
google/vit-base-patch16-224, fine-tuned with a 2-class head - Data: balanced subset of the Defactify image dataset (real MS-COCO photos vs images from SD 2.1, SDXL, SD 3, DALL-E 3, Midjourney v6)
- See training script for exact config and metrics
Limitations
- v1 is trained on general scenes, not damage photos specifically. A damage-specialized v2 (fine-tuned on real vs manipulated claim photos) is planned.
- Like all AI-image detectors, accuracy degrades on generators newer than the training data.
- Not a legal proof of fraud — a triage signal for human adjusters.