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.