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 v3
v3 of the claim-photo integrity detector. Continued fine-tuning of v2, which scored 97% validation accuracy but still misclassified real everyday phone photos (indoor scenes, people, video frames) as AI-generated.
v3's key addition: the builder's own everyday phone photos as "real" training examples - exactly the distribution v2 failed on - plus a stronger fine-tuning schedule (3 epochs at 2e-5 vs v2's 2 epochs at 1e-5) so the decision boundary actually moves. Training-time augmentation (JPEG recompression, noise, blur, brightness/contrast jitter) is kept.
Usage
from transformers import pipeline
pipe = pipeline("image-classification",
model="obuladinnesai/claim-photo-integrity-detector-v3")
pipe("path/to/claim_photo.jpg")
Limitations
- A triage signal for human review, not proof of fraud.
- Accuracy degrades on generators newer than the training data.