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 v4
v4 of the claim-photo integrity detector. Continued fine-tuning of v3, which fixed the classroom false positive but overcorrected: with no photorealistic AI people in its fake diet, v3 learned "people = real" and waved AI headshots through (91.83% real).
v4's fix: ~5,000 AI-generated faces added to the fake diet - 2,500 StyleGAN faces (140k Real and Fake Faces, HF mirror) + 2,500 diffusion faces (DeepFakeFace text2img/insight, HF) - covering both GAN and diffusion generators, since detectors trained on one family fail on the other. A few hundred AI faces were also added to validation so the val metrics actually measure face detection.
Training-time augmentation kept: JPEG recompression, noise, blur, brightness/contrast jitter, flips.
Usage
from transformers import pipeline
pipe = pipeline("image-classification",
model="obuladinnesai/claim-photo-integrity-detector-v4")
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.
- Check dataset licenses (140k/CC, DeepFakeFace/OpenRAIL) before commercial use.