AIDE_FINE_TUNED_98_acc

AIDE Fine-Tuned 98 Acc

This repository contains a fine-tuned AIDE checkpoint for AI-generated image detection.

Checkpoint

• Main published checkpoint in this repo: checkpoint42.pth
• Base architecture: AIDE (hybrid SRM/DCT artifact features + ConvNeXt semantic features)
• Parameter count: about 54.4M

Training Summary

This checkpoint was resumed from a previously trained AIDE model and then fine-tuned on a balanced custom dataset built from:

• real images sampled from the earlier AIDE multisource dataset
• fake images from recently downloaded Nano-Banana / Pico-style generated images

Balanced dataset used for this run:

• train: 39,600 real / 39,600 fake
• eval: 2,200 real / 2,200 fake
• test: 2,200 real / 2,200 fake

Reported eval metric during training for this run:

• best seen in run: 98.82% at epoch 41
• published checkpoint here: epoch 42
• epoch 42 eval accuracy: 98.55%

Files

• checkpoint42.pth: uploaded trained checkpoint
• inference.py: CLI inference for a single image
• app.py: simple Gradio image test app
• models/ and data/: required custom AIDE model code

Quick Start

Install dependencies:

pip install -r requirements.txt

Run inference on one image:

python inference.py --image /path/to/image.jpg

Run the local image test UI:

python app.py

Output

The model predicts:

• real
• fake

and returns the fake probability.

Notes

This is a custom PyTorch checkpoint, not a Transformers checkpoint. The included Python files are required for loading and inference.

For a fully hosted browser demo on Hugging Face, the recommended next step is to place app.py in a separate Hugging Face Space.

License

The upstream AIDE project is MIT licensed, and the included code in this repo follows that license.