meet4150/50_epoch_aide

50 Epoch AIDE

This repository packages the 50-epoch AIDE checkpoint from:

/home/meet/Aivsre_001/AIDE/output_pico_balanced_49k_run1/checkpoint-50.pth

The model is exported as:

• model.safetensors for safer deployment
• checkpoint-50.pth as the original PyTorch training snapshot

Model

This is the same hybrid AIDE architecture used in the original repository:

• fixed 30-filter SRM high-pass module
• dual ResNet-50-style forensic encoders
• frozen OpenCLIP ConvNeXt-XXL visual trunk
• final MLP fusion head for binary classification

The model predicts:

• 0 -> real
• 1 -> fake

Input Preparation

Inference follows the original AIDE pipeline:

1. convert image to RGB
2. convert to tensor in [0, 1]
3. generate four DCT-based reconstructed views with DCT_base_Rec_Module
4. normalize all five views with ImageNet mean/std
5. stack in this order:

[x_minmin, x_maxmax, x_minmin1, x_maxmax1, x_0]

The provided inference.py keeps the exact same preparation logic.

Training Run

This packaged checkpoint comes from the local run:

• data_path=/home/meet/Aivsre_001/aide_data/train_pico_balanced_49k_v1
• eval_data_path=/home/meet/Aivsre_001/aide_data/eval_pico_balanced_49k_v1
• batch_size=8
• blr=1e-4
• weight_decay=0.0
• epochs=52
• checkpoint exported here: epoch 50

Metrics

From output_pico_balanced_49k_run1/log.txt:

• epoch 49 validation/top-1 accuracy: 99.2076
• epoch 50 validation/top-1 accuracy: 99.1061
• epoch 50 validation loss: 0.0730
• epoch 51 validation/top-1 accuracy: 99.1264

This repository specifically publishes checkpoint-50 because that is the requested 50-epoch model snapshot.

Files In This Repo

• model.safetensors
• checkpoint-50.pth
• config.json
• model.json
• preprocessor_config.json
• inference.py
• models/
• data/
• requirements.txt
• LICENSE

Usage

Install requirements:

pip install -r requirements.txt

Run local inference:

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

Or use it programmatically:

from PIL import Image

from inference import load_model, predict_pil_images

model = load_model(".")
image = Image.open("example.jpg").convert("RGB")
result = predict_pil_images(model, [image])[0]
print(result)

Notes

• This repository is for weight hosting and local inference
• the architecture is custom and not a native transformers AutoModel
• use the provided code files with the exported weights

Credits

Derived from the original AIDE implementation:

• Project: https://github.com/shilinyan99/AIDE
• Paper: https://arxiv.org/abs/2406.19435