RAID-techjam/raid-detector-fusion

RAID Image Detector Models

Experimental AI-image detection models trained on the SID_Set dataset.

Model files

• semantic_stream.pt: pretrained ViT-B/16 semantic stream checkpoint.
• bayar_srm_stream.pt: Bayar+SRM low-level forensic stream checkpoint.
• bayar_srm_head.pt: standalone Bayar+SRM classifier head for low-level evaluation.
• detector_fusion.pt: fusion/classifier checkpoint trained with frozen semantic and Bayar+SRM streams.
• base_config.yaml: training configuration.

The stream checkpoints and fusion checkpoint must be used with the matching source code from the RAID repository. These are experimental weights, not a production detector.

Download the models

python -m pip install huggingface_hub
hf auth login
hf download RAID-techjam/raid-detector-fusion --repo-type model --local-dir checkpoints

This downloads the files into the local checkpoints directory.

Install the project

git clone https://github.com/Beastarz/RAID.git
cd RAID
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt

For an NVIDIA GPU, install the CUDA-enabled PyTorch wheel appropriate for the machine before installing the remaining requirements. CPU execution also works, but is considerably slower.

Run standalone Bayar+SRM evaluation

The Bayar+SRM stream expects native-resolution crops with raw RGB values in [0, 1]. It is not interchangeable with the semantic stream's normalized input. Evaluation requires a local CSV manifest with this format:

image_path,label
data/example.jpg,0
data/generated.jpg,1

Labels are binary: 0 is authentic and 1 is AI-generated or manipulated.

Set the manifest path in configs/base_config_bayar.yaml, then run:

python -m training.evaluate_bayar_srm `
  --config configs/base_config_bayar.yaml `
  --crop-size 256 `
  --backbone resnet_shallow

Reproduce fusion training

The fusion stage loads the semantic and Bayar+SRM stream checkpoints, freezes both streams, and trains only the fusion and classifier layers:

python -m training.train_fusion `
  --config configs/base_config.yaml `
  --semantic-checkpoint checkpoints/semantic_stream.pt `
  --frequency-checkpoint checkpoints/bayar_srm_stream.pt `
  --epochs 5 `
  --steps 0

The resulting checkpoint is written to checkpoints/detector_fusion.pt.

Important limitation

The current predict.py entry point still targets the original semantic plus frequency-stream DetectorPipeline. The Bayar-aware fusion checkpoint is not yet wired into that CLI. Use the training/evaluation scripts above until a Bayar-aware inference wrapper is added.

Reported experiment

On a 10,000-image SID_Set subset, the Bayar+SRM model reached approximately 0.8738 clean validation AUC. Resize robustness improved after training with resize augmentation, reaching approximately 0.8459 AUC at 0.5 scale, while more aggressive 0.25 and 0.35 scale conditions remained challenging.