CPID

Cross-Prototype Intervention for Generalizable AI-Generated Image Detection

Overview

AI-generated image detectors can learn correlations that change across image generators. CPID organizes generated samples into feature-space prototype groups and uses the group structure to construct controlled authenticity-space interventions. The detector is trained to preserve the real/fake decision under these cross-prototype changes.

Architecture

The model uses a frozen CLIP vision encoder followed by two branches:

• Authenticity branch: maps image tokens to a representation for binary real/fake prediction.
• Grouping branch: maps generated samples to normalized features and assigns them to fake-only prototypes.

For a fake anchor, CPID selects a sample from another prototype group and translates the anchor using the difference between the corresponding authenticity-space centers. The translated feature shares the authenticity classifier with the original feature. Fake-target supervision and prediction consistency provide the intervention objectives.

The inference path retains only the authenticity route:

image → frozen CLIP → authenticity encoder → real/fake classifier

Code

• train_main.py: model construction, prototype assignment, center-difference intervention, training losses, and optimization loop.
• module/image_model.py: image representation and authenticity model.
• module/prototypes.py: EMA prototype bank.
• infer.py: single-image and directory inference.
• figures/CPID_overview.pdf: method overview figure.

The training interface accepts frozen CLIP tokens with shape [batch, 257, 1024] and binary labels, where 0 denotes real images and 1 denotes generated images.

Requirements

• Python 3.10+
• PyTorch 2.x
• torchvision
• transformers
• Pillow

pip install -r requirements.txt

The default vision backbone is openai/clip-vit-large-patch14.

Inference

python infer.py \
  --input /path/to/image-or-directory \
  --checkpoint /path/to/checkpoint.pth

Native 224-pixel views are aggregated by averaging their logits before computing the fake probability. Use --view center for a single center view.

License

See LICENSE. The code is provided for non-commercial research use. Redistribution, commercial use, and submission of modified code as another person's work require written permission.