Deckard18/detect-ai-image-models

AI-generated image detectors (ONNX)

Three small binary classifiers (fake / real) exported to ONNX for client-side inference with ONNX Runtime (Web or Python). They run the demo at detection.taizenda.uk, where images are analyzed in the browser and never uploaded.

Folder	Base model	Size	Input	Reported results
pe_tiny	Perception Encoder PE-Core T16 (Meta), 6M parameters	12.5 MB	3×256×256	94.6% balanced accuracy on DeepDetect-25 faces (92.0% after compression or blur). Catches about 65% of fakes from the newest generators.
pe_small	Perception Encoder PE-Core S16 (Meta), 24M parameters	47 MB	3×256×256	93.6% balanced accuracy on DeepDetect-25 faces. Catches about 70% of fakes from the newest generators.
resnet50-6ch	ResNet-50 with a 6-channel input	98 MB	6×224×224	97.8% accuracy (AUC 0.997) on a held-out split of its own training data.

"Newest generators" means GPT-Image-1.5, Nano Banana Pro, Seedream 4.5, Imagen 4 and FLUX.2.

Each folder holds model.onnx and a config.json that describes the preprocessing for the web app.

• Output: 2 logits, index 0 = fake, index 1 = real. Apply a softmax to get probabilities.
• pe_tiny, pe_small: resize the shorter side to 256 (bilinear), center crop 256×256, RGB scaled to [0, 1], no mean/std normalization.
• resnet50-6ch: resize to 224×224 (bicubic), then 6 channels: R, G, B / 255; HSV saturation / 255; SRM noise residual of the grayscale image + 0.5, clipped to [0, 1]; FFT log-magnitude of the grayscale image, min-max normalized to [0, 1].

Usage (Python)

import cv2
import numpy as np
import onnxruntime as ort
from PIL import Image
from scipy.ndimage import convolve
from torchvision import transforms

LABELS = ["fake", "real"]

pe_transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(256),
    transforms.ToTensor(),
])

def pe_input(path):  # pe_tiny, pe_small
    return pe_transform(Image.open(path).convert("RGB")).unsqueeze(0).numpy()

SRM = np.array([[-1, -1, -1, -1, -1],
                [-1,  2,  2,  2, -1],
                [-1,  2,  8,  2, -1],
                [-1,  2,  2,  2, -1],
                [-1, -1, -1, -1, -1]], dtype=np.float32) / 16

def forensic6_input(path):  # resnet50-6ch
    rgb = np.asarray(Image.open(path).convert("RGB").resize((224, 224), Image.BICUBIC))
    gray = cv2.cvtColor(rgb, cv2.COLOR_RGB2GRAY).astype(np.float32) / 255
    sat = cv2.cvtColor(rgb, cv2.COLOR_RGB2HSV)[..., 1].astype(np.float32) / 255
    srm = np.clip(convolve(gray, SRM, mode="constant") + 0.5, 0, 1)
    mag = np.log1p(np.abs(np.fft.fftshift(np.fft.fft2(gray))))
    fft = (mag - mag.min()) / (mag.max() - mag.min() + 1e-8)
    x = np.concatenate([rgb.transpose(2, 0, 1) / 255, sat[None], srm[None], fft[None]])
    return x[None].astype(np.float32)

def predict(model_dir, x):
    session = ort.InferenceSession(f"{model_dir}/model.onnx")
    logits = session.run(None, {session.get_inputs()[0].name: x})[0][0]
    probs = np.exp(logits - logits.max())
    probs /= probs.sum()
    return dict(zip(LABELS, probs.tolist()))

print(predict("pe_tiny", pe_input("photo.jpg")))
print(predict("resnet50-6ch", forensic6_input("photo.jpg")))

Intended use and limitations

These models are meant for education, research and personal checks of images. A score is a probability, not proof.

• Errors happen both ways: real photos flagged as generated, and generated images missed. Recent generators are missed about 30-35% of the time by the PE models.
• resnet50-6ch is measured on a split of its own training data. Expect much lower accuracy on images from other sources.
• Compression, resizing, screenshots, filters and edits change the scores.
• Do not use a score as the only basis for a decision about a person (legal, disciplinary, professional, journalistic or other).

Training data

To complete before publishing: list the datasets each model was trained on, with links and their licenses, and say whether resnet50-6ch started from ImageNet-pretrained weights.

License

The models in this repository (model.onnx and config.json in each folder) are licensed under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0):

• You may use, share and modify them for non-commercial purposes.
• You must credit "Taizenda Detection", link to this repository and to the license, and say if you changed them.
• For commercial use, ask first by opening a discussion on this repository.

pe_tiny and pe_small are built on Meta's Perception Encoder, released under the Apache License 2.0. Its text is in pe_tiny/LICENSE-pe_tiny.txt and pe_small/LICENSE-pe_small.txt, and it still applies to Meta's original models, which you can get from Meta directly. See NOTICE for third-party attributions.

The models are provided as is, without any warranty.

Citation

If you use pe_tiny or pe_small, please also cite Perception Encoder:

@article{bolya2025PerceptionEncoder,
  title={Perception Encoder: The best visual embeddings are not at the output of the network},
  author={Daniel Bolya and Po-Yao Huang and Peize Sun and Jang Hyun Cho and Andrea Madotto and Chen Wei and Tengyu Ma and Jiale Zhi and Jathushan Rajasegaran and Hanoona Rasheed and Junke Wang and Marco Monteiro and Hu Xu and Shiyu Dong and Nikhila Ravi and Daniel Li and Piotr Doll{\'a}r and Christoph Feichtenhofer},
  journal={arXiv},
  year={2025}
}