Checks images
Synas Detect 1: real photograph or AI-generated?
An image classifier that labels a picture as a real photograph or AI-generated, and reports how confident it is.
- Type
- Image classifier (Vision Transformer)
- Base model
- openai/clip-vit-large-patch14
- Size
- About 304 million parameters, Safetensors, F32
- You give it
- An image
- You get back
- A label (0: AI Generated, 1: Real Photograph) and a confidence
- Training data
- ComplexDataLab/OpenFake, Core partition
- License
- MIT
What it does
Synas Detect 1 is a binary classifier. You give it an image, and it returns one of two labels, AI Generated (Fake) or Real Photograph, along with a confidence computed from the model's output scores.
It was fine-tuned from OpenAI's CLIP ViT-L/14 on the Core partition of the OpenFake dataset. The model card names Midjourney, Stable Diffusion, DALL·E and Imagen as examples of the generators it targets.
Reported accuracy, and how to read it
The model card reports about 96% accuracy on images like its training data and about 92% to 95% on out-of-sample images, with a self-reported accuracy of 95.0% on the OpenFake Core dataset.
These are the author's own figures, measured on one dataset. Accuracy depends on the generator and on how an image has been handled, and it can fall on images unlike the training data. Test the model on your own images before relying on it. Our guide on how AI image detection works explains why even an accurate detector raises many false alarms when AI images are rare.
Run it yourself
With the Transformers library, the whole thing is a few lines.
from transformers import pipeline
detector = pipeline("image-classification", model="Atugil-Intelligence/Synas-Detect-1")
print(detector("photo.jpg"))The model card also shows a native PyTorch version using AutoImageProcessor and AutoModelForImageClassification, and links Google Colab and Kaggle notebooks. The model ID is case-sensitive on some systems, so copy it from the card exactly.
Limits you should know
These come straight from the model card.
- Compression lowers confidence. Heavy JPEG compression and social media recompression can reduce it.
- Edits remove signal. Cropping, resizing and post-processing can strip the forensic signals the model relies on.
- Not hardened against evasion. It was not explicitly trained against adversarial perturbations or deliberate attempts to fool it.
- A score isn't proof. Don't treat a confidence as absolute proof of authenticity.
It is a two-way classifier of real photographs versus AI-generated images. It doesn't locate edits inside an otherwise real photo.
Who it is for
The model card suggests three uses: media authenticity checks for journalism, publishing and digital archives; content moderation that flags AI-generated uploads; and forensic research into generation artifacts. In every case, use it as one signal next to the image's source, provenance data and human review. Because the weights are published, you can measure it on your own images and run it where your images already live.
Questions
What does Synas Detect 1 output?
A label and a confidence. The model is a binary classifier: label 0 is "AI Generated (Fake)" and label 1 is "Real Photograph". The confidence comes from a softmax over the two scores.
How accurate is it?
The model card reports about 96% accuracy on images like its training data, and about 92% to 95% on out-of-sample images. These are self-reported figures from the card, with an evaluation of 95.0% accuracy on the OpenFake Core dataset. Real-world accuracy depends on the generator and on how the image has been handled, so test it on your own images.
Does a result prove an image is fake?
No. The card says confidence scores should not be treated as absolute proof of authenticity. Use the result as one signal alongside the image's source, provenance data and human review.
Does it work on screenshots and compressed images?
Less well. The card lists heavy JPEG compression and social media recompression as reducing confidence, and notes that cropping, resizing or post-processing can remove forensic signals.
Is it robust against deliberate evasion?
Not demonstrated. The card says it was not explicitly trained against adversarial perturbations or deliberate evasion.
What license does it use?
MIT. The card credits the base model, OpenAI's CLIP ViT-L/14, as copyright 2021 OpenAI and the fine-tuned weights as copyright 2026 Atugil Intelligence.
Try it
Open the model card for the weights, usage code and notebooks. No sign-up needed.