Checks images
Synas Detect 001: AI-generated image detection
An open-weight model that looks for signs an image was AI-generated or edited, and tells you how sure it is.
- Type
- Image forensics model
- You give it
- An image
- You get back
- A calibrated score and its confidence
- Looks for
- Diffusion fingerprints, inpainting seams, metadata mismatches
- Weights
- Open-weight; license published with the release
- Access
- Waitlist
What it looks for
Diffusion fingerprints
Image generators leave statistical traces in the pixels they produce, much as a camera sensor leaves its own noise pattern in a photograph. Detect 001 looks for the traces associated with diffusion-based generators.
Inpainting seams
When part of an image is regenerated or pasted in, that region's noise, texture and compression history often differs from the rest of the picture. Detect 001 looks for the boundary.
Metadata mismatches
File metadata can disagree with the image itself, or with its claimed source. A camera record that doesn't fit the pixels, or an edit history that doesn't add up, is a signal worth surfacing.
How to read a result
Detect 001 reports a calibrated score together with its confidence, never a bare yes or no. Calibrated means the number carries real-world meaning: when a calibrated detector says 0.8, roughly 80 in 100 such images are AI-generated, on images that resemble the conditions it was tested under.
The confidence tells you how far to trust the score for this particular image. A low-confidence result on a heavily compressed screenshot is a different thing from a high-confidence result on an original file.
Treat the output as evidence that decides what gets a closer look. The guide to how AI image detection works explains why even a good detector raises false alarms when AI images are rare.
Limits you should know
- Compression and screenshots weaken the signal. Every re-encode erases some of the traces detectors depend on.
- New generators can slip through. A detector can only recognize traces it has learned to look for.
- Heavily edited real photos can look suspicious. Aggressive retouching, denoising and upscaling change the same statistics a detector reads.
- A score isn't proof. Don't use it as the only basis for a high-stakes decision. Combine it with the image's source, provenance data and human review.
Who it is for
Trust and safety teams, newsrooms and fact-checkers, marketplaces and platforms that handle user images, researchers studying generated media, and anyone who needs a second opinion before using an image they didn't make.
Because the weights are open, you can test Detect 001 on your own images and your own threshold for risk before relying on it, and run it where your images already live. Learn what open-weight means.
Questions
Does a high score mean the image is fake?
It means the image shows signs that Detect 001 associates with AI generation or editing, at the confidence reported with the score. It is evidence for a closer look, not proof.
Does Detect 001 work on screenshots and compressed images?
Less well. Compression, resizing and screenshots weaken the signals every detector relies on, which is why the confidence reported with each score matters as much as the score.
Can I run Synas Detect 001 on my own hardware?
That is what open weights are for. Join the waitlist to hear when the weights and their license are released.
Get early access
Join the waitlist and we'll email you when the models and their licenses are released. One email, no spam, remove yourself any time.