Frequently asked questions

Atugil Intelligence is an independent AI company. It builds open-weight vision models: Synas Image 001, which generates images, and Synas Detect 001, which checks whether an image is AI-generated or edited. Read more about the company.

They are two open-weight vision models that work as a pair. Image 001 creates and edits images, with deterministic seeds and style locks so results can be repeated. Detect 001 looks for diffusion fingerprints, inpainting seams and metadata mismatches, and returns a calibrated score with its confidence. Compare the two models.

Open-weight means the trained model parameters, the weights, are published so you can download and run the model yourself. It isn't the same as open-source, which also asks for the training code and a description of the training data. Our guide explains the difference.

Join the waitlist with your email address. You don't need an account to read this site. We email waitlist members when each model and its license are released.

Reading this site is free. Pricing for any hosted access hasn't been announced, and waitlist members hear first. Running open weights on your own hardware costs only your own compute.

Each model's license is published with its release. Open-weight licenses vary a lot, so read the text before you build on a model. We say open-weight rather than open-source for that reason.

Yes, that is the point of open weights. Running the models on your own hardware means the images you generate or check never have to leave your infrastructure.

Three kinds of signal: diffusion fingerprints, which are statistical traces some image generators leave in pixel data; inpainting seams, where part of an image was regenerated or pasted in; and metadata mismatches, where a file's metadata disagrees with its content or claimed source. See how Detect 001 works.

It depends on the generator, the kind of image and how the file has been handled. Compression, resizing and screenshots weaken the signals detectors rely on, and no detector catches everything. That is why Detect 001 reports a calibrated score with a confidence instead of a yes or no, and why a single headline accuracy figure would mislead. Our guide covers where detection fails.

No. A score is evidence, not proof. Use it alongside the image's source, provenance data such as content credentials, and human review, especially when the decision affects people.

No single check is reliable. Look for provenance data such as content credentials, find where the image first appeared, check details that some generators still get wrong, such as hands, text and reflections, and use a detector as one more signal. Our guide covers each method and explains content credentials.

A style lock pins a seed together with a frozen style embedding, so a series of prompts renders in one consistent look. Same lock, same style. Read the guide to seeds and reproducibility.

It is stored so we can tell you about releases and access, and for nothing else. We don't sell it or share it, and you can ask to be removed at any time. See the privacy policy.

No. The site sets no cookies and runs no analytics or advertising trackers.

This website has no image upload feature, so nothing you do here sends an image to us.