About Nebula AI Detector
An independent project. It exists because telling AI-generated images apart from real photographs has become genuinely hard, and most existing detectors are either closed-source, expensive, or both.
The approach
The approach here is deliberately boring: two independent signals, shown side by side instead of merged into one opaque number.
- File metadata — AIGC labels under the TC260 standard, generator signatures (Midjourney, Stable Diffusion, DALL·E, ComfyUI, NovelAI, Flux and others), and camera EXIF.
- A pixel-level model — which works even when the metadata has been stripped.
When the two disagree, you see that they disagree — that is more honest than a single confident score. A hit on a generator signature is conclusive; pixel evidence on its own is inference. The tool keeps the two apart rather than blending them together.
How it runs
It runs on modest hardware and is funded out of pocket, which is why free checks are limited to 3 per day. There is no account system, so nothing about you is tied to a profile. Images pass through an image host and auto-expire within 5 minutes; our server never stores the original.
Accuracy
No claims of 100% accuracy. Re-compressed, screenshotted, or partially repainted images will fool any detector, this one included. Treat the result as evidence, not as a verdict.
That is not a disclaimer — it is a design trade-off. We wrote a dedicated article explaining "why AI image detectors fail", including why this one gets things wrong too. If you want the technical detail, start there.
Feedback
Feedback, bug reports, feature requests and corrections to this page are welcome on the GitHub repository linked in the footer.