About this tool
Check if any image was made with AI
Drop any photo or image to find out whether it was created by an AI. The check starts with the image's built-in origin signature — the same invisible record that Adobe, LinkedIn, and some cameras embed — and shows a verdict in seconds. If there is no signature, you can opt in to a deeper pixel-level analysis that looks for AI generation patterns directly in the image data.
Why use OmegaPix
-
Origin signature first : If the image carries a built-in signed record of how it was made, you get a verified answer immediately with no extra downloads. Shows the AI tool used, who signed it, and when.
-
Pixel-level backup : For images with no signature, an optional 165 MB pixel detector analyses the image data itself. The model is downloaded once and cached — every image you check after the first is instant.
-
Honest about uncertainty : Results show a probability band (Likely AI / Possibly AI / Inconclusive / Likely real) rather than a binary yes/no. Heavy compression or editing can reduce confidence and the tool says so.
-
Nothing uploads : Both checks run entirely in your browser. The image never leaves your device. Verify in DevTools, Network.
How it works
Drop an image
JPG, PNG, WebP, HEIC, or AVIF. Any size.
Read the signature
The tool instantly checks for a built-in origin record. If one is found, you see whether AI was used, which tool created it, and who signed it.
Run a deep check if needed
No signature? Click "Run AI check" to download the pixel detector once. It analyses the image and returns a confidence band.
Frequently asked questions
How can you tell if an image is AI?
Two ways. First: many AI tools and editing apps embed a tamper-evident origin record in the file. It explicitly states whether AI was used, which tool, and who signed it. If that record is present, the verdict is instant and based on a cryptographic signature — not guesswork. Second: a pixel-level AI detector trained on millions of AI-generated and real images can spot statistical patterns invisible to the human eye. That takes longer and carries more uncertainty, which is why we show a probability band rather than a binary answer.
What if no origin record is found?
Absence of a record does not mean the image is real. AI images often have their records stripped before sharing, and many generators do not embed one at all. In that case, you can run the pixel-level check, but keep in mind that a heavy JPEG re-compression or screenshot can reduce its confidence.
Why is the pixel detector 165 MB?
It uses a SwinV2 transformer model — the same class of architecture behind many modern image AI tools. Smaller models are not accurate enough on real-world images. The 165 MB is a one-time cost: after the first download it is cached in your browser and never downloaded again.
What is "Inconclusive" and why might I see it?
Inconclusive means the pixel pattern sits in a range where the detector cannot be confident either way. This happens most often with heavily JPEG-compressed images (social media screenshots, forwarded photos), images that have been cropped or colour-graded heavily, and images from generators the model was not trained on. It is the honest answer when the signal is ambiguous.
Does this upload my photo?
No. Both checks run entirely in your browser. The origin-record reader is a self-hosted WebAssembly module. The pixel detector runs in a Web Worker on your device. Open DevTools, Network and you will see no outbound request carrying your image.
How accurate is the pixel detector?
In benchmarks on Stable Diffusion 2.1, SDXL, SD3, DALL-E 3, and Midjourney v6 images: AUROC 0.88-0.89, about 5% false-positive rate on uncompressed real camera photos. Performance drops on heavily-compressed or re-edited images. That is why results are shown as a band with a confidence percentage, not a hard yes/no.
When to use this tool
Verifying a news photo
Check whether an image in an article carries a signed origin record, or run the pixel check to look for AI patterns.
Checking profile pictures
See if a headshot or avatar carries signs of AI generation before trusting the identity behind it.
Reviewing submitted work
Quickly check whether images submitted in a contest, assignment, or publication came from a camera or an AI generator.
When not to use this tool
Forensic / legal evidence
A probability band is not court-level evidence. For legal purposes, consult a qualified digital forensics expert.
Screenshots and re-edits
A screenshot of an AI image is a real-photo capture of a screen. The pixel detector may not flag it reliably.
Heavily compressed images
A JPEG saved at very low quality loses the subtle patterns the pixel detector looks for. Expect Inconclusive.
Technical details
Stage 1: origin signature
Reads the embedded JUMBF provenance container (present in JPEG APP11, PNG caBX, WebP C2PA chunks). Verified by a self-hosted WebAssembly toolkit. Checks the IPTC digitalSourceType assertion for AI-generation flags. Instant — no additional downloads required.
Stage 2: pixel detector
LPX55/detection-model-1 — a SwinV2 transformer, 165 MB q4f16 quantised ONNX. Input: 256x256, ImageNet-normalised RGB. Runs in a dedicated Web Worker via ONNX Runtime Web (WASM). Output: softmax probability, displayed as a calibrated band rather than a raw percentage to reflect model uncertainty.
Model caching
The pixel detector is downloaded once and stored in the browser's Cache Storage API. Subsequent checks use the cached copy with no network request. Within a single session, the ONNX session is kept in-memory so inference on the second image is instant.
Your files stay on your device
Both checks run entirely in your browser. The origin-signature reader and the pixel detector both operate client-side — your image never reaches a server. The AI model is downloaded from a CDN and cached in your browser; only the model file is fetched, not your image. Verify in DevTools, Network.