GPTOne vs Illuminarty: The Heatmap Detectors Compared

Both show which regions are AI-generated. Illuminarty names the likely model, GPTOne runs on-device and reads C2PA.

Illuminarty and GPTOne are the two consumer tools that answer "which part of this image is AI" rather than just "is it". Illuminarty adds model attribution, attempting to name which generator produced an image. GPTOne runs analysis on-device so the file never leaves your browser, and reads C2PA provenance alongside the pixel assessment.

If region-level explanation is what you need, these are the two to compare. Here is where they diverge.

Key Takeaways

• Both localise AI regions. Illuminarty states it identifies which regions of an image were generated, and GPTOne returns a pixel-level heatmap.
• Illuminarty attempts model attribution, estimating which public generation model produced the image.
• GPTOne analyses on-device. The image is processed in your browser rather than uploaded to a server.
• GPTOne reads C2PA Content Credentials alongside the pixel analysis, giving you provenance and pixels in one pass.
• Illuminarty also does text detection and offers an API, with a browser extension in development.

What Illuminarty does well

Illuminarty has been in this space a long time and it built the right feature early.

Its stated capability set is specific: give it an image and it estimates the probability of AI generation, attempts to identify which public AI model generated it, and marks which regions of the image were generated. It combines several computer vision approaches rather than relying on one classifier.

The model attribution is the genuinely distinctive part, and no other consumer tool markets it as prominently. Knowing an image likely came from a particular generator changes what you do next, because different models have different tells, different watermarking behaviour and different typical use.

It also runs a text detector using the same approach, marking the passages most likely to be AI-generated, and offers an API for automated detection.

Treat model attribution results with appropriate caution. Identifying which of many generators produced an image is a harder problem than deciding whether any generator did, and confidence degrades as generators converge in output quality. A named model is a lead, not a conclusion.

Where GPTOne differs

Three things.

Analysis runs on-device. Your image is processed in the browser. It is not uploaded. If you are checking a client's property photographs, an insurance claim, a legal exhibit, an HR submission or anything covered by a confidentiality obligation, uploading it to a third-party service is a problem in itself. This is the difference that matters most in professional use, and it is the one people notice last.

C2PA provenance reading is built in. Alongside the pixel heatmap you get whatever Content Credentials the file carries, read in the same pass. Provenance and pixels answer different questions, and having both at once is how you actually reach a conclusion. We wrote up the provenance side in C2PA Content Credentials and how to verify them, and why metadata alone is insufficient in whether removing C2PA makes an image undetectable.

Pro features on the free tier. The heatmap, C2PA reading and private analysis are included rather than reserved for a paid tier. The limit is volume, not capability.

To be straightforward: the free image allowance is tight. One image scan costs 1,000 credits, the entire starting grant on a free account, so a new free account covers a single image. If you check dozens, Illuminarty's allowance or a paid plan on either service will serve you better. Ours start at $7.99 a month.

Why the heatmap is the whole point

Both tools get this right, and it is worth explaining why it matters so much, because most of the market still ships a bare percentage.

Real verification questions are rarely "is this entire image synthetic". They are:

Is this a genuine photograph with a swapped face? Is this a real property with a generated sky and a tidied lawn? Is this an authentic news photo with an object removed? Is this a real person against a background that never existed?

A single confidence score cannot separate those from a fully generated image, and it cannot separate any of them from a false positive on a heavily edited but genuine photograph. A region map can. When the highlighted area is a face and nothing else, you are looking at a different problem from a uniformly lit heatmap.

We covered this in detecting AI-modified images and the distinction between manipulation types in deepfake versus AI-generated image.

On comparing accuracy

We are not going to give you a head-to-head accuracy table, because we have not run a controlled test against Illuminarty's current models and a fabricated one would be worse than none.

Our own first-party work is a 60-image benchmark published on Medium, testing our detection across generators. It is our study, not an independent evaluation, and we label it as such. The write-up is in how accurate AI image detectors really are.

We also decline to publish a fixed accuracy percentage for image detection at all, and we would encourage scepticism toward any tool that does. Performance swings by generator, by resolution, by compression and by post-generation editing. One number spanning all of that is a slogan. What you should get from a detector is a confidence assessment plus visible evidence you can evaluate yourself.

If accuracy on your particular image type decides it, run both against twenty of your own images. That test answers your question better than any published table.

Which to choose

Choose Illuminarty if model attribution matters to your workflow, if you want image and text detection from one vendor, or if you need an API.

Choose GPTOne if the images are confidential and cannot be uploaded, if you want C2PA provenance and pixel analysis together, or if you want the heatmap without an account.

For occasional verification, use both. Two independent region maps agreeing is strong. Two disagreeing tells you the image sits in genuinely ambiguous territory, which is exactly when you should stop treating any score as an answer. The wider field is covered in our roundup of AI image detection tools.

Verifying the feature comparison

Both products document what they do, and feature claims are the part of any comparison you can check directly.

Illuminarty's feature list sets out its capability list: probability of AI generation for a given image, identification of which public model generated it, and marking of which regions were generated. It also describes a text detector using the same approach and an API for automated detection. If you want to confirm anything we said about Illuminarty, that page is the source.

The provenance side is worth understanding too, because it is where the two tools diverge most. the C2PA specification defines how Content Credentials are structured, signed and validated. Reading a manifest is a separate operation from analysing pixels, and a tool that does both in one pass gives you two independent lines of evidence rather than one.

That matters because the two answer different questions. Pixel analysis asks whether the image data shows generator artefacts. Provenance asks what the file claims about its own history. Agreement between them is strong evidence. Disagreement is interesting and worth investigating, particularly when a file carries credentials claiming camera capture while the pixels suggest otherwise.

On accuracy, we are deliberately not producing a comparison table. Image detection performance varies by generator architecture, image resolution, compression level and post-generation editing, and a single percentage spanning all of those has averaged away the variables that decide your case.

Our own first-party 60-image benchmark is published and labelled as ours rather than as independent evaluation.

If accuracy on your images is the deciding factor, gather twenty of them, half generated and half genuine, and run both tools. Twenty minutes of your own testing beats any vendor's table, including ours.

FAQ

Does Illuminarty show which parts of an image are AI?

Yes. It states that it identifies which regions of an image were generated, alongside an overall probability.

Can Illuminarty tell which AI model made an image?

It attempts model attribution. Treat the result as a lead rather than a conclusion, since attribution is harder than detection.

Does GPTOne upload my image to a server?

No. Analysis runs on-device in your browser.

How many free scans does GPTOne give?

One image scan costs 1,000 credits, the whole starting grant on a free account, so a new free account gets one. The text detector allows up to 50,000 characters per scan on every plan including Free.

Which is more accurate?

Neither has published a controlled head-to-head against the other. Image detection performance varies by generator and compression, so test both on your own images.

The short version

These are the two tools that show their working. Illuminarty adds model attribution. GPTOne keeps your image on your machine and reads provenance in the same pass.

Try a scan at GPTOne and look at the heatmap.

Meta description: Both show which regions are AI-generated. Illuminarty names the likely model, GPTOne runs on-device and reads C2PA.