Stable Diffusion Detector

Add the image to the Stable Diffusion detector above. A Stable Diffusion image has no signed record unless the app that made it adds one, and none of our test files had Content Credentials or an AI label in their metadata. So the answer comes from a detection model that judges the pixels, Sightengine first and Hive as a backup, and it uses one of your 3 free checks.

The reason is how Stable Diffusion is used. Its model weights are open, so anyone can run it on their own computer or inside an app, and nothing in between has to sign the result. Your browser still reads the file's metadata first, the same flow as the AI image checker on our home page.

We tested one Stable Diffusion model only: Stable Diffusion XL 1.0, with 5 original files from a public dataset and 1 screenshot. Sightengine and Hive both called all 6 AI, the lowest Sightengine score was 0.99, and the checker got 6 right. That is a good result on a narrow sample, not a verdict on Stable Diffusion in general.

The dataset's card says it holds 10,000 images made with Stable Diffusion XL for training an AI image detector. That cuts both ways. The images are typical SDXL output, but detection engines may have been trained on similar pictures, which could make them look easier than images from a model the engines have seen less often. We had no way to rule that out.

Many Stable Diffusion images come from models we never tried. Treat any Stable Diffusion detector answer on them as untested ground, and a Can't tell as a fair result rather than a fault.

• Other base models. No Stable Diffusion 1.5, 2.x or 3.x images.
• Fine-tunes and LoRAs. No community fine-tunes, LoRAs or merged models, which can change a model's look a lot.
• Edits. No img2img, inpainting or upscaled images, which mix generated and real pixels.

Sightengine says it aims to support Stable Diffusion 1.x, 2.x, SDXL and SD3. That is the vendor's statement, not something we measured.

Some local Stable Diffusion apps write the prompt and settings into the image file. AUTOMATIC1111's web UI, for example, adds the generation parameters to a PNG as a text chunk. Our checker reads a PNG text field named parameters, but it never lets that field decide the answer.

Here is exactly what happens. In your browser, the checker looks at the parameters text. If it names a known AI tool, such as Stable Diffusion, SDXL or ComfyUI, the checker notes it as a label anyone can write, and the image still goes to the detection model. A settings line that only lists a model file name isn't recognized as a label at all. Either way the text is unsigned, anyone can delete or edit it, and it is usually gone after a screenshot or a conversion to JPEG. None of our 5 SDXL originals carried a label like this.

Stability AI's model card for SDXL 1.0 lists the model's own weak spots, and they make a useful checklist when the Stable Diffusion detector says Can't tell. They describe the base model; newer models and fine-tunes may not share them.

• It doesn't reach perfect photorealism.
• It can't render legible text.
• Faces and people may come out wrong.
• It struggles with prompts that combine objects, such as one shape stacked on another.

None of these proves anything on its own. Our visual checklist covers the rest, and the Midjourney and FLUX pages show how other generators did in the same test.

A Stable Diffusion detector result is one signal, and on models we didn't test it is a weaker one. Use it to decide whether to dig further, not to accuse anyone.

• Likely AI-generated on an SDXL-style picture matches what we saw in our test.
• Can't tell is a fair answer on fine-tunes, edits and heavily compressed copies, none of which we tested. Look for the original post or a larger copy.
• Likely a real photo doesn't clear a picture from a model we didn't test.

If you need to check images in bulk, see what we learned choosing a detection API.

It reads a PNG text field named parameters in your browser. If that text names a known AI tool, it is listed as a label anyone can write, and the detection model still decides. It never settles the answer on its own.

We don't know: we tested only Stable Diffusion XL 1.0. Treat answers on other models as untested, and Can't tell as a fair result.

Its weights are open, so it often runs on someone's own computer, where no company signs the output. Some apps may add credentials or labels, but none of our test files had them.

No. A missing label tells you nothing, because many generators and most copies carry none. The model's answer is what counts, read with its range.