ChatGPT Image Detector

No detector can identify the model behind an image from appearance alone with certainty. This checker estimates whether a picture is AI-generated and shows model-related scores only when the analysis supplies them. To start with your own file, check this photo with the free detector and compare its explanation with what you know about the image.

Check a Photo in Seconds

Open the checker and upload the clearest copy of the image you have. It accepts a photo from your device, so you can use a saved picture rather than relying on a cropped preview in a social post. The checker sends the image for analysis and shows an overall AI probability with a cautious verdict. If the provider returns generator-specific scores, those may appear as additional signals. You can see available model signals in the result instead of inferring a creator from the picture's style alone.

Read the overall assessment first. A model-related score is a comparison with patterns associated with a generator family; it is not a receipt from that company's service. If the result is inconclusive, keep that uncertainty in your conclusion. A low-quality screenshot, a crop, or a heavily edited export can give the analysis less of the original image to work with, even though the checker can still accept the file.

Before uploading, decide what you need to know. If the question is whether a scene probably came from a camera, the overall result is relevant. If someone claims a named service made the image, a model-related score may help you investigate that narrower claim when one is available. Save the source post or original file if you can, because the surrounding evidence may matter more than a small difference between model scores.

Why "Which Model" Matters

People ask about the model because the answer can change how they trace an image. A claimed ChatGPT image might have been posted as an original photograph; a supposed Midjourney image might actually be a camera photo with aggressive editing. Knowing the likely creation route can point you toward the original creator, an earlier upload, or a record that explains how the image was made. It can also tell you whether a source's caption is making a claim the picture itself cannot support.

Names are easy to misuse. ChatGPT is a product where people can create or edit images, while DALL·E and GPT image generation refer to generator families. Gemini is an app that has used Google's image models, including the Nano Banana family. These product names do not all describe a single fixed image pipeline, and the models behind a product can change. A detector must compare what is in the file with examples of supported generators; it does not learn which button a person pressed.

Model attribution is also different from asking whether an image was entirely generated. A real photo can be edited with AI, and a generated background can be placed behind a photographed subject. The final pixels then reflect more than one process. If the claim has legal, journalistic, or personal consequences, ask for the original file and publishing context rather than treating one label as a complete creation history.

Checking by Model

There is no short list of colors, faces, or lighting styles that reliably names an image generator. Classifiers instead look for patterns across pixels and compare them with examples associated with different generators. Some differences may involve fine texture, edges, or reconstruction behavior; they are usually less obvious than an extra finger or a misspelled sign. For a practical visual pass, see how to tell if an image is AI generated, then use the model sections below to understand what the detector can reasonably suggest.

ChatGPT and DALL·E

If an image is described as "made by ChatGPT," ask whether the claim refers to ChatGPT's image feature or specifically to a DALL·E generation. They are related OpenAI routes, but a product name is not a unique visual signature. The detection service lists GPT image generation and DALL·E separately among supported generator families, so a returned score may reflect which examples the pixels resemble. That comparison can guide a follow-up question; it cannot establish who generated the image or which ChatGPT conversation produced it.

Look closely at any part of the image that seems unusually coherent or inconsistent, but resist treating polished text or clean hands as a ChatGPT marker. Modern generators can improve or change these details, and ordinary editing can change them again. Separate provenance checks can sometimes provide stronger information for supported OpenAI files, but this site's analysis uses image pixels rather than reading embedded origin records. A missing record or a missing model score should not be converted into a claim that OpenAI was uninvolved.

Midjourney

If someone calls an image Midjourney-made because it looks cinematic, set that impression aside and check the file itself. Composition, color grading, and dramatic lighting can be copied by photographers, editors, and other generators. A model classifier can look for less conspicuous pixel relationships in examples associated with Midjourney outputs, including texture and edge patterns. Those relationships can shift between versions and after resizing or retouching, so a Midjourney-related score is a lead, not a maker's signature.

Check the source as well as the picture. An original post, creator account, or earlier version may tell you more about a claimed Midjourney image than its appearance does. If only a low-resolution repost remains, record that limitation before interpreting a model score. Avoid calling a picture Midjourney-made solely because it has the kind of glossy finish you have seen in another gallery.

Stable Diffusion

For a possible Stable Diffusion image, inspect the highest-quality copy and note any fine texture that changes strangely between nearby regions. Stable Diffusion models generate in a compressed latent space and decode the result back into pixels; that reconstruction route can leave subtle patterns in texture or image frequencies for a forensic classifier to analyze. A viewer should not expect to see a consistent grid or other obvious badge in every output. Different versions, checkpoints, decoders, and editing workflows can alter the trace.

The name also covers a family of models and custom setups rather than one unchanging source. A person can combine Stable Diffusion with a photograph, rerender only one area, or export through another editor. The resulting file may carry features of both the original photo and the generated part. Treat a Stable Diffusion-related score as a clue about resemblance to that family, then ask which parts of the picture were actually created or changed.

Gemini

If a picture is said to come from Gemini, ask for the original export or the post where it was created before trying to judge it by style. Gemini is the user-facing app; Google's image-generation models, including Nano Banana variants, do the image work. A pixel-based classifier may compare the result with examples of supported Google generator families, but that does not make every Gemini edit visually distinguishable from every other tool. A photo lightly edited in the app can retain many characteristics of the original camera image.

Google also offers a separate way to check supported Google origin signals, including an invisible watermark, when those signals are available. That verification method is different from this site's pixel analysis: the checker here does not inspect watermarks or embedded provenance information. A positive Google-origin signal can answer a narrower provenance question, while the absence of one does not prove a camera made the image or that another AI system was not used.

Sora

If someone says a still picture came from Sora, ask for the original video or its source rather than trying to recognize Sora from one frame. Sora was a video-generation product, and a saved frame removes the sequence of motion and continuity that could help assess a video. The image-detection service does not list a Sora-specific image score, so this page cannot promise Sora attribution from an uploaded still. OpenAI says the Sora product is no longer available, although frames from earlier videos can still circulate.

A frame can be cropped, color-corrected, or combined with other material before it reaches you. It may also have been labeled "Sora" by someone who did not create it. Judge the visible frame as an image if that is all you have, and keep the claimed video origin separate. If the original clip is available, examine it in its own context rather than making the still image carry evidence it no longer contains.

What a Detector Can and Can't Tell You

A detector can compare image pixels with patterns it learned from real and generated examples. It can return an overall assessment and, when the service supplies them, scores related to supported generator families. These results are useful for deciding what to inspect next. They cannot certify the exact model, prompt, user, account, or time of creation, and they cannot reconstruct every edit that happened after generation.

Closely related models can produce overlapping results. A screenshot can remove context; a crop can hide the part that explains a strange edge; a real photo edited with AI may not fit a clean real-versus-generated label. A high model-related score does not mean the other scores are impossible, and a missing score does not establish that the named model was never involved. Do not turn a comparison score into a statement of ownership or authorship.

This checker does not inspect file metadata, provenance credentials, or embedded watermarks. Those can provide a different kind of evidence when they are present and verifiable, but their absence is not proof of camera origin. For an important image, preserve the original file, find its earliest credible source, and compare any detector result with the visible scene and the claim attached to it. If the evidence conflicts, the honest answer is that the source remains uncertain.

FAQ