Is this photo real, or did a machine make it?
Drop an image in. SynthCheck takes the file apart — credentials, metadata, how it was encoded, what size it came out at — and shows you what it found. Then it tells you how much that evidence is actually worth.
No image handy? Run an example
Nothing is uploaded — the file is read on your device · No account, no limit, no catch · Works with the internet switched off
How the check works
Nothing is uploaded. When you drop a file in, your own browser opens it and walks through it in four passes, from what the file says about itself down to how it was physically written.
- 1Pass one — the container
- The file is taken apart at the byte level: JPEG segments, PNG chunks, WebP boxes, ISOBMFF atoms. This finds every place metadata can hide, including compressed PNG chunks that most tools skip because they have to be inflated first.
- 2Pass two — what the file declares
- Signed C2PA Content Credentials, IPTC Digital Source Type, XMP packets. These are formal declarations of origin. When one is present, the question is largely settled — the generator labelled its own work.
- 3Pass three — generator leftovers
- Stable Diffusion, ComfyUI, InvokeAI and Fooocus save the whole prompt, the seed, the sampler and the model name into the file and nobody ever removes it. We read it back word for word. Any generator that names itself anywhere in the metadata is matched against the list below.
- 4Pass four — the shape of the file
- This is the one that still works after a social platform has scrubbed everything. Output dimensions that only diffusion models produce. JPEG quantization tables that no camera firmware would write. Progressive coding and thin Huffman tables that betray a software encoder. Each is weak alone; several together are a pattern.
Generators recognised by name
When a file names its maker anywhere in its metadata, we identify it. This list is generated from the detector’s own source at build time, so it is exactly what the code looks for — never a marketing claim.
- Adobe Firefly
- AUTOMATIC1111 WebUI
- Canva Magic Media
- CogView
- ComfyUI
- DALL·E
- DeepSeek (Janus)
- Draw Things
- FLUX (Black Forest Labs)
- Fooocus
- Freepik AI
- Google DeepMind / SynthID
- Google Gemini
- Google Imagen
- GPT Image
- Grok / xAI
- HiDream
- Hunyuan
- Ideogram
- InvokeAI
- Kimi (Moonshot AI)
- Kling
- Kolors
- Krea
- Leonardo.Ai
- Luma AI
- Meta AI (Emu)
- Microsoft Designer / Bing Image Creator
- Midjourney
- Nano Banana (Gemini image model)
- NightCafe
- NovelAI
- OpenAI
- Pika
- PixArt
- Playground AI
- Qwen-Image
- Recraft
- Reve
- Runway
- Sana
- Seedream (ByteDance)
- Sora
- Stable Diffusion
- Veo
- Wan
Claude is absent on purpose: it does not generate images, so there is nothing to find. Anthropic’s watermark announced for 2 August 2026 applies to generated text, not pictures. When metadata has been stripped, no name can be read at all — but output dimensions often still narrow the field to two or three candidates, and that shortlist is shown in the report.
How to read the result
Every check ends in one of six verdicts. None of them is a percentage, because a percentage would be invented.
- AI-generated
- The file says so itself. This marker was written by the generator that made the image, and nothing else produces it.
- Consistent with AI generation
- No provenance markers survive, but the structure of the file itself lines up with generated output on several independent counts — see below. This is a pattern, not a declaration: each trait has innocent explanations, and it is their combination that is telling. Treat it as a strong lead, not a finding.
- AI-assisted
- The metadata points to generative editing applied on top of other source material, rather than an image conjured from nothing.
- Camera capture evidence
- Capture fields consistent with a physical camera, and no AI markers anywhere. Worth knowing: EXIF is plain text and can be copied from a real photo onto a fake one, so this is supporting evidence rather than proof.
- Conflicting signals
- This file carries AI origin markers and camera capture fields at the same time. Usually that means generated content was composited into a real photo, or metadata was transplanted between files. Read the evidence carefully — this is the interesting case.
- Inconclusive
- Nothing survives in this file to read, and its structure carries nothing distinctive. That is not a hint in either direction — it is an absence of evidence, and it is the most common outcome for anything saved off social media.
Under the verdict, every piece of evidence is filed by how much weight it can carry. Definitive means the producing software wrote it and a camera cannot. Supporting points the same way but is forgeable or indirect. Contextual is file history that narrows nothing on its own. Read the definitive block first; if it is empty, the answer is a judgement rather than a fact.
What this can and cannot prove
Every other detector shows you a confidence percentage. Ask where that number comes from and you will not get a straight answer. We would rather show you the evidence and let you judge it.
Solid ground
- Confirming AI origin when the generator labelled its own output
- Reading signed C2PA manifests and every action recorded inside
- Recovering prompts, seeds and model names left in the file
- Showing camera body, lens, exposure and GPS
- Spotting generator-preset dimensions and software encoders after metadata is gone
Out of reach
- SynthID. Google DeepMind's watermark lives in the pixels, not the metadata, and there is no public reader. Use Google's own SynthID Detector.
- A verdict from pixels alone. We do not run a classifier and we will not invent a percentage.
- Certificate validation. We show you the manifest; the official verifier checks the signing chain.
- Anything at all, sometimes. A screenshot of a screenshot carries no information, and no tool can change that.
The one thing worth remembering
A clean file is not a real photo. Metadata is trivial to remove — taking a screenshot does it by accident. Absence of an AI marker is never evidence that an image is authentic. It only means this particular file has nothing left to say.
Markers are strong evidence in one direction and no evidence at all in the other. Any tool that treats silence as innocence is selling you certainty it does not have.
Why this suddenly matters
Article 50 of the EU AI Act took effect on 2 August 2026. Providers of generative AI must mark synthetic output so machines can detect it, with penalties up to EUR 15 million or 3% of worldwide annual turnover.
The markers this tool reads are moving from a voluntary courtesy to a legal obligation, and the share of AI images that declare themselves is climbing fast. Systems already on the market before that date have until 2 December 2026 to comply with the machine-readable marking requirement, and content generated before 2 August 2026 needs no retroactive labelling.
Stripping those markers to pass synthetic media off as real runs against the transparency regime, and against the terms of service of every major model provider. This tool reads markers. It does not remove them.
Questions
Is my image uploaded anywhere?
No. Every byte is parsed by JavaScript inside your own browser tab. There is no backend, no storage, no logging. Want to verify that yourself? Turn off your internet and use the tool anyway — it still works. That is the easiest proof there is.
Can you detect SynthID from Google Gemini?
No, and be sceptical of anyone claiming otherwise. SynthID hides a statistical watermark in the pixels themselves and Google has not released a public reader. If a Gemini image still carries its C2PA manifest or IPTC tag, we will find it. If those were stripped, the SynthID watermark may still be sitting in the pixels, invisible to us. Google runs its own SynthID Detector for that.
I checked a photo and got Inconclusive. Is it broken?
No — that is the honest answer for that file. Instagram, Facebook, WhatsApp, Telegram, Discord and X all re-encode what you upload and throw away the metadata. Screenshots lose everything too. When that happens we fall back to reading the file's structure, which catches a lot of generated images, but not all of them. Try the original file rather than a version saved off social media and you will usually get much further.
Can the markers be faked?
In both directions, and we say so on every result. EXIF is plain text that can be copied from a genuine camera file onto a generated one, which is why camera evidence is reported as supporting rather than conclusive. C2PA manifests are cryptographically signed, so forging one convincingly is far harder — but verifying the signing certificate needs the official verifier, which we link to on every report.
Which formats work?
JPEG, PNG, WebP, AVIF, HEIC/HEIF and TIFF. JPEG and PNG get the deepest treatment, including compressed zTXt and iTXt chunks and full quantization table analysis. AVIF and HEIC get container parsing plus a raw marker sweep.
What does it cost, and what is the catch?
Nothing, and there isn't one. The site is paid for by advertising. Since no file ever leaves your device there is no server cost per check, so there is no reason to meter you, gate features or ask for an email address.