AI Photo Detector
Drop a PNG, JPEG, or WebP into this AI Photo Detector to inspect generator metadata, camera EXIF, C2PA credentials, and pixel-level clues. You get a 0–100 probability plus a readable evidence log—not a black-box guess.
The workbench on this URL is the product: files are parsed locally unless you opt into a paid vision check. Use it when a listing, news image, or profile picture looks too clean and you need traces you can actually cite.
Drag & drop a photo here
or browse files — PNG · JPEG · WebP · max 20 MB
Try a real photo
Try an AI image
Core forensic pipeline runs in your browser — your photo never touches our servers unless you opt into the cloud check.
No specimen analyzed yet. The evidence report renders here after your first scan.
How to use this AI Photo Detector
Drop in the photo
Drag any PNG, JPEG, or WebP (up to 20 MB) onto the intake panel, or click to browse. The file is read with the browser File API. Nothing is uploaded for the default scan. Try a sample camera still if you want to see a report first.
Hit Analyze and scan
Choose Analyze Image. Four channels run in sequence: generator text in PNG chunks, camera EXIF, C2PA/JUMBF blocks, and canvas pixel statistics. Tick the optional cloud check only if you want a vision model on top—that path spends credits and sends the file to a server.
Read the report card
You get a 0–100 probability, a plain-English verdict, and a weighted evidence log. Download a PNG card for a client or editor. Treat mixed scores as “keep looking,” not a courtroom stamp.
What this detector actually inspects
Generator metadata analysis
Many diffusion workflows write parameters, prompts, or software names into PNG tEXt/iTXt chunks. When those strings match known pipelines (Stable Diffusion, ComfyUI, Flux, and others), the finding is weighted heavily because it is documentary, not a vibe check.
Camera EXIF cross-check
JPEG stills from phones and DSLRs usually carry Make, Model, Software, and timestamps. A coherent camera fingerprint leans authentic. A wiped IFD, or Software that names an image generator, leans synthetic. Partial EXIF is logged as moderate—not as proof either way.
C2PA credential scan
Content Credentials (C2PA / JUMBF) can record capture, generation, or edits when the producer signed the file. Presence of a block is a provenance clue, not automatic guilt: cameras, editors, and generators can all emit credentials. The log tells you a credential was found so you can inspect the claim.
Pixel-statistics heuristics
A canvas pass looks at noise uniformity, micro-contrast, chroma in smooth regions, and common generator canvas sizes. These signals are labeled heuristic and weighted lowest. Heavy denoise, social recompression, and HDR polish can mimic “too clean” pixels on a real capture.
Private by architecture
Metadata parsing, C2PA scanning, and pixel math run in your browser. The photo never touches Pixwit servers unless you enable the cloud check. Keep embargoed newsroom files and personal IDs off that optional path.
Shareable report card
Every completed scan can export a PNG: gauge, verdict, and the weighted log. Use it as a paper trail, not as a substitute for talking to the photographer. If metadata was stripped, the card will say so instead of inventing certainty.
How to read forensic signals without overclaiming
People search for an AI Photo Detector when a file already looks suspicious: a product shot with no sensor grain, a headshot that never existed on a camera roll, a meme that jumped from a generator Discord into a marketplace. This page is that check—four channels, a probability, and a log you can quote. It is not a lab that recovers deleted originals or a model that “understands” the scene. Generators improve, exporters strip chunks, and editors rewrite EXIF. Start with what the bytes still contain.
Why metadata still beats a vibe check
If a PNG still holds a prompt, sampler, or UI signature, you are reading a note the software left behind. That is stronger than “the hands look weird.” JPEG EXIF from a named camera body plus a plausible timestamp is equally useful in the other direction. The common failure mode is a re-saved social download: Instagram, WhatsApp, and many CDNs strip both generator chunks and camera IFDs. A blank metadata panel then means “the trail was cleaned or never written,” not “this must be real.”
What C2PA can and cannot prove
The Coalition for Content Provenance and Authenticity publishes an open standard so capture devices and creative tools can attach signed history. A credential might say a phone took the frame, or that a generator produced it, or that Photoshop edited it. Absence of C2PA is normal for older cameras and casual exports. Presence is a lead: open the claim, do not treat the mere bytes as a verdict. For the spec’s own overview, see the background reading below.
When pixel heuristics help—and when they lie
Diffusion outputs often show unusually even noise, flattened chroma in skies or walls, and canvas sizes that match training defaults (square 1024, 768, and similar). Those patterns can also appear after aggressive NR, AI denoise in a camera app, or a heavy Instagram filter on a genuine capture. That is why this scanner marks pixel stats as heuristic and keeps their weight below metadata. If the only red flag is “too smooth,” ask for the original file from the device.
Formats, size limits, and what to send
PNG, JPEG, and WebP up to 20 MB are supported. PNG is richest for generator parameters; JPEG is richest for camera EXIF. WebP still gets C2PA and pixel passes. Screenshots of screenshots and chat-bubble crops have the weakest trail—ask for the camera or generator export when you can.
Honesty about accuracy and abuse
No consumer scanner is 100% accurate. Metadata can be forged or stripped; heuristics misfire on film scans, night shots, and beauty-mode portraits. Use the score beside reverse-image search and the original file. This is a first-pass aid, not a biometric ID check, a medical test, or a litigation lab.
Optional cloud check versus the local pipeline
The default path never leaves the device. The optional cloud check spends credits and sends the image to a vision model that may catch stylistic tells the local channels miss. Use it when metadata is empty—not as a substitute for reading the log.
Background reading: C2PA (Wikipedia)
Who actually needs an AI photo check
Marketplaces and second-hand listings
Run a listing image before you pay a deposit. Generator chunks or missing camera EXIF plus stock-smooth pixels is a reason to ask for a handheld shot with a handwritten note—not an automatic ban by itself.
Newsrooms, fact-checks, and community notes
A local scan that finds Firefly or Midjourney software tags is citable. A blank result is also citable: say the file was inconclusive, then keep geolocation and eyewitness work in the story. Download the card for the editor’s file.
Hiring, profiles, and creator portfolios
Recruiters and brand partners can ask talent for a camera original when a portrait looks airbrushed beyond a retoucher’s usual pass. The report is a conversation starter, not a reason to reject someone in silence.
Design QA before you publish a campaign
Teams mixing stock, shoots, and generated comps need a quick gate so a synthetic extra does not ship labeled as photography. Drop each hero still through the same three steps, keep the cards with the DAM entry, and only escalate to legal when the brief promised real talent or a real location.
Frequently asked questions
Related tools
Need motion after you vet the still?
Once a frame is cleared or flagged, continue in Pixwit with prompt- or image-based clips for product and campaign footage.