AI image detector
Adobe Firefly & Generative Fill detector
Adobe attaches Content Credentials to everything Firefly generates, including Generative Fill and Generative Expand edits in Photoshop. GPTTrace reads them and tells you whether a picture was generated outright or only partly edited.
Free · no sign-up · your file never leaves your device
Drop a suspected Firefly image, click to choose, or paste
JPEG · PNG · WebP · AVIF · HEIC (Safari) — checked on your device
The most consistently labelled AI images
Adobe has pushed provenance harder than any other company. It co-founded the Content Authenticity Initiative in 2019 and the C2PA standards body in 2021, and it attaches Content Credentials to Firefly output across its products: the Firefly web app, Photoshop’s Generative Fill and Generative Expand, Illustrator’s generative features and Adobe Express. The credential is a signed manifest stored in the file, naming the tool and recording that generative AI was used.
That makes Firefly images the easiest AI images to identify — when the credential survives. GPTTrace reads the manifest from JPEG, PNG and WebP files, shows the claim generator and the signer, and lists every recorded action. The C2PA viewer shows the full history, including earlier manifests for the photo that was edited.
Generated versus edited
A distinctive thing about Adobe’s tools is how often they are used for partial edits. A photographer removes a tourist from a landscape, extends the sky to fit a banner, or replaces a background. Those images are mostly genuine photographs with an AI-made region. Content Credentials capture the difference using IPTC source types: trainedAlgorithmicMedia for wholly generated images and compositeWithTrainedAlgorithmicMedia for real images with AI parts. GPTTrace reports the two separately so you don’t mistake a touched-up photo for an invented one.
Without a credential, a partly edited photo is hard for any pixel classifier, because most of the pixels are real. If you suspect a specific area — a removed person, an added object — crop to it and check the crop.
When there is no credential
Firefly output without a manifest has usually passed through a screenshot, a social network or an export that dropped metadata. Visually, Firefly images tend toward clean stock-photo aesthetics: bright even lighting, tidy compositions and polished surfaces, which suits its common use for marketing material. The neural classifier in GPTTrace handles stripped Firefly images like any other generator, with the usual drop in confidence for small or heavily compressed copies.
If you publish Firefly work and want it recognised as such, keep credentials enabled on export and host the original files where possible.
Checking a Firefly image step by step
Drop the original file you received — ideally the one exported from Photoshop or downloaded from the Firefly site. If the result says “Content Credentials: created by AI” and names Adobe as the signer, you have a definitive answer. If it says “edited with generative AI”, open the C2PA viewer to see which steps were generative: the ingredient list shows the original photo and the edits applied on top. If there is no credential, look at the classifier score and the generator-size hint; Firefly’s default outputs are common aspect ratios such as square, 4:3 and 16:9 at around two megapixels. A stock-photo-perfect image with no camera metadata and a high classifier score is worth asking about, especially in advertising or news contexts where AI use should be disclosed.
What Adobe Firefly leaves in a file
These are the traces GPTTrace checks for. “Some files only” means the trace is often missing — a re-save, screenshot or social-media upload removes metadata — so its absence proves nothing.
| Generator | Where | What to look for | How reliable |
|---|---|---|---|
| Adobe Firefly | C2PA manifest | Content Credentials signed by Adobe, naming Firefly or Photoshop Generative Fill as the generator | Documented by the vendor |
| Adobe Firefly | XMP | xmpMM history and photoshop:Credit entries when exported through Photoshop | Some files only |
How accurate is it? Our measured numbers
We test GPTTrace on labelled image samples and publish the results, including where it does badly. It is tuned to keep false accusations rare, so it misses some AI content rather than flag real work.
image check: AUC 0.938 (cross-validated)
679 labelled samples (399 AI, 280 human), run 2026-10-08. At the “Likely AI” line it caught 64% of AI samples and wrongly flagged 5% of human ones.
| Source | Truth | Samples | Result at “Likely AI” |
|---|---|---|---|
| gemini-nano-banana | AI | 40 | 40% caught |
| midjourney-v6 | AI | 40 | 68% caught |
| midjourney-v5 | AI | 40 | 73% caught |
| flux-dev | AI | 40 | 13% caught |
| flux-schnell | AI | 40 | 48% caught |
| sdxl | AI | 40 | 100% caught |
| gpt-image | AI | 40 | 30% caught |
| kling | AI | 39 | 97% caught |
| leonardo-stablecog | AI | 40 | 98% caught |
| bitmind-imagine-mix | AI | 40 | 80% caught |
| fullsize-photos | Human | 40 | 10% wrongly flagged |
| open-images-photos | Human | 40 | 5% wrongly flagged |
| lfw-faces | Human | 40 | 0% wrongly flagged |
| caltech-objects | Human | 40 | 3% wrongly flagged |
| coco-photos | Human | 40 | 0% wrongly flagged |
| ffhq-faces | Human | 40 | 0% wrongly flagged |
| celeba-faces | Human | 40 | 18% wrongly flagged |
Data sources and method: methodology & accuracy.