Flux image
detector

Flux is an image generator developed by Black Forest Labs, founded by researchers who previously worked at Stability AI (creators of Stable Diffusion). Unlike traditional diffusion models, Flux uses a flow-matching architecture that produces images with exceptional photorealistic quality.

Flux 1.1 Pro is considered by many experts as the generator that best mimics documentary and journalistic photography. Its output lacks the classic visual tells — it does not oversaturate like Midjourney, does not have the "illustrative" look of DALL·E, and produces skin, vegetation and architecture textures nearly indistinguishable from a real camera.

This makes Flux the most dangerous model for disinformation targeting media organizations, fact-checkers, and newsrooms that rely on quick visual assessment.

Although Flux is the hardest model to detect visually, its flow-matching architecture leaves its own statistical fingerprints:

1. Flow-matching DCT signature. Flow-matching generation produces a frequency coefficient distribution different from both traditional diffusion and real camera sensors. ScanTrace has trained its spectral model on thousands of labeled Flux images.

2. Residual noise pattern. Flux produces residual noise with higher spatial autocorrelation than real photography. This signal is weak but consistent — and is the key to detecting Flux 1.1 Pro, which eliminates almost all visual artifacts.

3. EXIF and metadata inspection. Flux images lack camera metadata. Combined with spectral signals, this absence contributes to the forensic verdict.

4. Contextual reasoning. A language model interprets numeric data from the three previous layers and generates a readable explanation of the verdict.

ScanTrace detection rates as of April 2026:

Stable Diffusion XL: 97% — pronounced artifacts, easiest to detect.

DALL·E 3: 95% — distinctive DCT patterns, strong text-area signal.

Midjourney v7: 93% — consistent spectral signature despite extreme realism.

Flux 1.1 Pro: 91% — the hardest. Weaker signal, minimum 768x768 recommended. Flux Schnell and Dev have rates of 94% and 92% respectively.

Flux's flow-matching architecture represents the next frontier in AI image detection — and the reason forensic tools must continuously evolve.

Although Flux 1.1 Pro has eliminated most visual artifacts, these signals may occasionally appear: reflections on metallic or glass surfaces at inconsistent angles, depth-of-field transitions that do not follow real lens optics, partially readable background text with invented characters (less frequent than in Midjourney), and symmetries in objects that should be asymmetric (leaves, stones, clouds).

However, relying solely on visual cues to detect Flux is insufficient. Automated forensic analysis is the only reliable path.

Suspect Flux when receiving an image with documentary photography aesthetics but no verifiable source: natural-looking portraits without apparent editing, urban or conflict scenes with perfect lighting, hyper-realistic landscapes, and complete absence of EXIF metadata.

ScanTrace analyzes the image in under 15 seconds — compatible with any editorial or verification deadline.

What is Flux and why is it dangerous for disinformation?

Flux is an image generator by Black Forest Labs that uses a flow-matching architecture instead of conventional diffusion. Its Flux 1.1 Pro variant produces the closest aesthetic to real documentary photography, making it the most dangerous tool for creating visual disinformation targeted at journalists and newsrooms.

Can ScanTrace distinguish between Flux Pro, Dev and Schnell?

ScanTrace identifies the Flux family signature. Flux Schnell produces lower-quality images with more obvious artifacts. Flux Pro and Dev share a similar spectral fingerprint with subtle differences in detail level that ScanTrace factors into the analysis.

Does Flux include EXIF metadata?

Images generated with Flux through the Black Forest Labs API or services like Replicate typically lack camera EXIF metadata. This absence is one signal ScanTrace combines with spectral analysis to produce the verdict.

Does it work on Flux images uploaded to Instagram or Twitter?

Yes, with limitations. Social media platforms recompress and strip metadata. If the original had sufficient resolution (1024x1024+), the Flux spectral signature usually survives compression. At very low resolutions the verdict may be INDETERMINATE.

Is the Flux detector free?

Yes. 10 free analyses per month, no credit card. Each analysis returns a verdict, 0–1 confidence score and forensic explanation of the signals detected.