Midjourney image
detector

Midjourney, particularly since version v6, produces images with a level of photographic realism that surpasses most generators. According to a University of Waterloo study, humans correctly identify Midjourney v5+ images as synthetic only 38% of the time — worse than flipping a coin.

What makes Midjourney especially dangerous for disinformation is its ability to generate scenes that look like real photojournalism: natural lighting, off-center compositions, organic textures, and a color palette that avoids the artificial oversaturation of earlier generators.

Midjourney v7, released in early 2025, pushed photorealism further — making automated forensic analysis the only reliable detection path.

ScanTrace runs three independent forensic layers in parallel:

1. Pixel-level spectral analysis. Midjourney images show characteristic anomalies in high-frequency DCT coefficients. While a real camera sensor produces random thermal noise (PRNU), Midjourney generates smooth, spatially correlated noise patterns — invisible to the human eye but statistically detectable.

2. EXIF metadata inspection. Images downloaded directly from Discord (Midjourney's official channel) lack camera metadata. Some include the tag "Software: Midjourney" in EXIF fields. The total absence of lens data, ISO, and GPS is a signal ScanTrace weighs alongside spectral analysis.

3. Contextual AI reasoning. A language model receives numeric data from layers 1 and 2 — never the raw image — and generates a human-readable explanation of why the image is likely synthetic, including the specific signals detected.

Midjourney v4: visible artifacts in hands and eyes, repetitive background textures, unnaturally soft edges. Relatively easy to detect both visually and forensically.

Midjourney v5: significant leap in realism. The chromatic oversaturation ("Midjourney look") is the most recognizable visual signal. Forensically: distinctive DCT patterns and very uniform residual noise.

Midjourney v6: major improvements in text rendering and hands. Reduces oversaturation but introduces a particular skin texture ("porcelain skin") detectable spectrally. Images in --style raw mode are the most challenging to detect.

Midjourney v7: the most advanced model as of April 2026. Nearly indistinguishable photography. Detection relies almost exclusively on statistical analysis — visual cues are minimal. ScanTrace maintains a 93% detection rate on v7 thanks to persistent frequency-domain anomalies.

In March 2023, a Midjourney v5 image of Pope Francis wearing a white Balenciaga coat was shared more than 4 million times before being debunked. In 2024, fake images of Trump being arrested, generated with Midjourney, were published by international outlets as real during the first hours of circulation.

The European Broadcasting Union reported in 2025 that 77% of European newsrooms have inadvertently published at least one AI-generated image. Midjourney is the most frequently cited tool in these incidents due to its photorealistic output quality.

For journalists and fact-checkers, running every unverified image through a forensic detector is no longer optional — it is part of the editorial workflow.

Check any image from an unverified source that displays: exceptionally high quality without sensor noise, portraits with unnaturally smooth skin, scenes that look like editorial photography but circulate without credit, and breaking-news material from social media.

ScanTrace delivers the verdict in under 15 seconds — fast enough for any editorial deadline.

Can ScanTrace distinguish between different Midjourney versions?

Yes. Midjourney v4, v5, v6 and v7 leave slightly different spectral fingerprints in the frequency domain. ScanTrace identifies patterns common to all versions plus version-specific nuances — the characteristic oversaturation of v5, the porcelain skin texture of v6, and the photographic hyper-realism of v7.

Does it work on Midjourney images that have been edited or inpainted?

Yes, as long as the image retains enough original pixels. Partial inpainting (less than 50% of the image area) usually preserves the detectable forensic signature. Extensive edits or full regenerations may lead to an INDETERMINATE verdict.

Can you detect a Midjourney upscale?

Yes. Midjourney's upscaling process adds its own statistical signature on top of the generative base. In fact, upscaled images are often easier to detect because they contain more pixels carrying the model's fingerprint.

What if the Midjourney image was compressed by WhatsApp or social media?

Compression degrades forensic signals. On high-resolution originals (1024x1024 or above), the spectral signature typically survives moderate JPEG compression. With extreme compression (below 200 KB) the verdict may shift to INDETERMINATE.

Is the Midjourney detector free?

Yes. The free plan includes 10 analyses per month with no credit card. Each analysis returns a verdict, 0–1 confidence score, and forensic explanation.