Midjourney vs Real Photos: How to Tell Them Apart
Straight to the point: on the original export, Midjourney images carry a stable machine-readable trace; the "style" you can spot with your eyes is far less reliable. This page separates the two by how much you can actually trust them.
Lots of people search for "how to spot Midjourney" expecting a checklist of visual tells. The reliability of such lists decays fast with each model version. Since v6, Midjourney output quality is close enough to real photography that "it looks too perfect" is no longer a usable signal.
The useful split is between machine-readable evidence (metadata) and visual characteristics that still apply, but only in specific conditions.
Layer 1: conclusive metadata evidence
This is the most reliable marker of a Midjourney image. Exported PNGs and JPEGs carry a parameter string in their metadata, typically containing:
- Software: Midjourney — the generator identifier itself.
- A full parameter string — for example
--v 6 --ar 16:9 --style raw --s 250 --q 2, including version, aspect ratio, stylisation, seed. - The prompt — sometimes the full prompt text as well.
- A job ID — a unique identifier for the generation.
Read those in the original file and the conclusion is settled — no visual inspection required. That is exactly the layer this site checks.
But this layer is trivially erased
Screenshots, social platform transcoding, re-saving as JPEG, and any image compression tool all strip container metadata. If the image you received has already been processed by a platform, this evidence does not exist and you are down to the weaker layer.
Layer 2: Midjourney's visual tendencies
The characteristics below still exist in current versions, but they are only useful under specific conditions and should never be used alone.
Tendency 1: the default aesthetic is too "correct"
Midjourney's default taste leans cinematic: a centred subject, shallow depth of field, warm highlights against cool shadows, and a subtle vignette. The frame almost always has a clear subject, a defined focal point and deliberately designed tonal separation. Real photography has such images too — just in different proportions.
The point is the tendency, not a feature: computational photography is converging on the same look, and the gap keeps narrowing.
Tendency 2: lighting is "too coherent"
Generative models tend to produce dramatic yet internally consistent lighting: clean rim light, soft shadow edges, rounded highlights that never blow out. Real light is messier — mixed sources, reflected spill, lens flare, slight exposure unevenness. Every light source being exactly plausible can itself be the tell.
Tendency 3: texture is over-smoothed
Large areas — skin, fabric, walls — often come out uniformly grain-free, while real photographs retain sensor noise or lens texture even after beauty retouching.
Tendency 4: secondary detail is suspiciously plausible
Background signage, book spines, bystanders' clothing — elements that need not be legible — often come out roughly the right shape and entirely fake in detail. Zoom into secondary regions rather than the subject. The subject is what the model is good at; the background is where it struggles.
Tendency 5: compositional habits
Default Midjourney output favours central composition, symmetry and a single clear focal point. Real photography is looser and more random. But with --style raw or image guidance, this tendency weakens noticeably.
Tendency 6: Chinese characters and complex fonts
Short English text is now largely fine, but Chinese, Japanese, Arabic and decorative fonts still fail often enough to be useful — strokes fusing, components misplaced, meaninglessness. When such text is present, this cue still works well.
Versions and parameters change things a lot
Treating "a Midjourney image" as one category is imprecise. These variables all change the picture significantly:
| Variable | Effect on detection |
|---|---|
| Version (v5 / v6 / v7) | Newer is harder to judge by eye, but the version string in metadata stays explicit |
| --style raw | Substantially reduces the cinematic tendency; visual tells become faint |
| --s (stylisation) | Lower values produce more photorealistic output and more false positives |
| --stylize / --chaos | High chaos pushes composition off-convention, weakening visual tells |
| Reference guidance (--sref / --cref) | May blend real source material; metadata remains decisive |
| Whether it was re-processed | Determines whether metadata evidence still exists |
The version number and parameter string — both in metadata — are worth more than any visual guess.
A quick-reference table
Ranked by how much you can trust it:
- Metadata contains a Midjourney identifier and parameter string — conclusive.
- Metadata contains any AI generator signature — conclusive; it need not be Midjourney specifically.
- Metadata contains a complete camera EXIF — strong evidence for a real photograph.
- Background detail errors, garbled text, mismatched reflections — inference; read it in context.
- "It looks too perfect / too cinematic" — a very weak subjective cue. Do not rely on it.
If the image has already been processed
Then the honest answer is "cannot tell", not a score. Without the original, metadata evidence does not exist and visual characteristics offer only weak signals — any high-confidence conclusion at that point is unreliable.
To confirm, try asking for the original file. In social media contexts, a reverse image search helps too — Midjourney output often appears on public galleries before it appears in the wild.
Frequently asked questions
Can you tell a Midjourney image by eye?
It is hard, and it gets harder with each version. Since v6 the output quality is very close to real photography. The reliable route is checking metadata for the Midjourney identifier and parameter string — conclusive on the original file. If it has been screenshotted or transcoded, you are left with weaker visual cues.
What metadata does Midjourney leave behind?
Usually a Software identifier, a full parameter string (version, aspect ratio, --style raw flag, stylisation, seed), sometimes the complete prompt text and a job ID. It sits in the PNG tEXt chunk or the JPEG XMP/EXIF block, and it disappears with a screenshot or a re-encode.
Can Midjourney v6 images still be detected?
On an unprocessed original export, yes — the version and software identifiers in the metadata are definitive. Once the image has been compressed or transcoded by a platform, the metadata disappears and you are left with pixel analysis and visual cues, with noticeably lower accuracy.
What if it was generated with --style raw?
Visually harder, because that flag markedly reduces the default cinematic tendency. But the metadata parameter string still records --style raw, which tells you it is Midjourney output — and that is still conclusive evidence.
Does Midjourney add a watermark?
Midjourney writes generation identification into the exported file's metadata as a parameter string, and offers a "no signature" option that does not remove the generation information from the file metadata. Note that this differs from pixel-embedded watermarks like Google SynthID: metadata lives in the file container and dies in a screenshot, whereas a pixel watermark is copied along with the pixels.
Related articles
- Why AI Image Detectors Fail (And What to Do)
- How to Tell If an Image Is AI-Generated: 7 Methods
- AI Image Detection: Metadata vs C2PA vs Watermarks
- AI Image Metadata: AIGC Labels, Signatures and EXIF
Check an image yourself
Upload an image and check its metadata for a Midjourney or other generator signature
Open the detector