Midjourney detector

How Midjourney images give themselves away

Midjourney has a recognisable house style: cinematic lighting, rich colour grading, flawless skin and a painterly smoothness even in “photographic” prompts. People learn to spot it, but versions 6, 6.1 and 7 have closed most of the obvious gaps — hands have five fingers, text is often legible and skin has pores. Looking for weird fingers no longer works as a test.

What still works is checking the file and the statistics of the pixels, which is what GPTTrace does.

Metadata on original downloads

Images saved straight from Midjourney’s website commonly carry the prompt text and a job identifier in their description metadata. GPTTrace reads PNG text chunks, XMP and EXIF fields and flags any that name Midjourney. If you see a long prompt in the PNG chunks section of your result, that is the strongest evidence you can get without Midjourney’s own records.

Telltale dimensions

Midjourney renders at fixed sizes for each aspect ratio — 1:1 grids, 2:3 portraits, 16:9 landscapes — and its upscaler doubles them. Sizes like 1024×1024, 1456×816 and 2048×2048 are common, while phone photos come out at 4032×3024 or similar. GPTTrace notes when an image has a typical generator resolution, but treats it as a hint: anyone can crop.

Pixel statistics

Diffusion models build images through a decoder that leaves faint, regular structure in the frequency spectrum, and they render noise and texture differently from a camera sensor. The neural classifier picks up these patterns even when the picture looks photographic. You can see part of what it sees in the Spectrum tab, where a generated image often shows symmetric bright points away from the centre.

Checking a viral “photo” that might be Midjourney

The images that cause trouble are the ones made to look like news: a celebrity at an event that never happened, a disaster scene, a politician in an unlikely place. Midjourney was behind several of the best-known examples. When one lands in your feed:

• Find the earliest copy. A reverse image search often leads to a Midjourney showcase, a Reddit post with the prompt, or the artist’s account.
• Run it here anyway. Even a recompressed copy usually keeps enough pixel signal for the classifier, and the result lists each reason.
• Look for consistency. Real events are photographed by many people from many angles. A single striking image with no other coverage deserves suspicion whatever any detector says.

If you are an artist who uses Midjourney and wants to be transparent, keep the original files: their metadata is the easiest way to prove what you made and when.