Midjourney Detection Benchmark

Why Midjourney gets its own benchmark

Midjourney's images are produced by a hosted, versioned service rather than a locally-run open model, and its outputs have a recognizable stylistic tendency that has evolved across versions. Folding Midjourney into a single blended AI-image detection benchmark would average its detection difficulty together with structurally different generators (for example, open-source diffusion models with far more variable post-processing), obscuring whether a detector's accuracy is stable across Midjourney's own version history specifically.

What this benchmark scopes to test

The evaluation is scoped to Midjourney outputs specifically, across the version range still in common circulation, and to the common ways those images reach an end user — direct export, common upscaling, and social-platform re-upload.

Relationship to the general AI-image detection benchmark

This benchmark shares the same evaluation protocol conventions (test set size, scoring threshold, tie-handling, reproducibility disclosure) as PhotoProof AI's general AI-image detection benchmark, so results are comparable in method even though the test sets are disjoint. See the general benchmark for the multi-generator baseline this one is scoped narrower than.

Does this replace the general AI-image detection benchmark?

No. It complements it. The general benchmark measures cross-generator performance; this one isolates Midjourney specifically, since a multi-generator average can hide generator-specific weaknesses.

Will this benchmark be updated as new Midjourney versions release?

That is the intent — a generator-specific benchmark that isn't revisited as its target model changes stops being representative. See the revision history on this page for what has actually been updated so far, rather than assuming it is current.