Anyone else noticing detection accuracy just falling off a cliff since midjourney v7 went wide??
I run a stock photo review service (small operation, just me and one other person). We’ve been using Hive + manual review as our detection pipeline and it was catching ~90% of AI submissions through january. since v7 dropped we’re probably missing 30-40% based on spot checks we’ve been doing.
the images are just… better. the skin texture thing is mostly gone. hands are usually fine now. that weird “AI sheen” on surfaces is way less noticeable.
we’re looking at requiring C2PA verification for all submissions now. if you cant prove provenance it doesnt get accepted. feels harsh but i honestly dont see another option.
what are other platforms doing? anyone managing submissions at scale want to share how you’re handling this?
I photograph weddings for a living and honestly even I struggle to tell v7 from real photos at first glance now. The quality jump from v6 to v7 was massive.
Requiring C2PA seems like the right call for stock. Bit harsh for individual photographers who havent set it up yet though — maybe a grace period + a guide on how to start signing their work?
We’re a larger stock platform (cant say which, sorry) and yeah we’re seeing the same thing. Our detection pipeline is multi-layered — Hive + internal model + manual review for flagged items — and our false negative rate has roughly doubled since v7.
We haven’t gone full C2PA-required yet but it’s being discussed seriously in leadership. The concern is we’d lose 60%+ of our contributor base overnight since most photographers don’t use it yet.
For now we’ve added a self-certification checkbox and are investing heavily in updating our internal models. But it feels like playing whack-a-mole honestly.
this is kind of inevitable tho right? the detection vs generation arms race was always going to favor generation eventually.
the real solution isnt better detection its better provenance. prove what IS real instead of trying to catch what isnt. C2PA, process documentation, camera-level signing. flip the burden.
easier said than done i know but detecton-first is a losing strategy long term imo
Some data points from my own testing (I run detection benchmarks quarterly):
MJ v6: Hive detected 87%, Illuminarty 82%, Sensity 79%
MJ v7: Hive detected 61%, Illuminarty 54%, Sensity 58%
That’s a ~25 percentage point drop across the board. Pretty alarming.
The biggest improvement in v7 seems to be texture consistency and lighting. Earlier versions had this “everything is slightly too smooth” look. V7 introduces realistic imperfections — sensor noise, slight chromatic aberration, natural vignetting. It’s actively mimicking camera physics.
@henry.nomad yeah a grace period makes sense. gonna draft a contributor guide for C2PA setup this week.
@w.moses.flint.54 appreciate you sharing that, even anonymously. the self-certification approach is interesting — at least creates a paper trail even if people lie.
@sleek.Protocol those numbers are really helpful, thanks. closely matches what we’re seeing. might start doing quarterly benchmarks ourselves.
gonna share the C2PA guide here when its ready if anyone wants to adapt it for their own platform.