midjourney vs real photos: a side-by-side i did so i would stop guessing

tech · 2024-09-27 · 6 min

i used midjourney for a month and kept guessing whether an output looked real. that is a bad way to work. so i did a side-by-side: my own photos on one side, generated images on the other, same subjects, similar framing, and a simple question for the people who looked at them. which one is real?

the test was not scientific. it was useful.

the test

ten subjects: a cup on a desk, a street corner, a portrait, a plant, a window, a meal, a pair of shoes, a bookshelf, a bicycle, and a wall with a poster. for each subject i took a photo with my phone, then asked midjourney for the closest match i could get with a text prompt. same subject, similar light, similar crop. twenty images total, shuffled.

five people looked at them and guessed. i did not tell them how many were generated. i did tell them the images were from the same set, because otherwise the question is unfair.

what surprised me

texture and light are not the tell anymore. the generated skin, the generated wood, the generated window light: all good enough to pass at a glance. the old tells, plastic skin and impossible shadows, are mostly gone. if your test is "does it look like a photo", the answer is often yes.

the tells moved to the edges. hands, text, thin structures, and anything with a specific place. a sign with letters, a bicycle wheel with the right number of spokes, a street that exists. the model is excellent at the average of a thing and weak at the particular thing.

consistency is the real gap. one image can pass. a series cannot. the same person across five images changes face, the same room changes layout, the same product changes label. if you need a set, the generated set is where the illusion breaks.

people guessed better than chance, but not by much. they were right about fourteen out of twenty. the misses were mostly generated images that were slightly too clean: too little noise, too perfect a composition. the real photos had the small mess that a camera catches and a prompt forgets.

where generated images win

mood and abstract. a background, a texture, a concept image, an illustration for a post. the model is fast, cheap, and does not need a release form. for a blog header, it is often better than a stock photo, because it can be the exact color and composition you need.

iteration. i can try ten directions in ten minutes. a photographer cannot. for early exploration, the generated image is a sketchbook, and a sketchbook is a legitimate tool.

the impossible. a product that does not exist yet, a scene that would cost a shoot, a visual metaphor. the model is not replacing photography there. it is doing a job photography was never able to do.

where real photos win

trust. a real photo of a real place, a real team, a real product is evidence. a generated image is decoration. for a business site, the difference is not aesthetic. it is the difference between "here is what we made" and "here is a picture of something like what we made".

specificity. the actual storefront, the actual dish, the actual view from the room. a generated image can be beautiful and still be wrong, and for a local business, wrong is expensive.

rights and disclosure. a real photo has a photographer and a license. a generated image has a model and a policy. both can be used, but the paperwork is different, and clients increasingly ask which one they are looking at.

what i do now

for a post header or a mood image, i generate. for anything that claims to show a real thing, i shoot. for a client, i say which one it is, before they ask. that last part is not legal advice. it is just the version of the conversation that does not end badly.

the side-by-side stopped the guessing. the rule that came out of it is simpler than the test: use the image that can tell the truth about the thing you are showing. if the truth is "a feeling", generate it. if the truth is "this place, this product, this person", photograph it.

for the browser side of the same shift, see the year of the ai browser. for the model side, see what is an llm.