I expected this comparison to be uglier. Meta Muse, Gemini Nano Banana 2, and ChatGPT Images 2.0 sounded like a perfect setup for plastic faces, mangled hands, fake products, and posters written in haunted alphabet soup. Instead, they were mostly competent, which somehow made the whole thing more suspicious.
These aren’t identical tools wearing different logos. Meta pitches Muse Image as a social image model living inside Meta AI and its apps. Google frames Nano Banana 2 around speed, editing, and Gemini’s broader knowledge. OpenAI sells ChatGPT Images 2.0 on text rendering, visual control, and stronger prompt handling. Different ambitions, same polished little showroom.
The uncanny valley got quieter
The first prompt asked for a tired office worker eating instant noodles at midnight. The detailed version added a messy kitchen, a laptop, dishes in the sink, and harsh refrigerator light. Meta and Gemini generated quickly. ChatGPT gave me the most distinct simple image, then mostly repeated that mood when I added more detail.
That wasn’t a failure, exactly. It obeyed. The problem was that all three seemed more interested in quality than realism. The rooms were composed. The lighting was handsome. The exhaustion looked art-directed by someone who’s never eaten noodles over the sink at midnight.
Readable text became the easy win
The poster test should have been a mess. I asked for a fake coffee shop poster for “Bad Wi-Fi Café,” then gave the detailed version exact text to include. All three made decent posters, and the text was readable enough to count as progress. AI image tools used to treat letters like cursed decoration, so I’ll give them that.
The finish still had the usual stain. Everything carried that warm AI yellow tint, as if every café, kitchen, and bedroom had been lit by a sponsored sunset. Gemini made what looked like a photo of a poster rather than the poster itself. ChatGPT had the roughest showing, taking more than three minutes, failing three times, and only working after I started a new chat.
The cat, suitcase, and umbrella prompt was cleaner. The models followed placement instructions well. Product imagery was shakier. A sleek earbud photo was easy, but “open-ear wireless earbuds” pushed all three into generic premium earbud mush.
Manila looked familiar through a filter
The Manila street food prompt could’ve collapsed into tourist-board soup. Meta and Gemini did better in the runs I completed. They included plastic stools, wet pavement, motorcycles, steam, tarps, and the cramped casualness of eating outside while the city keeps moving.
That made the weakness easier to see. The ingredients were right, but the finish still felt too smooth, like the street had been cleaned for a brochure shoot five minutes before the rain started.
None of these tools collapsed in the old funny ways. The scenes were coherent. The text was readable. The objects usually went where I asked them to go. The failures have moved from competence to taste.
That leaves a weirder question than which model “wins.” In the post-truth era, does it matter which tool is the most realistic, or which one follows instructions best, when all three can already make a fake thing look plausible enough to survive a quick scroll?
These models have learned how to make images look expensive before they’ve learned how to make them feel lived in. The fake premium era is less funny than nightmare fingers, but it’s probably more useful, which is exactly what makes it harder to ignore.