I use Gemini or ChatGPT to generate images for my posts. And they are getting really good at this, especially Gemini’s Nano Banana 2.
I used Nano Banana 2 to generate two images. With this prompt:
A 10-year old kid playing a conversational game with a ZX Spectrum. The monitor is an old TV. The image is photorealistic, but with aesthetics from the 80s, and a Polaroid-like look. Make the proportions 16-9.
The first one was this:
The mix of right and wrong details is amusing. The general aesthetic is very good, the key elements (ZX Spectrum, game on old TV, kid) are quite good (the Spectrum is too big). The game text is legible and consistent with conversational games.
On the other hand, too many posters and books mean lots of garbled text or deformed faces.
This is the second one:
This one is even better, in part because most of the text is out of focus. Some text in the foreground is garbled, but some out-of-focus text looks good (the cassette on the TV seems to read “The Hobbit”, and the imperial trooper in the background is in front of something that reads like Star Wars). The Spectrum has better proportions, and everything looks consistent with the date helpfully included in the Polaroid border.
My favourite detail is the game on the TV: it is actually The Hobbit, probably the best known conversational game for the Spectrum, but with graphics that look more like those of an NES.
As a check, I used the same prompt with ChatGPT Images:
While the worst image quality is consistent with an 80s Polaroid, everything in the image is worse: the Spectrum-like computer, the generic non-game-like text in the TV, the Rubik cube with a wrong perspective…
I’ll stick with Nano Banana 2 for the future.