April 26, 2026
Saturday project: a daily newspaper for Gradient Bang, generated with gpt-image-2. Notes and code in the thread.
I borrowed the idea, here, from a newspaper "digest" posted to a WhatsApp group I'm in. I wanted to experiment with using gpt-5.5 to build something a bit different from the stuff I'd been working on, and I wanted to try out gpt-image-2.
So I had Codex write a little script to pull a day's worth of events from the game database (~1M records) and look for interesting "story lines". The initial one-shot of this was pretty good!
Then I suggested some visual references (an example newspaper front page, a game screen shot, the ship images from the game front end) and asked Codex to write a script to generate a prompt for gpt-image-2. The initial one-shot of that was pretty good, too.
I worked a little bit on the prompting for both story-line extraction and image generation. Codex and I settled on two editorial voices: straight news and a gossip column. We refined the story topics a bit, fixed some corner cases, and tweaked a few other things. @jonptaylor suggested making the character and ship names cyan, which is a nice touch.
The main issue was that the text fidelity was good (especially compared to previous image models I've experimented with), but not perfect.
Here's an example of one-shot generation. Really good, but lots of small text rendering issues.
So I kept poking at this while doing other things.
Codex built a few versions with gpt-image-2 doing a graphical layout with "holes" for programmatic text rendering. Then I asked Codex to build a version that generated all of the individual story boxes, plus the masthead and footer, as separate gpt-image-2 renders, then just stitched them all together. That required standardizing on a layout. Which was okay, we had a layout I liked.
The result is near-perfect (as far as I can tell) text fidelity, excellent visual consistency, and nice visual balance in each story box.
As always, there's more stuff I'd like to do here. But this is pretty good. I'm really, really impressed with gpt-image-2.
Code is here: https://t.co/KYOh9g0IUC
