Same prompt tests GPT Image 2.5 vs Nano Banana Pro on counting
A 24-person Victorian library prompt checks exact counting and layout: GPT rendered about 23 people and Nano Banana Pro about 20, and neither hit 24.
Ran the same prompt on GPT Image 2.5 and Gemini Nano Banana Pro The prompt: A massive Victorian library viewed from a high balcony. Exactly 24 people occupy the room: 6 reading at tables, 4 climbing ladders, 3 standing near the fireplace, 5 walking between bookshelves, 2 librarians behind the desk, 2 looking through a telescope, and 2 sitting on the staircase. Every person must have a unique pose, outfit, hairstyle, and position. Three spiral staircases connect different levels. Hundreds of bookshelves form symmetrical geometric patterns. Maintain accurate perspective and consistent architecture throughout the entire image. And this is what I got Headcount → GPT: ~23 people → Nano Banana Pro: ~20 people → Neither got 24 Reading at tables (asked for 6) → GPT: 5 ❌ → Pro: 5 ❌ Climbers on their own ladder (asked for 4) → GPT: 2 ❌ → Pro: ~3 ❌ Fireplace (asked for 3) → GPT: 3 ✅ → Pro: 1 ❌ Telescope (asked for 2) → GPT: 2 ✅ → Pro: 2 ✅ Staircase sitters (asked for 2) → GPT: 2 ✅ (one is sitting a little too dangerously 😭) → Pro: 1 ❌ Librarians (asked for 2) → GPT: 2 ✅ → Pro: 2 ✅ (plus an extra person nearby) Ngl, the detail here is wild. I started hallucinating too while counting people, and I wouldn't blame the models if they did too 💀 But Nano Banana 2.5 is genuinely impressive. Can't wait for the release to see what people make with it.
This post quotes another post: x.com/abhinavflac/status/2106241851804328202