atareh

AI Workflow · APR 21, 2026

ChatGPT Image 2.0 Generations + Prompts

OpenAI just shipped a new native image model. I ran 12 prompts through it — magazine covers, gag manga, Hindi recipe cards, 360° Apollo 11 panoramas, Japanese posters. Every prompt verbatim, every image back. Scroll the slideshow or steal the prompts.

atareh
@atareh
APR 21, 2026 · 6 min read
AIImage GenerationPrompts

OpenAI just shipped ChatGPT Images 2.0 — a new native image model inside ChatGPT. It's better at long text, multilingual typography, reference-photo fidelity, and ultra-wide / ultra-tall aspect ratios than anything before it.

I ran 12 prompts through it — some with reference photos, some pure text-to-image — to see where it actually lands. The slideshow below is everything, with a copy-prompt button on every slide so you can steal whatever's useful.

Flip through the slideshow

Use ← / → or the on-screen arrows. Tap Copy prompt (top-right) to grab the exact text I sent. On mobile, swipe.

← / → · swipe on mobile · tap Copy Prompt to grab anyOpen in new tab ↗

What the model is good at

Four things jumped out across the 12 prompts:

  1. Long-form text inside images. Magazine covers, posters, recipe cards, lecture slides — legible paragraphs, not smeared glyph soup.
  2. Non-English typography. Hindi and Japanese came back readable. Previous-gen models butchered these.
  3. Reference-photo identity across many edits. The fashion sheet held my face across 8 outfits. The follow-up (“zoom into the first look”) held it again.
  4. Weird aspect ratios. 2:1 panoramas, 1:3 ultra-tall stickers, multi-page manga stitched into one image. Respected without prompting tricks.

Patterns that worked

  1. Aspect ratio up front. “Portrait 2:3”, “ultra-tall 1:3”, or “2:1” — say it explicitly and the model composes accordingly.
  2. Identity first, edits second. Uploading a reference photo with “use my photo only for identity” gives the model permission to restyle heavily without losing your face. That phrase changed my hit rate.
  3. Follow-ups inherit context. One prompt later (“I like the first look, zoom in”) and the model held the exact face, outfit, and styling. No re-uploading.
  4. It will make meta content. The duct-tape prompt asked it to advertise itself using fake social-media quotes and a working QR code. It delivered both.
  5. Panel layouts come back as one image. Three-page manga, side-by-side fashion sheets, multi-language typography — no stitching tool needed.

Try it yourself

Head to chatgpt.com/images. Paste any prompt from the slideshow as-is — they're copy-ready. Swap the subject (me → you, bakery → your brand, Apollo 11 → anything else) and the rest of the scaffold still holds.

Read the official launch post for OpenAI's pitch and feature list.

If you make something cool, send it over on X or Instagram. I'll reshare the best ones.

atareh

Written by

@atareh

AI architect & creator. Writing, designing, and producing in AI and tech. Previously head of product at a healthtech SaaS; background in molecular science. Founded gogray.today in 2017.

Related

Keep reading.

Made by @atareh · x / twitter · instagram