"a cat" and "an orange cat lounging on a windowsill, afternoon sunlight, shallow depth of field, film grain" produce completely different images. Text-to-image models aren't mind readers — the more specific you are, the more specific the result.

The Four Building Blocks

No formulas to memorize. Just cover four things:

  • Subject: what's in the frame. More specific beats generic — "an orange cat" over "a cat"; "a programmer wearing glasses" over "a person".
  • Setting: where. "on a windowsill", "a neon-lit street", "a plain white background".
  • Style: what texture. "watercolor", "film grain", "3D render", "flat illustration". Leave it out and you get photorealism by default.
  • Composition/light: how it's shot. "close-up", "overhead view", "warm afternoon side light", "shallow depth of field". This one shapes the mood the most.

String the four into a natural sentence — no need for comma-separated keyword soup. Modern models read plain language just fine.

Choosing Between the Two Engines

The tool ships with two engines, each with a personality:

  • GPT Image 2: better with long prompts and complex scene descriptions — use it when the image needs to tell a specific story.
  • Nano Banana 2 Lite: faster and lighter — great for quick drafts and mood pieces.

Run the same prompt through both once, and you'll know which fits your use case.

Match the Ratio to the Destination

  • Social media posts: 1:1 or 4:3
  • Phone wallpapers / stories: 9:16
  • Blog headers / banners: 16:9

Picking the ratio before generating preserves far more of the image than cropping afterwards.

Iterate — Don't Chase Perfection

The first image won't be perfect, and it doesn't need to be. Change one variable at a time: subject looks off, add descriptive detail; mood is wrong, adjust the light words; composition feels stiff, try another angle. Three rounds usually gets you something usable.

Open Text-to-Image in AI Image Toolkit and try the four building blocks from this post.