Artificial Lexicon
LLM guide

Prompting chat image models

Image models moved into the chat apps. Nano Banana lives in Gemini, GPT-Image lives in ChatGPT, and they read sentences, not keyword stacks. The walls still feed them.

Free Nano Banana GPT-Image 2 min read

What changed

Midjourney and Stable Diffusion take a stack of keywords and render it. The chat image models take a conversation. Google's Nano Banana and OpenAI's GPT-Image-v2 sit inside Gemini and ChatGPT, read your prompt as a sentence, remember the image they just made, and change it when you ask.

That shifts the craft in two ways. Prompts read better as described scenes than as comma lists. And the reroll loop becomes an edit loop: instead of regenerating from scratch, you say what to fix.

Same words, different dialect

The vocabulary transfers. A style word, a lighting word, and a concrete detail steer these models the same way they steer Midjourney. What changes is the packaging: wrap the words in a sentence instead of stacking them.

keyword dialect (MJ/SD)a tea house in the mountains, ukiyo-e, golden hour lighting, paper lanterns, mist --ar 3:2
chat dialect (Nano Banana, GPT-Image)Make a wide image of a tea house in the mountains in the style of an ukiyo-e print. Golden hour light, paper lanterns by the door, mist in the valley.

Note what happened to the parameters. There is no --ar; you say wide or portrait in words. No --no either; say no people in the sentence. The walls and the builder still supply the nouns; the sentence supplies the glue.

The edit loop

This is the real advantage. The model keeps the image in context, so the one-change-at-a-time habit becomes a conversation.

follow-upsKeep everything, but make it dusk. Same scene, swap ukiyo-e for gouache. Add a cat on the porch. Small, not the focus.

Each follow-up is one experiment, and you never lose the parts that already worked. When a change goes wrong, say so and the model backs it out. This replaces seeds for most day-to-day work.

What they do that the others don't

  • Reference images. Upload a photo and ask for it restyled, extended, or repainted in a wall word: redraw this in linocut print.
  • Legible text. They render signs, labels, and titles far better. Put the exact wording in quotes.
  • Consistency. Ask for the same character or product across several images and they mostly hold the design.
  • Questions. They can ask what you meant. Answer instead of rewriting the whole prompt.

When Midjourney and Stable Diffusion still win

Exact aspect ratios, seeds, and batch variations are still Midjourney territory; --chaos exploration has no chat equivalent. Stable Diffusion still owns fine-tunes, LoRAs, and local control. The Midjourney and Stable Diffusion pages cover those parameters.

A working split: explore composition and style in a chat model where edits are cheap, then take the settled prompt to Midjourney or Stable Diffusion when you need control over frame, seed, or volume.

Published here first.

Keep going