Most AI image prompts fail for one of two reasons: they are too vague (“a nice photo of a city”) or too long (a 200-word paragraph the model half-ignores). A prompt built from six explicit parts gets you a usable image in 1-3 generations instead of 15. This guide gives you that structure, the three mistakes that flatten output, and which tool to pick for your style as of June 2026.
TL;DR
- Write every prompt as six parts: subject, style, composition, lighting, mood, use-case (aspect ratio).
- Keep it to roughly 40-80 words. Most models start under-weighting tokens past that.
- On the second pass, change one component, not three, so you can tell what moved.
- Pick the tool to match the job: Midjourney for art direction, ChatGPT Images 2.0 for prompt-following and legible text, Gemini “Nano Banana” for fast iteration and editing.
The six components
Order them this way in your prompt. Each line does one job.
- Subject — one specific noun, not a category. “A red fox sitting on a mossy rock,” not “an animal.”
- Style — one named reference. “Watercolor illustration,” “1970s film photography,” “isometric vector art.” One, not three.
- Composition — where things sit in the frame. “Centered subject, low angle, rule of thirds, shallow depth of field.”
- Lighting — this is what makes an image feel a certain way. “Warm golden-hour sidelight” reads differently from “cool overcast diffuse.”
- Mood — 2-3 adjectives. “Calm, melancholic, intimate.”
- Use-case — aspect ratio and purpose. “16:9 blog hero, no text overlay, space on the left for a headline.”
The use-case line is the one beginners skip and regret. Generating a square then discovering you need 16:9 means starting over.
Pick the right tool (June 2026)
The big three handle the same prompt structure but reward it differently. Pricing and model details below are current as of June 2026.
| Tool | Model / version | Strength | Aspect ratio control | Entry price |
|---|---|---|---|---|
| Midjourney | V7 (V8.1 in alpha) | Art direction, painterly look | --ar 16:9 flag | Basic $10/mo (~3.3 fast GPU hrs, ~200 images) |
| ChatGPT | Images 2.0 (GPT Image 2) | Prompt-following, legible in-image text | Plain words: “16:9 widescreen” | Plus $20/mo (also on Free) |
| Gemini | Nano Banana 2 / Pro | Fast iteration, conversational editing | 9 fixed ratios incl. 16:9, 9:16, 21:9 | Free tier ~500 images/day in AI Studio |
Notes that change the choice:
- ChatGPT Images 2.0 shipped April 21, 2026 and topped the Image Arena text-to-image leaderboard at launch. Its standout upgrade is text rendering, so it is the one to use for posters, ads, menus, or UI mockups where words must be readable. OpenAI retired DALL-E 2 and DALL-E 3 on May 12, 2026; GPT Image 2 is now the only model behind ChatGPT image generation.
- Midjourney V7 rewards natural-language art direction over keyword lists, and it exposes dials the others hide:
--s(stylize, 0-1000; most pro work sits at 200-400),--chaos(0-100 for variety), and--noto exclude an element. See the Midjourney beginner guide to set up an account and your first prompt. - Gemini “Nano Banana” is the cheapest to experiment with and the best at “now change just the hat” style edits. Walk through that loop in the Nano Banana image editing tutorial.
A real workflow: a blog hero in four passes
Goal: a 16:9 hero image for a post about remote work.
- Pass 1 — full six-part prompt at 16:9, generate 4 variants.
- Pass 2 — pick the closest variant. It reads too corporate, so change only the mood line (“warm, lived-in, quiet morning”). Regenerate.
- Pass 3 — composition is busy. Change only the composition line to add “negative space on the left for headline text.” Regenerate.
- Pass 4 — lock it. Upscale, then add the headline in an editor, not in the prompt.
If you are working inside ChatGPT specifically, the iterate-one-variable loop in that UI is covered in this ChatGPT image tutorial.
The three traps that flatten output
- The wall of text. A 200-word paragraph describing everything at once gives the model nothing to prioritize. Keep prompts to about 40-80 words and let structure carry the detail.
- Stacking styles. “Oil painting, anime, Pixar, hyperrealistic” cancels itself out into mush. Pick one named style and commit. If you want a blend, weight it deliberately rather than piling on adjectives.
- Asking for text inside the image. Text rendering improved a lot in 2026 (GPT Image 2 is genuinely usable for short labels), but it is still the least reliable thing you can ask for. For anything longer than a few words, generate the image clean and set type in an editor.
Advanced tips
- Save winners as templates. Lock five of the six lines and change only the subject for brand-consistent series.
- Reuse the seed. AI image generation is stochastic, so the same prompt rarely repeats. If your tool exposes a seed (Midjourney does, via
--seed), capture it and iterate from there for near-identical framing. - “Candid” beats “portrait” for real-looking people. A “candid photo” prompt usually reads as less staged than “studio portrait.”
- Negative prompts are subtractive, not magic. In Midjourney,
--no text, watermarkremoves elements, but the model ignores a--nothat contradicts your subject.
Copy-ready template
Replace each bracket with your own line and paste the result as one prompt. Brackets are placeholders, not literal syntax.
Subject: [one specific thing, e.g. a red fox on a mossy rock]
Style: [one named style or era, e.g. 1970s film photography]
Composition: [framing and angle, e.g. centered, low angle, rule of thirds]
Lighting: [direction, temperature, mood, e.g. warm golden-hour sidelight]
Mood: [2-3 adjectives, e.g. calm, melancholic, intimate]
Use-case: [aspect ratio and purpose, e.g. 16:9 blog hero, no text overlay]For Midjourney, append the aspect ratio and dials as flags, for example --ar 16:9 --s 300.
FAQ
Which tool produces the best images in 2026? It depends on the job. Midjourney V7 wins for artistic and stylized work, ChatGPT Images 2.0 (GPT Image 2) wins for following exact instructions and rendering readable text, and Gemini Nano Banana wins for cheap, fast iteration and conversational edits. Test all three on your own style before committing a subscription.
Why does the same prompt give different results every time? AI image generation is stochastic, so each run samples differently. If your tool exposes a seed value, save it and iterate from that seed to keep framing consistent.
How long should an image prompt be? Aim for roughly 40-80 words across the six components. Past about 80 words most models start under-weighting later tokens, so extra detail often gets dropped rather than honored.
Can AI reliably put text inside an image now? Short text yes, long text no. GPT Image 2 (shipped April 2026) renders short labels and headlines far better than older models, but for paragraphs or precise typography you should still generate the image clean and add the text in a design editor.
Do I have to pay to follow this guide? No. ChatGPT Images 2.0 is available on the Free plan with limits, and Google AI Studio offers a free Gemini image tier (around 500 images per day as of June 2026). Midjourney has no free tier; its cheapest paid plan is $10/month.