That Midjourney Prompt Won't Do the Same Thing in GPT Image 2

Why That 1,600-Prompt Pack Disappointed You

Search for image prompts and you get collections — 128 photography prompts, 1,600 Midjourney prompts, 60 best prompts for AI art. Every one of them is written as though a prompt is a portable thing you paste anywhere.

It is not. The same string produces materially different output depending on which model receives it, because the models were built around different strengths:

A dense, keyword-stacked Midjourney prompt — the comma-separated style of "cinematic, 8k, hyperdetailed, octane render, trending on artstation" — does something useful on Midjourney and close to nothing on GPT Image 2, which is reading your sentence as an instruction rather than as a mood board.

The Six Parts of a Prompt That Travels

Write these explicitly and the prompt survives a change of model. Leave them implicit and you are relying on whatever that model defaults to.

1. Subject — what is in frame, and what it is doing. Specific nouns beat adjectives.
2. Composition — where the camera is. Close-up, wide, overhead, eye level, three-quarter.
3. Lighting — the single highest-leverage word in most prompts. Golden hour, overcast, single hard key light, candlelit, studio softbox.
4. Setting — where, and what time. Including what is in the background, because something will be.
5. Medium or style — photograph, oil painting, flat vector, 3D render, pencil. State it or the model picks.
6. Exclusions — what must not appear. More useful than most people expect, and ignored by some models.

Lens and camera language — 35mm, f/1.8, shallow depth of field — is a seventh that works well on photographic models and is noise on vector or illustration ones.

Prompt 1 — Realistic Photograph

A [subject] [doing what], photographed [composition — e.g. three-quarter view, eye level]. [Setting], [time of day]. Lit by [specific light source], [quality of light — soft, hard, directional]. Shot on 35mm, f/2.0, shallow depth of field. Natural skin texture, visible grain, no retouching. No text, no watermark, no logo.

The part that does the work: "natural skin texture, visible grain, no retouching". Models default to a glossy, over-processed look because that dominates their training data. Asking for imperfection is what moves it toward photographic.

Best on: GPT Image 2 for adherence, Midjourney if you want it to look intentional rather than accurate.

Prompt 2 — Product Shot

[Product], centred, on [surface]. Studio lighting: large softbox at 45 degrees camera left, white bounce card camera right, subtle rim light from behind. Background [colour], seamless, slightly out of focus. Product in sharp focus throughout. Realistic material rendering — [specify: brushed metal, matte plastic, frosted glass]. Square crop. No props, no hands, no text.

The part that does the work: naming the lighting positions. "Studio lighting" alone gets you a generic glow; describing a three-point setup gets you something that looks shot rather than generated.

Best on: Seedream 4.5 if you need 4K and consistency across a catalogue, GPT Image 2 for precise material description.

Prompt 3 — Social Media Graphic With Text

A [format — square post, vertical story] graphic. Headline text reading exactly: "[your headline]". Text is the dominant element, [describe: bold sans-serif, high contrast, centred]. Background: [simple description — solid colour, soft gradient, abstract shape]. Leave clear margin around the text. Nothing else competing for attention.

The part that does the work: "reading exactly" plus the quoted string. Models that handle text well still paraphrase unless you mark the string as literal.

Best on: Ideogram 4.0, which exists for this. GPT Image 2 if the text is not in Latin script.

Prompt 4 — Illustration in a Specific Style

[Subject], illustrated. Flat vector style with [number] colour palette: [list the colours]. Thick uniform line weight, no gradients, no shadows. Geometric shapes, simplified forms. Centred on [background colour], generous white space. Consistent with a brand illustration set.

The part that does the work: naming your actual colours. "Flat vector style" gets you a generic one; four named hex-adjacent colours gets you something that matches the rest of your material.

Best on: Recraft V4 if you need the file to be an actual SVG. Anything else gives you a picture of a vector illustration, which is not the same thing and will not scale.

Prompt 5 — Keeping a Character Consistent

Using the attached reference image(s) as the definitive appearance of this character, generate: [new scene description]. Keep facial features, hair, build and clothing identical to the reference. Change only the setting, pose and lighting as described. Same illustration style as the reference.

The part that does the work: separating what must stay from what may change, explicitly. Without that split, models drift on both.

Best on: FLUX.2 for multi-reference conditioning, Seedream 4.5 which accepts up to six reference images, or Nano Banana 2 where you can correct drift conversationally rather than re-prompting.

Prompt 6 — Fixing One Thing Without Redoing Everything

Keep this image exactly as it is. Change only: [the one specific thing]. Do not alter composition, lighting, colour grading, or any other element. If the change is not possible without affecting something else, tell me what would be affected before doing it.

The part that does the work: the last sentence. It turns a silent full regeneration into a question, which is the difference between editing and starting again.

Best on: Nano Banana 2, built around conversational editing. Most other models will quietly redraw the whole thing.

What I Have Not Done Here

I have not run these six prompts across all six models and compared the output. Doing that properly means a fixed seed where available, identical settings, and a published grid — and until someone does it, any claim about which produced the better image is an opinion with a confident tone.

What this page gives you is prompt structure, which is model-independent, and what each vendor documents its model as being built for. The pricing and capability detail behind that table comes from our text-to-image tools comparison, where every figure is sourced to the vendor.

Take the structure, run it on whichever model you already pay for, and keep the version that works. That is more useful than a pack of 1,600 prompts written for a model you may not be using.

FAQ

What are the best AI prompts for realistic photos?

Ones that ask for imperfection. Specify the light source and its quality, add lens language like 35mm and f/2.0, and explicitly request natural skin texture, visible grain and no retouching. Models default to a glossy, over-processed look, so the realism comes from what you ask them not to do.

Do AI image prompts work across different generators?

Partly. Subject, composition, lighting and setting travel. Style keyword stacks do not — a comma-separated Midjourney prompt reads as a mood board to Midjourney and as literal instruction to GPT Image 2. Write the six components explicitly and the prompt survives the move.

What is the best prompt for an AI image generator for social media?

One that marks the headline as literal — "text reading exactly: [your words]" — and states that text is the dominant element with clear margin around it. Ideogram 4.0 is built for typography; GPT Image 2 handles non-Latin scripts.

How do I get consistent characters across images?

Supply reference images and separate explicitly what must stay identical from what may change. FLUX.2 supports multi-reference conditioning, Seedream 4.5 accepts up to six references, and Nano Banana 2 lets you correct drift conversationally.

Why does the same prompt give different results on different models?

Because they were built around different strengths. Midjourney reads aesthetic direction, GPT Image 2 reads instruction, Ideogram is organised around typography, Recraft outputs vector. The prompt is the same; what the model is optimising for is not.

Are long prompts better than short ones?

Specific is better than long. Six well-chosen elements beat forty stacked keywords, and on instruction-following models the keyword stack actively hurts, because every term is treated as a requirement rather than a hint.

These six templates are a starting set. Our library has 744 prompts across 123 collections for ChatGPT, Claude, Gemini, Grok and Kimi.