Generating images with AI

What an image model actually does

A model trained on millions of images and captions builds a new image from your sentence. It does not search, it does not pull a file from anywhere, and it does not find an image; every output pixel is made at that moment.

That distinction has one practical consequence most tutorials skip. Run the same prompt twice and you get two different images, not a better one. So the common habit, hitting generate until one of the outputs comes out well, is buying lottery tickets. The real lever sits somewhere else.

That lever is the reference input: one or more images you send alongside the prompt so the model takes shape, style and character from them. It is the only way a product or a person stays the same across ten images in a row, and it is what fills the gap between a fun picture and a deliverable set. The ceiling on reference images differs by model, and the numbers are in the next section.

One more thing that is invisible on day one: these models do two separate jobs and they are not equally good at both. Generating an image from text is one job; editing an existing image is another, and the model order differs between them. If your work is mostly editing, the text-to-image ranking is not the answer to your question.

And the last thing worth saying up front: the output is a frame, not a delivery file. Cropping, format, weight, final colour and any text that has to sit on the image are still your job. An image that goes straight from the model to a client site is usually both heavier than it should be and not quite what was asked for.

Which models actually have numbers today

The AI section of this site already ranks six image models with published weights and sources, and we are not rebuilding that table here; the image model ranking lives where it lives. What a designer should take from it is three lines.

Line one: GPT Image 2 is today the only model that wins both measured boards, text-to-image and editing. Line two: Nano Banana Pro, the expensive Google model, sits below Google own cheap model Nano Banana 2 and costs twice as much. Line three: of those six, exactly one publishes open weights you can run on your own hardware.

We re-read the prices from each vendor own page on the day this lesson was written, and all five matched what the AI section already records. The cheapest is FLUX.2 klein 4B at $0.014 per image. GPT Image 2 is $0.053 at medium quality and $0.211 at high, four times as much. Nano Banana 2 is $0.067 for a 1K image, FLUX.2 max starts at $0.07, and Nano Banana Pro is $0.134.

The reference-input ceiling has its own published numbers, and for repeat work it matters more than the rank: Nano Banana 2 accepts up to ten object images, Nano Banana Pro six, FLUX.2 up to eight on the API and klein 4B up to four. If one fixed product has to recur across a set, that column decides for you, not the quality column.

And one thing you only learn by running it: the row that came first on a leaderboard names a configuration, not a model. The first GPT Image 2 row is medium quality, the $0.053 one. Pick high quality and your cost is four times larger with no measured score behind that choice at all.

Why Midjourney appears in none of our tables

Midjourney is everywhere in Persian recommendation lists and nowhere in our table. Not because it is bad. Because we have read nothing of it.

We tried again on the day this lesson was written: the home page, the documentation and even the robots.txt file of the Midjourney site all return 403 to our server. So we cannot read its pricing page, we cannot read its terms of use, and we cannot even read the rule we are meant to obey. For comparison, the documentation sites of OpenAI, Google and Black Forest Labs all open from this same server.

And the part that matters more: Midjourney has no row on either measured image board. The AI section of this site last read those two boards on 12 August 2026: seventy-seven models scored on text-to-image and fifty-three on editing, and its name was on neither. So we have no facts and no evidence, and we do not build a row without numbers.

Our position is that plain: if you read somewhere that a tool is the best, ask where the number is. A list of ten tools with no measured column is SEO content, not a buying guide. This is not an argument against Midjourney; it is an argument against any recommendation without a source, ours included on the day we stop citing one.

How to write an image prompt

A good prompt is not a description, it is a brief. Instead of writing one poetic sentence, you fix six things explicitly and leave the rest to the model: subject, what it is doing, frame and camera distance, light, medium or style, and colour palette. Leave any of those six out and the model picks one for you, and you only find out which when you look at the output.

The part rarely said out loud: piling on adjectives usually does nothing. Ultra high quality, 8K, cinematic, award winning is a row of words that adds no real constraint. A real number or ratio does far more work: landscape frame, subject in the left third, empty space at the top of the frame for text. Those are the things you discover were necessary the moment you drop the image into a layout.

Write negative constraints too, but keep them short. A long list of what you do not want takes as long as the list of what you do want, and its effect is not predictable in advance. Two or three things that genuinely go wrong every time is enough.

And the last step, the one that makes the actual difference: when an output finally has the right look, keep it as a reference input and build the rest of the set from it. That is what the reference column in the previous section exists for, and it is the only way ten images in a set come out looking related. Anyone who does not know this is rolling the dice from zero every time.

Persian text inside an image, where this breaks

For an Iranian designer this is the single most important question, and the honest answer is that no vendor has published a number about it.

Google describes text in its models as advanced. OpenAI writes that output may still have trouble with text. Black Forest Labs calls one of its variants specialised for typography. Three adjectives, zero measurements, and none of the three says anything about Persian at all. We have not tested these six models on Persian either, and until we do we will not print a number.

There is one practical recommendation that needs no number, and it comes from years of putting images on Persian sites: do not generate the text inside the image. Build the image with no lettering at all and put the text on top in your design tool or directly in CSS. You win three things at once: correct letterforms, the ability to translate that same image into another language, and text a search engine and a screen reader can actually read.

There are cases where this does not work, such as text that has to sit on a curved surface or inside the scene. Even there the answer is not to hope the model gets it right; the answer is an editing pass in a design tool.

Who owns the image you just made

For personal work this question is not serious. For work that has an invoice attached, the answer has to be known before the image is made, not after it is delivered.

In the additional terms for the Gemini API, Google states plainly that it will not claim ownership over content you generate, and then immediately adds the second sentence: you acknowledge that Google may generate the same or similar content for others and reserves all rights to do so. What you get is ownership, not exclusivity. For a logo that distinction is the whole story.

Black Forest Labs states its licences more clearly, because it publishes model weights: FLUX.2 klein 4B is Apache 2.0, so commercial work is free and it runs on a GPU with roughly thirteen gigabytes of memory. The 9B variant of the same family is under their own non-commercial licence, and the dev variant is non-commercial too. One family name, three different terms; if you read somewhere that FLUX is free, ask which FLUX.

About OpenAI and Midjourney we owe you a limit of our own: the terms-of-use pages of both return 403 to our server. OpenAI technical documentation does open and we quote only that; we write nothing about what we have not read. If you made an image with either and it is going into client work, read the terms from your own account.

And one point no licence solves: owning something is not the same as it being original. An image a model made can legally be yours and still resemble two hundred others produced with the same model that week. For a brand identity, that is a risk in itself.

Where to use these images and where not to

Our position is simple: an AI image does not replace a photograph, it replaces a stock photograph.

Anywhere the image is decoration and nobody makes a decision from it, these models beat any stock library on both speed and cost: a first draft of an idea for a client, a slide background, an explanatory shape inside an article, a mood board. Speed is worth something there and originality is worth nothing.

Anywhere the image makes a claim of its own, no. The product shot a customer buys from, the team photo, the portfolio piece, any image that says this is us or this is your product. A fabricated image in those places is not only a legal problem; it is a trust gap that does not close once somebody notices.

There is a third layer most tutorials never mention: what these models produce is usually not unmarked. Google image models put an invisible watermark on every output, and some services attach a signed record to the file as well. If you want to know what is inside the file you are uploading, the lesson on AI image watermarks and SynthID opens exactly that.

And if the output has to be a publish-ready file rather than a raw frame, that is the work done in our graphic design service.