Text-to-image generation

Text-to-image generation means that an AI model makes a picture from a written description, the prompt. You type "a bicycle on a canal bridge at sunset, in watercolour" and get a new image. Businesses use it for social media visuals, illustrations, product mock-ups and mood boards. Before you publish one, check what the tool's terms allow and whether the image looks too much like existing work.

How does it work?

A text encoder first turns the prompt into embeddings, lists of numbers that capture its meaning. Stable Diffusion v1 uses OpenAI's CLIP for this. With Imagen (2022), Google found that a larger text encoder, the language model T5, improved the images more than a larger image model did (Saharia et al.). An image model then builds the picture. There are three main approaches:

• Diffusion models start from random noise and remove it step by step, steered by the text. Stable Diffusion and DALL-E 3 work this way.
• Autoregressive models write the image as a series of image tokens, the way a language model writes text. OpenAI's first DALL-E (2021) and Google's Parti (2022) did this (Ramesh et al., Yu et al.), and OpenAI calls the image generation in GPT-4o (March 2025) autoregressive (system card addendum).
• Flow matching, a training method that includes diffusion as a special case, is the basis of Black Forest Labs' FLUX.1 models, released in August 2024 (announcement).

How do you steer the image?

Open tools such as the diffusers library, which runs Stable Diffusion, let you set more than the prompt:

• a negative prompt, for what should not be in the image;
• the guidance scale, 7.5 by default: a higher value follows the prompt more closely, at the expense of image quality;
• the number of denoising steps, 50 by default: more steps usually give a better image but take longer;
• a seed for the random generator, so the same settings give the same image again.

ControlNet (2023) adds control over the layout, with edges, a depth map or a human pose.

Where does it go wrong?

The model card of Stable Diffusion v1 (2022) lists the typical weak points: no legible text, trouble with compositions such as "a red cube on top of a blue sphere", faces and people that come out wrong, weaker results outside English, and some memorisation of images that appear many times in the training data. Newer models do better on some of these. OpenAI says GPT-4o reliably puts text into images; for DALL-E 3 it called that capability unreliable.

Can a business use the images?

That depends on the tool. Midjourney's terms (May 2026) let you own the images you make, but a company with more than $1 million a year in revenue needs a Pro or Mega plan for that. Adobe says it never trained Firefly on customer content, pays the Adobe Stock contributors whose work it used, and calls the models commercially safe. Stable Diffusion v1 learned from subsets of LAION-2B(en), a public collection of more than 2 billion images with English captions. No contract gives more than the law allows: in the United States, the Copyright Office concluded in January 2025 that purely AI-generated material is not protected by copyright and that prompts alone do not give enough control. To mark images as AI-made, OpenAI said in March 2025 that it adds C2PA metadata to all GPT-4o images, and Google uses SynthID.

Example

Suppose a bakery wants a spring poster. The owner describes the scene, generates a few versions and picks one. The prices go in afterwards with a layout program, because the model may garble text. Before printing, she checks that the tool's terms allow commercial use and that the image does not resemble a well-known brand or artwork.