From words to pictures
AI image generators create new images from a text description, called a prompt. Most are built on a technique called diffusion, and some newer systems are built directly into large language models. You do not need the technical details, but a simple mental model helps you get better results.
During training, the model sees an enormous number of images paired with descriptions. It learns the relationships between words and visual patterns: what “golden hour” lighting looks like, how a “watercolor” texture differs from an “oil painting,” what a “close-up” frames. When you write a prompt, the model generates a brand-new image that fits those patterns. It is not searching for or pasting together existing pictures, though it can sometimes produce results that closely resemble its training data.
Why results vary
- Every generation is different. The same prompt usually produces different images each time. That is useful for exploring options, and frustrating when you want consistency (covered later in the course).
- Vague prompts lead to “average” images. If you only say “a dog,” the model fills in everything else with typical choices.
- Some details are hard. Hands, crowds, exact counts of objects, and precise text in images have historically been weak spots. Many current tools handle these much better, but check closely.
- The tools have rules. Most refuse certain content, such as realistic images of real people in misleading situations, explicit content or trademarked characters, and rules differ by tool.
Two ways to work
1. Chat-based image creation. Assistants like ChatGPT and Gemini can generate images inside a normal conversation. You describe what you want, then refine by chatting: “make the background blue,” “add a coffee cup on the table.” This is the easiest way to start.
2. Dedicated image tools. Tools like Midjourney, Adobe Firefly, Canva and others offer more control: style settings, aspect ratios, variations, editing brushes and integration with design work.
Example: chatting your way to an image
Create an image of a cozy independent bookshop at dusk, seen from
the street. Warm light glows through the windows, a cat sits on the
windowsill, and it has just stopped raining so the pavement reflects
the light. Realistic photo style, horizontal format.
Then refine:
Great. Make the shop sign say "Chapter & Verse" in elegant gold
lettering, and add a couple walking past with an umbrella.
Look closely at the result: Is the sign spelled correctly? Do the people look natural? Small fixes are often easier with a follow-up or an editing tool than by starting over.
What AI images are good for
- Social media graphics and blog illustrations
- Concept art, mood boards and early design ideas
- Backgrounds and scenes for marketing
- Visualizing products in different settings (with care to stay accurate)
- Presentations and teaching materials
What to be careful about
- Showing products inaccurately. If an AI image makes your product look different from reality, customers will notice, and it may count as misleading advertising.
- Real people. Do not create realistic images of real people without consent, especially in situations that could embarrass, deceive or harm them.
- Rights and licensing. Each tool’s terms explain how you can use what you make. More on this in the final lesson.
Practice task
Using any chat-based assistant that can create images (such as ChatGPT or Gemini), generate an image from a detailed scene description like the bookshop example. Refine it with at least three follow-up requests. Inspect the final image closely for errors in text, hands, object counts or physics, and note what the tool did well and badly.