AI Image Generation

AI image generators create new pictures from a written description.

How Text-to-Image Works

Most modern image generators use a method called diffusion. The idea, in simple steps:

1. During training, the model sees millions of images with captions.
2. It learns to take a picture full of random noise (like TV static) and remove the noise step by step until an image appears.
3. The text prompt guides each step, so the final picture matches the words.

So the model does not "search" for an existing photo. It builds a new image from noise, guided by your words.

Here is a toy version of "removing noise step by step" with numbers:

target = [10, 50, 90]          # the "clean image" the prompt describes
noisy = [80, 5, 30]            # random starting noise

for step in range(1, 6):
    noisy = [round(n + (t - n) * 0.5) for n, t in zip(noisy, target)]
    print("Step", step, noisy)

Step 1 [45, 28, 60]
Step 2 [28, 39, 75]
Step 3 [19, 44, 82]
Step 4 [14, 47, 86]
Step 5 [12, 48, 88]

Each step moves the values closer to the target. A real model does this for millions of pixels, and it *predicts* the target from the prompt instead of knowing it.

Writing a Good Image Prompt

A useful image prompt describes:

Put together:

A watercolour illustration of a young girl reading a book on a
rooftop in old Lahore, warm evening light, peaceful mood,
wide shot with the girl on the left, square format.

Editing Images with AI

Many tools now let you edit, not only create:

• Inpainting: select an area and replace it ("remove the bike from the background")
• Outpainting: extend a picture beyond its edges
• Style change: "make this photo look like a pencil sketch"
• Image to image: upload a rough sketch and get a polished design
• Chat-based editing: in multimodal chatbots, describe the change in words

Common Uses

• Social media posts and ad ideas for small businesses
• Product mock-ups before a real photo shoot
• Illustrations for presentations and blogs
• Quick concepts for logo and packaging design (a designer should still finish the final logo)

Problems to Watch For

• Text inside images can come out misspelled, so check every word.
• Hands, faces and small details can look wrong.
• Copyright: avoid asking for copies of famous characters, brands or a living artist's exact style for commercial work. Rules differ by tool and country.
• Fake images: realistic fake photos of real people or events can mislead. Many tools now add invisible watermarks or labels to AI images.

Exercise