Image Generation Basics

Text and image generation solve the same basic problem — produce plausible new content from a prompt — but the dominant technique for images works completely differently from the token-by-token approach in lesson 2.

Starting from noise

Most modern image generators are diffusion models. Generation starts from an image that's pure random static — visual noise, no structure at all — not a blank canvas.

Iterative denoising guided by a prompt

From that noise, the model repeatedly removes a small amount of noise, at each step nudging the image slightly closer to something that matches your text prompt, according to what it learned during training. After enough of these small steps (often a few dozen), what started as static has been sculpted into a coherent image.

Why the same prompt gives different images each time

Since generation starts from random noise, the same exact prompt run twice produces two different images, unless you fix the seed — the number that determines the model's starting noise pattern. Reusing the same seed and prompt together reproduces the same image; changing either one changes the result.

Note: text rendered inside a generated image (signs, labels, book covers) has historically been a weak point for diffusion models, because they're fundamentally reasoning about pixels and visual patterns, not about spelling — a model can be excellent at photorealistic faces and still misspell a word on a t-shirt.
PreviousText Generation Parameters