The one-sentence version
An image model was trained on millions of image-and-caption pairs until it learned the statistical link between words and what they tend to look like. When you prompt it, it starts from a field of random visual static — literal noise — and, guided by your words, denoises it step by step into a picture. You describe; it renders from static. That's the whole trick.
This one fact explains almost everything you'll notice. The model isn't pulling a photo from a library and it isn't drawing the way you would. It's making a confident guess about what "a hand-poured soy candle on a linen table in morning light" should look like, based on everything it has ever seen labeled with those words. Vague words, vague guess. Specific words, specific guess. That is why the six-slot recipe in Lesson 3 works so well — it replaces vague with specific, slot by slot.
Why hands used to be weird (and what it teaches you)
For a while, AI hands were a punchline — six fingers, fused knuckles, thumbs in the wrong place. The reason is genuinely useful to understand: hands appear in countless positions, from countless angles, half-hidden behind objects, so the "average" of all those hands is a mess. The model was confident about hand-ish and fuzzy about the exact count. Newer models have largely fixed this, but the lesson stands: the model is strongest where the training data was clear and consistent, and weakest where it was chaotic. Clean product shots, tidy flat-lays, and simple scenes come out reliably. Complicated hands, crowds, reflections, and precise text are where you still babysit.
Why text-in-image needs the right tool
Older models treated letters as shapes to imitate, not language to spell — which is why you'd get a "SALE" sign that said "SAEL." That has changed fast. As of 2026, a few models render clean, correct text on purpose (you'll meet them in Lesson 2), while others still garble it. The practical rule for now: if the words in the image matter, either use a tool built for text, or — better for most brand work — generate a clean background and add your text in Canva. We'll do exactly that for Pinterest graphics in Lesson 5.
Seed and style reference: your consistency levers
Two small controls do the heavy lifting for brand work, so meet them now:
- Seed is the specific starting patch of noise the image grew from. Same prompt + same seed ≈ the same image again; change the seed and you get a fresh variation. When something is almost right, holding the seed and tweaking one word lets you nudge instead of gambling.
- Style reference (names vary by tool) lets you point at an existing image — or a code that represents one — and say "make new images that feel like this." This is how you carry a look across a whole set instead of re-describing your aesthetic every time.
You don't need to master these today. Just know they exist, because they're the mechanism behind the brand-consistency system in Lesson 4 — the chapter that turns "nice one-off images" into "a feed that looks like one brand."
You are not commissioning an artist who reads your mind. You are directing a very fast, very literal renderer that guesses from your words. Your job isn't to be artistic — it's to be specific, then to pick the best result. Specificity and picking. That's the entire skill, and both are learnable.
What this means for the rest of the course
Everything ahead is built on this page. Lesson 2 picks the right renderer for the job. Lesson 3 makes you specific on purpose. Lesson 4 makes you consistent. Lesson 5 turns it into eight repeatable business recipes. Lesson 6 keeps you on the right side of the law. Lesson 7 hands you a hundred prompts so you rarely start from a blank box again.
Open whatever image tool you already have access to and generate one image from a deliberately vague prompt — just "a candle." Look at how generic and un-branded it is. Keep that image; in Lesson 3 you'll rebuild the exact same subject six ways and watch it go from forgettable to editorial. That "before" shot is your baseline.