From Words to Worlds
How AI turns a sentence into a picture — and pictures into video. Pick any concept to explore. Press Esc anytime for this menu.
A picture hidden in noise
AI image makers don't paint stroke by stroke. They start with pure random static — like an untuned TV — and, guided by your words, clean it up a little at a time until a picture emerges. Drag the slider to watch noise become an image.
Denoising step
Prompt
“a lighthouse at sunset”
The words decide what appears as the noise clears.
Same trick — meet AI
What “denoising” actually means
“Denoising” isn't magic — it's a loop. The AI looks at the messy image, predicts which part is noise, removes a little of it, and looks again. Repeat ~20–50 times and a clear picture is left behind. Step through one round at a time.
a bit of noise
Step 0 of 20
The loop
Look → predict the noise → subtract a little → look again. Each pass the picture gets a bit clearer. That repeated cleanup is denoising.
Same trick — meet AI
How the prompt picks the “noise”
Same static, different word, different picture. At every step the AI predicts what to remove so that what's left looks more like your prompt. Type “dog” and it steers toward a dog; type “cat” and the very same noise becomes a cat. The prompt is the steering wheel.
View
Your prompt
Step 0 of 20
Pick a prompt and press Step.
❌ The myth
“The dog is already hidden in the static, and AI rubs away the non-dog bits to reveal it.” There's no dog in random noise.
✅ The truth
Each step the AI predicts noise to remove so what's left looks a little more like your word. The picture is steered into being, not uncovered. Change the word → change every prediction.
It sketches, then paints
As the noise clears, the AI builds the picture in layers — first flat blocks of colour for the big regions, then the basic shapes, then the fine detail last. Like an artist blocking in a rough sketch before painting. Step through it.
Build stage
1 · Blocks of colour
First it just decides the big regions: sky on top, sea below, a tall something in the middle. Flat colour, no shapes yet.
Why it matters
Working big-shapes-first is why AI images look coherent — it decides the layout before sweating the details.
Words steer the picture
Same scene, different words. Toggle parts of the prompt and watch the picture restyle instantly. This is why prompting is a skill: every word nudges the result.
Time of day
Style
Your prompt
a lighthouse, day, photo
Anatomy of an image prompt
A great image prompt has parts: the subject, the style, the lighting, the shot. Add each one and watch a generic picture sharpen into exactly what you pictured.
Build the prompt
Prompt so far
(empty)
Sharpness: vague
Telling it what NOT to draw
As well as saying what you want, you can list what you don't want — a “negative prompt.” It pushes those things away: no text, no extra fingers, no blur. Toggle the things to banish and watch them disappear.
Remove from the picture
Negative prompt
(nothing banished yet)
Same trick — meet AI
Start from your own picture
You don't have to start from pure static. Hand the AI your rough sketch or photo, it adds a little noise instead of a lot, then denoises — so it keeps your composition but repaints it in a new style. Slide how much freedom you give it.
Creative freedom: low
Low = sticks close to your doodle. High = takes big liberties.
Why it works
Starting from your picture (plus a little noise) means the denoising keeps your shapes and layout — it restyles instead of inventing from scratch.
Same trick — meet AI
Change just one part
You don't have to redo the whole image. Paint a mask over one area — the sky, a face, an empty spot — and the AI regenerates only that, blending it in. That's how “remove the background,” “add a hat,” or “erase that person” work.
Pick an edit
Same trick — meet AI
The shape of the canvas
Before it draws, you choose the frame: square for posts, tall for phone screens, wide for banners. The shape changes how the scene is composed — and asking for the wrong shape is a common reason results look cramped or stretched.
Frame
Square 1:1
Balanced and safe — great for profile pics and feed posts. The subject sits centred.
Tip
Match the frame to where it'll be seen. A wide banner prompt squeezed into a tall frame will crop or distort the subject.
Same words, many pictures
Type the same prompt twice and you get different images — because each starts from a different random “seed.” Lock the seed and the picture repeats exactly. Change it and you explore endless versions.
Prompt (fixed)
“a mountain landscape”
4 different seeds
Same prompt, four random starting points → four different pictures. That's why you never get the same image twice.
Same trick — meet AI
It learned from captions
Nobody taught AI to draw with rules. It studied billions of pictures that each came with a caption — “a red apple,” “a dog on a beach” — until it learned which words go with which shapes and colours. Like a child with a picture book.
What it's picking up
Same trick — meet AI
Meaning is a map
Inside the AI, similar things sit close together on a “map of meaning.” That's why it understands a kitten is near a cat, and night is the opposite of day — and why it can blend concepts you never showed it together.
Try a concept blend
Concepts live on a map
Nearby = similar. The AI can move along the map to blend ideas — even ones it never saw combined.
Same trick — meet AI
Why it got so good, so fast
Just a few years ago AI images were melty, low-res nightmares. More data, bigger models and better methods turned them photoreal in record time. Drag through the years to watch the same prompt improve.
Year: 2021
2021 — blobby, dreamlike, barely there.
What changed
Same idea (“a corgi astronaut”), wildly different quality. More training images, bigger models, and smarter denoising = the leap you're seeing.
Same trick — meet AI
Video is a flipbook
There's no such thing as “video” — it's just still pictures shown fast. Flip ~24 of them every second and your eye sees motion. AI video makes each frame as an image, then plays them in order. Speed it up and down to feel it.
Subject
Frames per second: 1
Slow — you can see each separate picture.
Frame count
This 2-second clip = 48 separate images at 24 fps.
Same trick — meet AI
Keeping it steady
The hard part of AI video isn't drawing one frame — it's making the next frame match. If each is drawn fresh, colours and shapes jitter and “morph.” Good models keep frames consistent. Toggle it to see the difference.
Generation mode
Each frame drawn fresh
Watch the character flicker — colour, size and face jump frame to frame. This is the tell-tale “AI video” wobble.
Same trick — meet AI
Animating a still photo
Give a tool one photo and a direction — “zoom in,” “pan left,” “make the clouds drift” — and it invents all the frames in between that the photo never had. One picture becomes a moving shot.
Motion
Same trick — meet AI
Why video burns credits
One image is one picture to make. One second of video is ~24 pictures. A short 5-second clip is over a hundred — each one as much work as a full image. That's why free video credits vanish so fast.
What are you making?
1 image
A single picture — one unit of work. Cheap, often free.
The honest takeaway
Free tiers are generous with images, stingy with video — because video is dozens of images. Plan your clips before you spend.
Spot the deepfake
AI images are getting scary-good, but they still leave clues. Study the realistic render, then switch to the illustration to learn the tells — or load a real AI photo of your own. Click “Reveal” to highlight the giveaways.
Image
Common giveaways
Why it happens
AI paints what's plausible, not what's logical — so it nails the vibe but flubs the details that need counting or reading.
Why it fumbles hands
Ask for a hand and you might get six fingers. Ask for a sign and the letters look like a dream. It's not “dumb” — it learned what things look like, not the rules that a hand has five fingers or that words spell something. Toggle between them.
Show me
Hands are hard
It's seen millions of hands in every pose, so it knows the “look” — but not the rule that there are exactly five fingers. So it guesses, and often miscounts.
Getting better
Newer models are improving fast on hands and text — but it's still the quickest way to catch an AI image today.