Images and Beyond

The same season Plateful's chatbot launched, the marketing team quietly stopped booking photographers. Their new banner — steam rising off a bowl of ramen, chopsticks lifting noodles into golden light — came from a text description typed into a generator. No camera, no kitchen, no ramen. Iris's first question is the right one: is this a different technology, or the same trick from Topic 40 applied to a new kind of output?

The same trick. Text was only the first medium. The generate-by-prediction idea — produce content that fits the context — extends to images, voices, music, and video, each with its own twist on "predict what fits". This page tours the family without leaving the book's mechanics, because you already own every concept the tour needs.

One Idea, New Canvases

Strip Topic 40's loop down to its core and it says: given a context, generate content that fits it. For the chatbot, the context is the conversation and the content is words. For the marketing banner, the context is your text description — "steam rising off a bowl of ramen, warm light" — and the content is pixels. And you know from Topic 07 exactly what pixels are to a model: a grid of numbers. Reading that grid was Topic 37's job, recognizing dishes in photos. Generation runs the ambition in reverse — the grid is no longer being read; it is being written.

What carries over from text generation is the deep structure: a model trained on mountains of examples learns what fitting content looks like, then produces new content that fits a context it has never seen. What changes is the "how" — because unlike a sentence, an image does not arrive naturally one word at a time. The dominant trick for images deserves its own honest paragraph.

Diffusion in Plain Words

The trick is called diffusion, and it is best told as a story in two halves. In training, the model plays a strange game millions of times: take a real photo, spoil it with a little random static, and learn to undo the damage — over and over, at every level of spoilage, until it can look at pure noise-speckled mess and say "this part is probably an edge; that region is probably sauce". The model never learns to paint. It learns to de-noise — to remove what does not belong.

Then comes the reversal that sounds like a joke and works like magic: to generate, start from pure random static — an image of nothing — and ask the model to "restore" it, step by step, as if it were a photo that had been spoiled. Your description steers every step: guided by "steam rising off a bowl of ramen", the model removes the static that does not look like ramen. Rough shapes emerge, then edges, then detail, and after a few dozen refinement steps the nothing has become the banner photo. A sculptor of the old school claimed to see the statue in the marble and simply remove everything that was not the statue. Diffusion is that sculptor: it sees the dish photo in the static, and your description tells it which statue to find.

Pure staticno image at all

→

Refinerough shapes emerge

→

Refine againedges, color, detail

→

Finished photo"steam rising off a bowl of ramen"

Voices and Music

Sound joins the family the same way. Topic 07 told you what audio is to a model — a waveform, a long sequence of numbers — and generating it means the familiar move: produce sound that fits the context. Give the context as a few seconds of someone's voice, and the model continues in that voice — reading any text you like, in a voice that never spoke it. Give the context as a text description of a mood and a genre, and out comes music that fits the vibe. The same species of trick as the ramen banner, on a different grid.

Astonishing — and worth one sober sentence before moving on: a machine that can produce anyone's voice saying anything is the reason "hearing is believing" is quietly ending, and the deepfake conversation that follows from it gets its full reckoning in Chapter 10.

Multimodal — the Borders Dissolve

The newest step is that the media stopped being separate systems. Modern models increasingly read and produce across them: show a photo and get a description; hand over a rough sketch plus a sentence and get a refined image; talk out loud and get an answer spoken back. These are called multimodal models — one system working across text, images, and audio rather than one per medium — and that single sentence is all the vocabulary you need.

The honest note to close the tour: every new canvas inherits this chapter's diseases along with its powers. An image generator produces the likely, not the verified — which is how you get a photorealistic hand with six fingers, or a generated street sign spelling a word that does not exist. That is Topic 42's hallucination, in pixels. The medium changes; the mechanism, and its native failure mode, travel together.