Text-to-Video Basics
What AI video models actually deliver — clip lengths, motion description, and the honest list of what they can and can't do yet.
In 2023, AI video meant faces that melted mid-frame and hands with a suspicious number of fingers. Today, Sora 2 and Veo 3.1 produce shots that regularly fool professional editors. The catch: they reward people who think like directors, not people who type wishes. This lesson resets your expectations to reality — and reality is pretty exciting.
What One Generation Actually Gives You
A single text-to-video generation produces one continuous shot of roughly 4-12 seconds, usually at 24 frames per second. Veo 3.1 delivers 8-second clips at up to 1080p; Sora 2 goes up to about 10-15 seconds. That is not a limitation to work around — it is the same unit real filmmakers work in. The average shot in a modern movie lasts under 4 seconds. You are not generating films; you are generating shots, and films are edited sequences of shots.
Describe Motion, Not a Picture
The biggest beginner mistake is writing an image prompt and adding the word video. An image prompt describes a frozen moment; a video prompt describes change over time. Verbs are your main tool: pours, turns, accelerates, drifts, collapses. A solid video prompt covers five things — subject, one clear action, setting, camera behavior, and mood or lighting. If you already finished the prompt anatomy lesson, this is the same skeleton with a motion muscle attached.
- Subject: who or what the shot is about — one hero element, described concretely
- Action: exactly one verb phrase, e.g. slowly pours espresso into a glass cup
- Setting: where and when — sunlit kitchen at 7am, rainy Tokyo street at night
- Camera: static, slow dolly-in, handheld — one behavior (full menu in the camera language lesson)
- Mood and light: warm morning glow, harsh neon, soft overcast
The Honest Capability Map
| Models handle well | Models still struggle with |
|---|---|
| One clear action by one subject | Multi-step choreography (pours, THEN turns, THEN exits) |
| Atmosphere: weather, light, crowds as texture | Precise on-screen text and logos |
| Camera moves described in film language | Exact object counts (six dancers becomes roughly six) |
| Physics of everyday motion — water, cloth, momentum | Long dialog scenes and fine finger work |
Your first text-to-video prompt
A barista in a sunlit specialty coffee shop slowly pours steamed milk into a ceramic cup, forming a rosetta latte art pattern. Static camera, close-up on the cup, shallow depth of field, warm morning light through the window, gentle cafe ambience.
One subject, one action, one camera behavior. The milk pour is a physics showcase — fluid motion is where modern video models shine.
From here the path splits into two superpowers: animating a still image for maximum control, and speaking camera language so your shots stop looking like security footage. But first — generate something. Theory without renders is just reading.
Type one sentence of action and watch it move — your first clip takes about a minute. Generate your first video