Text to video: a clip from a single description

If animating a photo starts from a ready frame, then text-to-video creates a clip from scratch — you have nothing but a sentence, and out comes a moving scene. It's the most "magical" and at the same time the most temperamental generation mode: here the prompt decides everything.

How it works

You describe the scene in words — the network draws not one frame but a whole sequence, holding a single world between them: the same character, light, camera movement. In essence it's image generation stretched over time, plus an understanding of the physics of motion.

Because of this, text-to-video is more expensive and more temperamental than image-to-video: the model has to invent both the content and its movement at once. So short scenes (5–10 seconds) come out great, while a long coherent plot is still assembled from several clips.

Describe a scene in one sentence — get a clip. A hint for your first time: "a neon street at night in the rain, a slow forward dolly, reflections in the puddles, cinematic atmosphere". Video costs more than pictures: a clip is available after signing up and onboarding — which grant starter tokens.

What a good video prompt contains

An image prompt describes the frame. A video prompt also describes the movement and time. Keep five layers in mind:

1. Scene — what and where. "An old lighthouse on a rocky shore".
2. Movement in the frame — what's happening. "…waves crash on the rocks, gulls circle".
3. Camera — this is new and important. "…a slow dolly in", "an orbit", "a drone shot", "a static shot".
4. Light and time — "the setting sun, long shadows".
5. Style — "cinematic, like a film still", "3D animation", "documentary".

The main difference from a picture is the camera. It's the words about camera movement that turn a "living postcard" into a "film shot". Don't specify the camera and the model decides for itself, often badly.

Common beginner mistakes

• Too much action. "A person runs, jumps, turns and waves" in 5 seconds falls apart. One clear movement per clip.
• Text and captions in the frame. Still a weak spot of almost all models — letters "drift". Overlay text on the finished clip separately.
• Complex hands and crowds. A classic pain; the fewer of them in the frame, the more stable the result.
• Expecting a long plot. Think in "shots", not "scenes": assemble the clip from several short generations.

What's next

You can create a scene from scratch. A specific but very in-demand case is when there has to be a person in the frame who speaks. That's a separate genre with its own rules.

In the Twelver chat a video prompt is written like an ordinary message — the clip comes back in reply. Starter tokens for video are granted after signing up and onboarding.