Text to video: the complete guide

Text to video is exactly what it sounds like: you write a description, and a model generates a video clip from it. What is less obvious is how much of the result is determined by things other than the words — duration, resolution, aspect ratio, and above all which model you picked.

How a generation actually goes

1. You write a prompt and set parameters — duration, resolution, aspect ratio, whether to generate audio.
2. The request is queued and the model renders the clip. Most finish in well under a minute; longer and higher-resolution clips take proportionally longer.
3. You get back a video file — MP4, or MOV on the models that offer it — together with the prompt and parameters that produced it, so a good result can be reproduced.

Generation is stochastic. The same prompt run twice gives two different clips. That is not a defect to work around; it is the reason generating three takes and picking one is the normal workflow rather than a sign something went wrong.

How to structure a prompt

A prompt that works reads like a shot description from a treatment, not like a search query. Four components, roughly in this order:

• Subject — who or what is in frame, and what they are doing. A boxy 1980s four-wheel-drive rally car comes through a long left-hander.
• Setting — where, and when. A night stage of a gravel rally, in the pines.
• Camera — the shot itself. Low tracking shot from the outside of the corner, held.
• Light and mood — headlights raking the trunks, gravel spray backlit by the follow car.

Things that reliably help

• Verbs over adjectives. Dust lifting off the boards is worth more than dusty, atmospheric.
• One camera move per clip. A prompt asking for a push-in, a whip pan and a crane in three seconds gets none of them cleanly.
• Real lighting vocabulary. Backlit, hard key from the left, practical lamps, golden hour, overcast — these are understood precisely.
• A stated duration in the parameters, not in the prompt. Writing "a 30-second clip" in the prompt does nothing; the duration parameter does.

Things that reliably do not

• Stacking quality words — 4k, ultra realistic, masterpiece, best quality. Resolution is a parameter. The rest is noise.
• Negative instructions in the positive prompt. No text often produces text. Use the negative prompt field on models that have one.
• Twenty specific requirements. Models satisfy roughly the first handful and improvise the rest.

The parameters that change the result

Duration

The single biggest driver of both cost and quality. Longer clips cost more because they are more compute, and they drift more because each frame is conditioned on the last. Ask for the length you need and no more. Models differ enormously here: some generate one-second cuts, Seedance 2.5 runs long single takes, and Flux 3 Video has the widest range of the registry models.

Resolution

Draft at the lowest resolution the model offers and deliver at the highest your plan allows. Composition, motion and timing are all judgeable at 480p, and re-running a working prompt at 1080p is cheaper than discovering at 1080p that the composition was wrong.

Aspect ratio

Generate natively in the ratio you will publish in. Cropping 16:9 down to 9:16 throws away most of the frame and usually the subject with it. Note that models which accept a reference image generally follow the image's shape and ignore this parameter — crop the reference first.

Audio

Most current video models generate a soundtrack together with the picture rather than as a separate pass, which is why the sound lines up with the motion. If you are going to replace the audio in an edit anyway, turning it off is usually the cheaper generation.

Choosing a model

There is no overall best one. There are models that win specific jobs:

• Long single takes, consistent subject — Seedance 2.5.
• Physical plausibility, zero data retention — Veo 3.1, with Veo 3.1 Fast for iteration.
• Locked-off product shots at high resolution — Seedance 2.0.
• Very short cuts for fast-cut edits — Grok Imagine Video.
• Wide cinematic framing with a cheap draft mode — Flux 3 Video.
• Terse prompts you would rather not write out — Hailuo 2.3, which expands them for you.
• Frame-rate control and reproducible seeds — P-Video.
• Animating a still you already have — HappyHorse 1.1 or Wan 2.7.

The full catalog lists every one with its real specifications and example prompts.

What it costs

Generation is billed in credits, and the cost of a clip scales with the compute it consumes — model, resolution and duration. The exact figure is shown before you submit, so nothing is a surprise. A new account gets 50 credits with no credit card; Starter is $9.99/month with 1,000 credits, watermark-free export and a commercial licence. Full detail is on pricing.

A first session that will go well

1. Pick one shot you actually want. Not a montage — one shot.
2. Write it as subject, setting, camera, light. Four sentences at most.
3. Generate it at the lowest resolution and shortest duration on a cheap model.
4. Read what came back against what you asked for, and change one thing.
5. Repeat twice. Then re-run the working prompt on the model and settings you want to deliver from.