How do I generate a video with AI? A step-by-step guide

"Generating a video with AI" means two different jobs, and most confusion comes from mixing them up. A clip generator turns a prompt or an image into one shot that lasts a few seconds. A video generator turns an idea or a script into a finished video: several shots, a voice, captions and music, cut together.

The underlying models are the same either way. A video generator calls a clip model once per scene, then does the work around it: writing the script, splitting it into scenes, generating a voice, timing captions and editing. If you want to learn how prompting works, start with single clips: they are cheap and fast. Move to a full video when you know what you want to say.

A video model frames the shot for the aspect ratio you give it. Cropping a 16:9 clip into a 9:16 vertical keeps only 32% of the width, so subjects get cut in half and the composition falls apart. Decide where the video will be posted, then set the ratio before the first generation.

Then decide the length. A single AI clip lasts between 4 and 30 seconds depending on the model, so a 30-second post is usually 6 to 10 clips of 3 to 5 seconds each, one per sentence of the script. Short clips also fail less: the longer a clip runs, the more frames the model has to invent and the more the subject drifts.

Video models respond to what a camera could see. They do not understand intentions ("she misses her brother"), but they understand faces, objects, motion, light and lenses. A reliable prompt answers five questions in two to four sentences.

Rules that save the most retries

• One action per clip. A 5-second clip holds one clear action. "He opens the door, walks in and sits down" becomes three clips.
• Name the camera move once. "Slow push-in", "tracking shot", "static wide shot", "aerial drone shot". Two moves in one prompt usually means neither happens.
• Describe the end of the shot for longer clips. For 10 seconds and more, say where the action ends ("...and stops at the edge of the water"). Premium models also accept time ranges such as "0-5 seconds: ... 5-10 seconds: ...".
• Keep words out of the frame. Models draw letters as shapes, so signs and labels come out as gibberish. Add text later as captions or overlays.
• Do not name real people, brands or characters. Describe the look instead ("a tall man in a grey suit", "a red sports car"). Named people and protected characters are the main reason providers refuse a generation (see step 5).
• Write the style once, at the end. Mixing "photorealistic" and "anime" in the same prompt produces neither.

Copy a prompt that already works

These four prompts follow the formula. Each one opens in Revid with the prompt and settings filled in, on the cheapest tier, so you can test them before you spend more.

Build your own prompt

Fill in the five parts and the builder writes the prompt in the order models read best. Copy it, or open it in Revid.

Revid's view is that the prompt matters more than the model: the leading models are sold through APIs to every video tool, so two tools that send the same prompt to the same model get comparable clips. Models still differ in speed, length, price and what they are best at. These are the ones available on Revid in October 2026.

How long should each clip be?

Pick the shortest length that holds the action. Cost is billed per second, and waiting grows with length. For social video, 5 seconds per shot is the default that works: long enough for one action, short enough to keep the pace up. Use 8 to 15 seconds for slow, ambient shots (a landscape, a product orbit) and Seedance 2.5's 30 seconds only when you need one continuous take.

What it costs, with real numbers

• A 30-second video made only of Pro AI clips uses 300 credits (6 clips × 50).
• The same 30 seconds on the Low tier uses 90 credits.
• A 30-second video illustrated with Pro AI images and a voice-over costs about 40 to 80 credits. Switching only the 2 or 3 key shots to AI video keeps most of that saving.
• Revid's Growth plan includes 2,000 credits a month: about 200 seconds of Pro AI video, or 660 seconds on the Low tier.

Watch each clip twice, once for the whole and once frame by frame on the parts that usually break. Then change one thing in the prompt and generate again. If you change three things at once, you will not know which one fixed it.

1. 1Identity: does the subject keep the same face, clothes and colors from the first frame to the last?
2. 2Hands and bodies: count the fingers, check the joints, check that limbs do not merge.
3. 3Motion: is it physically plausible, at the speed you asked for?
4. 4Camera: did it do the move you named, and only that one?
5. 5Text and logos: anything written in the frame is probably wrong.
6. 6The last frame: does it end in a state that cuts well to the next shot?

Fix what breaks

In other words, when a generation fails, the content of the prompt or the image is usually the problem, not the service. Rewording is faster than retrying the same request. The full breakdown by model is in Revid's AI video model index.

A clip is a shot, not a post. People keep watching a short video because of what it says and how fast it gets there: a hook in the first two seconds, one idea per sentence, and a reason to stay until the end. This is the assembly order that works:

1. 1Write the script as narration. One sentence per shot. The first sentence is the hook: a question, a surprising number, or a promise.
2. 2Split it into scenes. One scene per sentence, 2 to 5 seconds each.
3. 3Generate one visual per scene, reusing the same style sentence so the video looks like one piece.
4. 4Add the voice-over. An AI voice, in one of 70+ languages on Revid, or your own cloned voice.
5. 5Add captions synced to the voice. Many people watch with the sound off, and captions carry the words.
6. 6Add music under the voice, low enough that every word stays clear.
7. 7Export in 1080p. Use 30 fps for narrated videos and 60 fps for fast action. Revid exports in 720p, 1080p or 4K.

Doing this by hand in an editor takes an afternoon per video. Prompt to Video does all of it from one sentence: it writes the script, picks a visual per scene, generates the voice and captions, and gives you a timeline to edit. These three videos were made that way, each from the one-line prompt shown under it.

Notice what the prompts do not contain: no camera directions, no scene list, no style keywords. A video generator writes those for you. When you generate single clips yourself (step 2), you are the one writing them.

Export at the size from step 1. A 4K export of clips generated at a lower resolution makes the file heavier without adding much detail, so pick 4K only when the clips were generated in 4K. Then check two things before you post:

• The first two seconds. Watch them with the sound off. If they do not make you want the next two, change the hook, not the visuals.
• The AI label. TikTok and YouTube ask creators to disclose realistic AI-generated content, such as a realistic person saying or doing something they did not. Use the platform's label when it applies. Obvious animation and fantasy usually does not need it, but the rules are each platform's to set.

Revid publishes directly to TikTok, Instagram and YouTube from the same place, or on a schedule, and shows the views and likes of each post afterwards, so you can see which prompts and styles actually work for your audience.

You can try it for free, but no serious AI video generator is free without limits. Every clip runs a large model on a GPU for seconds to minutes, so each one has a real cost. In practice, "free" means a small allowance of credits or clips, often with a watermark, a lower resolution or a slower queue.

On Revid, signing up is free, and new accounts usually get free credits to try video generation. A free account can also search the library of 2.77 million viral videos and posts and run up to 60 free deep analyses a day. Regular creation uses the monthly credits of a paid plan, from $39 a month ($32 billed yearly) for 2,000 credits.

How to get the most out of free credits

• Generate on the cheapest tier (15 credits per 5 seconds) until the prompt is right.
• Generate 5-second clips, not 10: half the cost, and they fail less.
• Start from an image when you can: on Revid it is 2.2 times faster on the fast tier, and the first frame is exactly what you want.
• Check the credit estimate before you click. Revid shows the cost of every generation and every video before it starts.

Frequently asked questions