How to make a video with AI: the complete guide

What an AI video model is

A video model is software that makes a new clip from what you give it: a description in words, a picture or a few sample images. Each model is built by a different maker and is good at different things. Some also create speech and sound, some reach 30 seconds, and some keep the same character across several clips.

The clips are short: here between 2 and 30 seconds, depending on the model. A longer video is put together from several clips in an editing app. In the studio you pick the model by its name, and see the exact price before every generation.

The five steps

1. Choose a model. By what the clip needs: speech and sound, length, a recurring character or a budget. The guide how to choose an AI video model goes through this need by need
2. Choose where to start. Text to video, Image to video, or References, which works from sample images. Each is explained below
3. Write a prompt. What you see, what happens, where, how the camera moves and how it sounds
4. Set the options. Duration, resolution, aspect ratio and sound. Every change updates the price at once
5. See the price and create. The exact price shows before you send, and the result is saved to your library when it is ready

All the steps are on one screen in the studio, and every setting has an i that explains what it does. While a clip renders you can leave the page, and the result will wait in your library.

Three ways to start a clip

Text to video

Write what you want to see, and the model creates everything: the characters, the place, the light and the motion. A good fit when you have no footage and want full freedom.

Image to video

Upload a picture, and it becomes the first frame of the clip. The prompt says what happens from there. A good fit for animating a photo, a product or an illustration. Some models also take an end image, and the model fills in the motion between the two pictures. An end image works only together with a start image.

Upload a sharp picture: the maker of Kling 2.6 notes that a sharper source photo gives a better video. These are some of the models that take both a start image and an end image:

• Veo 3.1
• Veo 3.1 Lite
• Kling 3.0
• Wan 3.0
• MiniMax H3
• LTX-2.3 Fast
• Ray 3.2

References: from sample images

Attach images of a character, a product or a style, and the model builds a new scene in which they stay alike. This is how you keep the same character or product across several clips. Veo 3.1 takes up to 3 images, Kling O3 up to 7, Seedance 2.0 up to 9 and Seedance 2.5 up to 30.

And an existing clip

Some models work with a clip you already have. Wan 2.7 Video Edit edits a clip from an instruction in words, such as a new style or a change of clothes, and Wan 2.7 can continue an existing clip. In Kling O3 the sample clip is used to create a new shot, such as the next shot, not to edit the clip you uploaded.

What each setting does

Settings change from model to model, and the studio shows each model only what it accepts. These are the common ones:

• Duration. How long the clip is, in seconds. In most models a longer clip costs more, and in some the price is the same for every length
• Resolution. 720p is enough for social media, 1080p fits most uses and 4K is for a large display. In some models the higher resolutions are upscaled from a lower resolution result
• Aspect ratio. 16:9 is wide, for a computer screen. 9:16 is tall, for phones, Reels and stories. 1:1 is square
• Generate audio. In models that create sound, this switch decides whether the clip comes with ambience, music or speech. In some models it changes the price, and the price updates at once
• Frames per second. In models that let you choose, 24 or 25 look cinematic, and 48 or 50 are smoother in fast motion
• Negative prompt. A short list of things you do not want in the result, such as text, blur or people in the background
• Prompt enhancement. In models that offer it, the model expands a short prompt before generating. When you want exact control, turn it off

How to write a prompt the model understands

The official prompting guide for Veo 3.1 suggests building every prompt from five parts, in this order. You can use the same structure in other models too, because it answers everything a video model needs to know:

1. Camera. The kind of shot and the camera movement, such as a close up, or a camera slowly moving in
2. Subject. Who or what is at the center of the picture
3. Action. What the subject is doing
4. Context. Where it happens and what is in the background
5. Style and mood. The overall look, the mood and the lighting

For a clip with sound, the same guide recommends putting every line a character says inside quotation marks, and describing effects and ambience separately. This is what a prompt built this way looks like:

Close-up, slow push in. A young barista in a small café pours latte art, morning light through the window, warm and calm mood. She looks up and says: "Your coffee is ready." Ambient noise: soft café chatter

Camera moves you can ask for

You can ask for camera moves in common filmmaking terms: push in, pull out, pan, a tracking shot that follows the character, or a static shot where the camera does not move. In the LTX models you can also pick a preset move from a list, such as moving in, moving out or rising.

A few more rules that help

• One action per clip. A few seconds hold one moment, not a whole story
• Write in English. The maker of Veo 3.1 and Gemini Omni Flash confirms full support only for English, and the maker of Wan confirms Chinese and English
• What you do not want goes in its own field. Write what must be missing from the result in the "Negative prompt" field, in models that have it
• Results close to each other. When a result works, keep its prompt and change one detail at a time, with the same start image or the same reference images

Sound and speech in a clip

Most video models here create sound together with the picture. Veo 3.1 creates speech, effects and ambience, in Kling 3.0 sound is an option you turn on, and Wan 3.0 creates dialogue, effects and background music. Models such as Ray 3.2 make clips without sound, and they fit when you add music or a voiceover in editing.

To get a particular sound, describe it in the prompt: what you hear in the background, which effects, and exactly what is said. For Gemini Omni Flash 1.1, for example, the maker says the model adds a matching soundtrack by default, and that you can describe the sound you want in the prompt.

Hebrew speech: no video model here officially supports it. For a clip with Hebrew narration, create the voiceover separately and join them in editing, as explained in the guide Hebrew in AI models.

What it costs

The price depends on the model, the length, the resolution and sometimes the sound. This is what a typical first clip costs right now:

The exact price shows before every generation, and that is what leaves your balance. A generation that fails gives all its credits back to your balance, automatically. The full explanation of prices is in the guide what creating with AI costs, and of refunds in the guide a failed or stuck generation.

Common mistakes and how to avoid them

• A whole story in one clip. A clip of a few seconds fits one action. Split a story into several clips and join them in editing
• 4K for a draft. Test the idea at a low resolution and in a budget version, and move to high quality once the prompt works
• The wrong aspect ratio. For Reels and TikTok choose 9:16 when you create, because cropping a wide clip loses part of the picture
• A blurry source photo. A sharp picture gives a better clip
• An end image without a start image. An end image works only together with a start image
• Expecting Hebrew speech. No video model officially supports it. Create the Hebrew voiceover separately
• A file that is too large. An upload must be up to 4 MB

Starting with a structured prompt, a budget version and a short clip gets you to the final result faster and with fewer credits.

Questions and answers