How to Make an AI Video: Step-by-Step Guide (2026)

To make an AI video, you pick a video model, describe the shot you want in a prompt (or upload an image to animate), choose a format and a duration, and generate. A good first clip takes a few minutes; a good final clip usually takes three to six iterations where you change one thing at a time. This guide walks through every step of how to make an AI video in 2026 — from choosing between text-to-video and image-to-video to audio, exporting and what it really costs.

The two ways to make an AI video

Every AI video starts from one of two inputs. Knowing which one you need saves most of the wasted generations.

Text-to-video

You write a description and the model invents everything: the people, the place, the light, the camera move. It is the fastest way to explore an idea and the best choice when you do not yet know exactly what the frame should look like.

The trade-off: you control less. Two generations of the same prompt will give you two different actresses, two different kitchens, two different colour grades.

Image-to-video

You upload a picture and the model animates it. The image becomes the first frame (or a reference for the look), so the composition, the face and the product are already locked. Use it when:

• you need a specific person or product to appear exactly as they are,
• you already made a great still with an image model and want it to move,
• you want several shots that look like they belong together.

Most professional workflows combine both: generate a still with an image model, fix it until it is right, then animate it. If you are starting from a photo you already own, our French guide on animating a photo with AI goes deeper into start and end frames.

Step 1: Choose the right AI video model

There is no single "best" model. Each one has a strength, and picking the right one is half the work.

Model	Best for	Watch out for
Seedance 2.5 (ByteDance)	Long single shots (4–30 s in one generation), recurring characters, dialogue with native audio, editing or extending an existing clip	Runs up to 720p in Framiq; a reference video adds to the cost
Kling 3.0 / Kling Omni (o3) (Kuaishou)	Realistic human motion, physical actions (dance, sport, fabric, water)	Shorter clips than Seedance 2.5
Veo 3.1 / 3.1 Fast / 3.1 Lite (Google)	Cinematic framing, understands film vocabulary well, native audio	Short clips; the Fast and Lite versions trade quality for price
Hailuo 2.3 (MiniMax)	Smooth everyday motion at a lower cost, lifestyle content	Less control over complex scenes
Wan 2.7 / 2.6 (Alibaba)	Volume work and stylised content on a budget	Shorter clips
Grok Imagine Video 1.5 (xAI)	Quick creative ideas, several reference images	Newer model, test before committing a campaign to it

A simple rule of thumb:

• Story, characters, dialogue, anything over 10 seconds → Seedance 2.5. It is the default model in the Framiq studio because it holds a character's face better than anything else we offer and does up to 30 seconds in one shot.
• A body moving in a physically convincing way → Kling 3.0.
• A "film look" shot with precise camera language → Veo 3.1.
• Lots of cheap variations to test an idea → Seedance 2.0 Fast or Mini, Hailuo, Wan or LTX Video, then re-render the winner on a premium model.

Side note on Kling 4.0: Kuaishou reportedly launched it in mid-October 2026, with Kling 4.0 Flash in early access from late September. It is not available in Framiq yet; Kling 3.0, Kling Omni and Seedance 2.5 cover the same use cases today. For a head-to-head of the two most popular families, read Seedance vs Kling.

Step 2: Write a prompt the model can actually follow

The prompt is where most AI videos fail. A video prompt is not a list of keywords — it is a short shot description, written like a director talking to a camera operator.

Use this structure, in this order:

1. Subject — who or what, with the details that matter (age, clothing, material).
2. Action — what happens, in the order it happens. One main action per shot.
3. Setting — where, when, weather.
4. Camera — shot size and movement (close-up, slow dolly in, handheld, drone).
5. Light and style — golden hour, neon, overcast, 35mm film, documentary.
6. Audio (if the model generates sound) — ambience, music, dialogue.

Here is a weak prompt and a strong one for the same idea:

a woman drinking coffee in a cafe, cinematic, 4k, beautiful

Medium close-up of a woman in her thirties with short curly hair and a
mustard sweater, sitting by the window of a small Paris café. She wraps
both hands around a white cup, takes a slow sip, then smiles at something
outside. Rainy morning, soft grey daylight, warm practical lights inside.
Slow push-in, shallow depth of field, natural colour grade.
Audio: rain on the glass, quiet café chatter, a cup set down on a saucer.

The second version gives the model a sequence (wrap hands → sip → smile), a camera move, and a light source. "Cinematic, 4k, beautiful" gives it nothing it can act on.

A few rules that save a lot of credits:

• One shot, one idea. If you need a cut, make it a second clip.
• Describe motion with verbs and speed: "slowly turns her head", "the camera rises over the rooftops".
• Name the camera move once. "Slow dolly in" is enough; three camera moves in one prompt produce a wobbly mess.
• Avoid negations when you can. "Empty street" works better than "no people".

If you would rather start from a one-line idea, the AI prompt generator button in the studio turns it into a full prompt you can then edit. For a deeper dive, see the AI video prompting guide.

Step 3: Add references (faces, products, style)

References are images (and, on some models, videos or audio) that you attach alongside the prompt. They are the difference between "a woman" and "this woman".

Seedance 2.5 accepts up to 30 reference images in Framiq, plus a reference video and reference audio. Typical uses:

• A character — several photos of the same person from different angles.
• A product — the packaging from the front and the side, so the label is not reinvented.
• A location or style — a mood frame you want the shot to match.

If the same person appears across many videos, create a Character in the studio. You upload their photos once (or import them from an Instagram, TikTok or X profile), optionally run a 3D scan that renders clean studio views you approve step by step, then tag them in any prompt with @Name. Every image and video model then receives the right references, which is how you keep the same face from shot to shot. More on that in consistent AI characters.

One firm rule: only use a real person's face if it is your own or you have their consent. No celebrities, no colleagues "as a joke".

Step 4: Choose duration and aspect ratio

Pick the format before you generate — reframing afterwards crops your subject.

Where it goes	Aspect ratio	Typical length
TikTok, Reels, Shorts	9:16	6–30 s
YouTube, website hero, presentations	16:9	5–30 s per shot
Instagram feed, ads, marketplaces	1:1 or 4:5	5–15 s

Longer is not always better. A 30-second single shot is great for a continuous scene — a walk-and-talk, a product demo, a monologue — but for fast-paced edits, three 8-second shots usually look more professional than one long take.

Seedance 2.5 generates 4 to 30 seconds in one go, so most short-form videos need only one generation per shot. If you need something longer, the Pro plan's Extra Long mode chains several Seedance clips with continuity and stitches them into one video of up to 5 minutes. It is experimental, so plan for a retry or two — see long AI videos.

Step 5: Add sound and dialogue

Several models now generate audio natively: Seedance 2.5, Veo 3.1 and Happy Horse (which also does multilingual lip-sync) produce ambience, effects and voices together with the picture.

With Seedance you can write sound directly in the prompt using its own notation:

• {…} for dialogue,
• (…) for music,
• <…> for sound effects.

@Lina sits on the edge of the bed and looks at her phone, then up at the
camera. @Lina says softly: {I wasn't going to call you. But here we are.}
(slow piano, barely audible) <a car passes outside>
Static medium shot, warm bedside lamp, night.

Always attribute a line to a speaker and give a delivery word (softly, shouts, calmly). If your character has a cloned voice linked in the Voice tab, switching on the Voices toggle makes Framiq perform each line with that voice and hand it to Seedance as reference audio — the rest of the soundtrack is still generated. It costs extra and is off by default.

For voice-overs, music beds or sound effects made separately, the Voice tab runs ElevenLabs models (Eleven v4, v4 Turbo, v3, multilingual v2) with audio tags such as [whispers], [laughs] or [pause].

Step 6: Generate, review, iterate

Your first generation is a draft. Watch it three times:

1. Once for the overall feel — is this the right shot at all?
2. Once for motion — hands, feet, faces, objects passing behind each other.
3. Once for details — text, logos, reflections, the last second of the clip.

Then change one variable and regenerate. If the motion is wrong, rewrite the action; if the face drifts, add references; if the look is off, change the light line. Changing five things at once tells you nothing about what fixed it.

Two time-savers:

• Test cheap, finish premium. Explore with a fast model, then re-render the winning prompt on Seedance 2.5, Kling or Veo.
• Extend instead of restarting. If a Seedance clip is great but too short, use Extend on it in the gallery: it uses the clip as a reference and continues the shot.

Common mistakes and how to fix them

Problem	Likely cause	Fix
The face changes mid-clip	No reference, or too many conflicting ones	Use a Character with consistent photos, tag it with @Name
Motion is frozen or too slow	Action described vaguely	Use concrete verbs and a sequence: "walks, stops, turns"
Camera shakes everywhere	Several camera moves in one prompt	Keep a single, named camera move
Product label is garbled	Model reinventing the packaging	Attach clear product photos as references
Sound covers the dialogue	Too many audio cues	Keep one ambience line and one music line

Step 7: Edit, export and publish

Need to change an outfit, the background or the lighting of a clip you like? Attach the existing video to Seedance 2.5 as a reference and describe the change; the model keeps the motion and composition and re-renders. Our guide on editing a video with AI covers this in detail.

When you are happy, download the file from the gallery and cut it in your usual editor (CapCut, Premiere, DaVinci Resolve) if you need titles, captions or several shots. Keep your clips organised with Project mode if you work for several clients or on a series.

Before publishing: TikTok and Instagram require you to label realistic AI-generated content (TikTok has an "AI-generated" label in the post settings). Use it — it is the platform rule, and audiences trust accounts that are upfront. For platform-specific tips, see AI video for social media.

How much does an AI video cost?

AI video is billed by what you generate: the model, the duration, the resolution and any reference video you attach. In Framiq:

• Free plan — models are sold at the standard rate, videos carry a watermark, and a new account comes with free credits to start.
• Pro plan (30/month) — every model at API cost, no watermark, a 200 GB library and Extra Long videos.

The balance is prepaid (you top up any amount) and failed generations are refunded. To give you an order of magnitude, a 30-second 720p Seedance 2.5 clip costs about $7 of API cost; shorter clips and fast models cost far less. All per-model prices are listed on the pricing page.

FAQ

Can I make an AI video for free?

You can start for free: a Framiq account comes with free credits, enough to try a few short generations. Free-plan videos carry a watermark. No serious AI video tool is unlimited and free, because each generation costs real compute.

How long can an AI video be?

It depends on the model. Seedance 2.5 does up to 30 seconds in a single generation; many other models stop around 5–10 seconds. For longer videos, you either edit several clips together or use a chaining feature such as Framiq's Extra Long mode (up to 5 minutes, Pro plan).

Which AI is best for making videos?

For characters, dialogue and longer shots, Seedance 2.5. For realistic body motion, Kling 3.0. For a cinematic look, Veo 3.1. The best approach is to test the same prompt on two models and keep the winner.

Can AI make a video from a photo?

Yes — that is image-to-video. You upload the photo, describe the motion ("she turns her head and smiles, hair moving in the wind") and the model animates it. Keep the motion subtle for portraits, and only animate people who have agreed to it.

Do AI videos have sound?

Models such as Seedance 2.5, Veo 3.1 and Happy Horse generate audio natively — ambience, effects and dialogue. Others are silent, and you add a voice-over or music afterwards.

Can I use AI videos commercially?

In general yes, for content you generated, as long as you respect each provider's terms and do not use someone's likeness, a brand or copyrighted characters without permission. For ads, read AI UGC ads.

Try it

The fastest way to learn is to make one. Open the Framiq studio, keep Seedance 2.5 selected, paste the café prompt above, and generate. Then change one line — the light, the camera, the action — and watch what moves. Ten minutes of that teaches more than any guide.