How to Make a TikTok Video with AI: Step-by-Step Guide

To make a TikTok video with AI, you write a hook that lands in the first one or two seconds, generate vertical 9:16 clips from a short script with a video model like Seedance 2.5, add a voice-over or spoken dialogue, then put captions on and post it with TikTok's AI-generated label. That is the whole loop. The rest of this guide walks through each step with the settings, prompts and mistakes that make the difference between a clip people scroll past and one they watch to the end.

If you are new to AI video in general, start with our pillar guide on how to make an AI video. This article is specifically about the short-form, vertical, sound-on world of TikTok.

Step 1: Start with the idea and the hook

On TikTok the first one to two seconds decide everything. The viewer has not chosen your video; the feed chose it for them, and their thumb is already moving. So before you open any generator, write down two things:

1. The idea in one sentence. "A cat explains why Mondays are a conspiracy." "A strawberry finds out her best friend is dating her ex." "Three ways to style one white shirt." If you cannot say it in a sentence, the video will wander.
2. The hook shot. What is literally on screen in second one? A face looking straight into the camera mid-sentence, an object falling, a weird juxtaposition, a text question. Not a slow establishing shot of a city skyline.

Good hooks for AI video tend to be visual surprises the model is good at: impossible scale, a transformation, an unexpected character speaking, or a close-up that starts in the middle of the action.

Hook shot: extreme close-up of a tiny astronaut standing on a slice of toast,
he turns to camera and shouts {Houston, we have butter.} Vertical 9:16,
bright morning kitchen light, handheld feel, starts mid-action.

Step 2: Set up for 9:16 from the start

Generate vertical. Do not generate 16:9 and crop later: you lose resolution and the model composed the shot for a wide frame, so heads get cut and subjects drift out of the safe area.

Recommended setup for a TikTok AI video:

Setting	Recommendation	Why
Aspect ratio	9:16	Native TikTok format, full screen
Clip length	5–15 s per shot, up to 30 s for a single take	Short shots keep pace; long takes suit talking or one continuous scene
Resolution	720p	Fine on a phone screen, faster and cheaper than higher tiers
Audio	Native audio on	TikTok is sound-on; silent AI clips feel unfinished
Framing	Subject in the middle third	Captions and the TikTok UI cover the top and bottom

In the Framiq studio, Seedance 2.5 is the default video model and accepts 9:16 directly, from 4 to 30 seconds per clip in one generation, up to 720p.

Keep the TikTok interface in mind: the username, caption and buttons sit on the right and bottom edges. Ask the model to keep the subject centred with "centred framing, headroom above, space below for captions".

Step 3: Write a short script

Even a 20-second TikTok benefits from a script. It does not have to be fancy. A simple beat structure works:

• 0–2 s, hook: the surprising image or line.
• 2–15 s, payoff or story: the joke, the tip, the twist.
• Last 2–3 s, loop or call to action: an ending that makes people rewatch or comment.

For a story or a skit, write the dialogue with who says it and how. That matters later when you generate voices.

[0-2s] @Mia stares at her phone, shocked. Mia shouts: {He did NOT just text that.}
[2-10s] Close-up of the phone screen, a message reads "u up?". Mia rolls her eyes.
[10-18s] Mia types fast, smirking. Mia whispers: {Watch this.}
[18-22s] Cut to her friend reading the reply and bursting out laughing.

Stuck for an idea? The AI prompt generator button in the studio turns a short idea into a full prompt, which you can then trim. For more on prompt structure, see our AI video prompting guide.

Step 4: Create a recurring character

The TikTok accounts that grow with AI content almost always have a recurring character. People follow a person (or a fruit, or a cat), not a style. Recurring characters, cliffhangers and short episodes are a big part of why the AI fruit soap-opera trend took off; we break that down in AI fruit videos.

The problem is that a video model invents a new face every time unless you give it references. In Framiq you create a character from photos in the Characters tab, and optionally run a 3D scan, where the studio renders clean studio views of the character (face close-up, then full body, then profiles and back) and you approve each stage. Then you tag the character in any prompt with @Name and the model receives its references, so the face stays the same across shots and episodes.

If the character is based on a real person, it must be you or someone who has given you clear consent. No celebrities, no lookalikes of people who did not agree to it.

More detail: consistent AI characters and our guide to keeping a character consistent in AI video.

Step 5: Generate your clips

Now generate. With Seedance 2.5 you have two strategies:

• One long take (up to 30 s). Good for a talking-head monologue, a single continuous scene or a product demo. One generation, no stitching, consistent lighting.
• Several short shots (5–10 s each). Good for skits, tutorials and anything with cuts. Faster to iterate: if shot 3 is wrong, you redo shot 3 only.

A practical prompt for one shot:

@Mia sitting on her bed at night, vertical 9:16, phone in hand, lit by the
phone screen and a warm bedside lamp. Medium close-up, centred framing,
space below for captions. She looks up at the camera and says calmly:
{Okay, I need your opinion on this.} Soft room tone, distant traffic.
Handheld, slight natural camera shake, realistic skin texture.

Tips that save generations:

• One action per shot. "She picks up the phone, walks to the window, opens it and waves" is four shots pretending to be one.
• Name the camera. "Static tripod", "slow push-in", "handheld selfie angle". TikTok viewers are used to selfie framing; it reads as authentic.
• Generate 2–3 variations of the hook. The hook is worth the extra cost; the middle shots usually are not.
• Use start and end frames when a shot has to match the next one exactly.

Want to compare models? Seedance 2.5 is our strongest option for character consistency; Kling 3.0 and Veo 3.1 are also in the studio. See Seedance 2.5 for its full capabilities.

On cost: a 30-second 720p Seedance 2.5 clip costs about $7 of API cost, so most creators generate short shots while iterating and only go long once the idea works. Plan prices are on the pricing page.

Step 6: Add voice-over or dialogue

Sound is half of TikTok. You have three options:

1. Native audio from the video model. Seedance 2.5 generates sound with the video: ambience, effects and spoken lines written in the prompt. Inside a Seedance prompt, {…} marks dialogue, (…) music and <…> sound effects.
2. Automatic dialogue voices. If your character is linked to a cloned voice, turn on the Voices toggle. Framiq voices each line written as @Mia says: {…} or Mia: with that character's voice and gives it to Seedance as reference audio, while the rest of the soundtrack is still generated. It costs extra and is off by default.
3. A separate voice-over. For faceless TikToks (facts, stories, tutorials), generate the narration in the Voice tab with ElevenLabs models and lay it under your clips in an editor.

Audio tags make a voice-over sound performed rather than read:

[whispers] Nobody tells you this about renting in Paris. [pause]
The landlord can't keep your deposit for that. [laughs] Yes, really.

Voice cloning on Eleven v4 works from about 10 seconds of clean speech. As with faces: clone your own voice or one you have permission to use. Our guide on generating an AI voice goes deeper.

Step 7: Add captions

Some viewers scroll with the sound off, others skim faster than they listen, and on-screen text also reinforces the hook. Add captions in TikTok's own editor or in a mobile editor like CapCut; both can auto-transcribe speech.

Caption rules that work:

• Short lines, two to five words, centred in the middle third.
• Big, high-contrast text with an outline or shadow.
• A text hook in the first frame that matches the visual hook ("POV: your landlord finds out").
• Do not ask the video model to render captions inside the video. Text in AI video is still unreliable, and you cannot edit it afterwards.

Step 8: Assemble the edit

AI generates the shots; a regular editor puts them together. Import your clips into CapCut, Premiere, DaVinci Resolve or TikTok's editor and:

• Trim the first frames of each clip so every shot starts on movement.
• Cut on action, and keep most shots under 3–4 seconds unless it is a single take.
• Balance levels: voice on top, music low.
• Add a trending sound if it fits, at a low volume under your own audio.
• Export vertical 1080×1920 or 720×1280.

If you need a longer piece, the Extra Long feature on the Pro plan chains Seedance clips into one video of up to 5 minutes. For TikTok, 30 seconds in one generation usually covers it. See long AI videos.

Step 9: Post, and label it as AI

When you post, write a caption that adds context or asks a question, use a few relevant hashtags, and pick a cover frame that shows your character's face or the hook.

Turn on TikTok's "AI-generated" label. TikTok requires creators to label realistic AI-generated content, and Instagram has similar rules. Labelling is not a penalty: unlabelled realistic AI content risks removal, and audiences react better to honesty than to being fooled. Never post AI content that makes a real person appear to say or do something they did not.

Post consistently: a numbered series that ends on cliffhangers gives people a reason to follow.

Step 10: Measure what matters

Views alone tell you little. In TikTok analytics, look at:

• Average watch time and the retention curve. A sharp drop at second 2 means the hook failed. A drop in the middle means the pacing failed.
• Completion rate. For short videos, the share of people watching to the end is one of the clearest quality signals.
• Rewatches and shares. Shares mean the idea worked, not just the visuals.
• Follows per video. This tells you whether your character and series are building an audience.

Then iterate: keep the character, change the hook. Generating three hook variations of the same video and posting them on different days is one of the cheapest experiments in short-form.

For broader strategy across platforms, read AI video for social media and our look at AI TikTok trends.

Common mistakes and quick fixes

Mistake	Fix
Generating 16:9 and cropping	Generate 9:16 from the start
Face changes between shots	Create a character and tag it with @Name
Silent clips	Turn on native audio or add a voice-over
Text rendered by the model	Add captions in your editor instead
Slow cinematic intro	Start mid-action, hook in the first second
Too many actions in one prompt	One action per shot
No AI label	Toggle TikTok's AI-generated label when posting

FAQ

Can I make a TikTok video with AI for free?

You can start for free: a Framiq account comes with free credits and the Free plan charges the standard rate, with a watermark. For regular posting without a watermark, the Pro plan at 30/month charges models at API cost. See pricing.

What is the best AI for TikTok videos?

It depends on the format. For recurring characters and up to 30-second takes with native audio, Seedance 2.5 is our strongest option. Kling 3.0 and Veo 3.1 are good alternatives for certain styles. Our Seedance vs Kling comparison covers the trade-offs.

How long should an AI TikTok video be?

Shorter is safer while you are testing: 10 to 30 seconds. With Seedance 2.5 you can generate up to 30 seconds in a single clip, which covers most TikToks without stitching.

Do I have to label AI videos on TikTok?

Yes, for realistic AI-generated content. TikTok asks creators to use its AI-generated label, and unlabelled realistic content can be removed. It is good practice to label anything clearly AI-made.

Can I make faceless TikToks with AI?

Yes. Combine AI-generated b-roll or a non-human character (an animal, an object, a cartoon) with a voice-over from the Voice tab and captions added in your editor. Faceless accounts live or die by the script and the voice, so spend your time there.

Can AI keep the same character across TikTok episodes?

Yes, if you give the model references. Create the character once from photos, optionally with a 3D scan, then tag it with @Name in every prompt so each episode uses the same face.

Try it

Pick one idea, write a two-second hook and generate it in 9:16 in the Framiq studio. Make three versions of the hook, keep the best one, and build the rest of the video around it. Your first AI TikTok will teach you more than any guide.