How Do You Make an AI Video From Scratch in 2026? (Photo, Prompt or Template)
Choose a starting point, write a prompt that describes motion, pick a model by budget, add voice and captions, fix the common failures, and export for each platform.
Making an AI video from scratch starts with one of three inputs: a text prompt, a photo or a template. Pick the one that matches what you already have, write a prompt that describes motion rather than appearance, pick a model by budget and length, then add voice, captions and music. The first draft should be short and cheap; the finished clip is a second render, not the first one.
Updated October 2026.
Pick the start: text, photo or template
The starting point decides how much control you have. Each of the three is the right choice for a different situation.
| Start | Best when | What you give up |
|---|---|---|
| Text prompt | You have an idea but no footage | Exact control over the subject and look |
| Photo | You have a real product, person or place | Freedom to invent a different scene |
| Template or effect | You want a known format quickly | Originality and a custom prompt |
Text is the most flexible and the least predictable. A photo anchors the result: the subject, the colour and the composition stay close to the source, which is why it is the right choice for a product or a person who must be recognisable. A template skips the prompt entirely and gives you a repeatable effect, which is useful when the format matters more than the content.
If you are unsure, start from a photo. A good photo narrows the problem and gives the model something concrete to animate. The trade-offs between the two main routes are covered in text to video vs image to video.
Write the prompt (subject, action, camera, light)
A working prompt has four parts: the subject, the action, the camera and the light. Leave one out and the model fills the gap with a guess.
- Subject. Who or what is in the frame, with the one detail that matters: "a woman in a green coat", "a ceramic mug with a chipped handle".
- Action. What happens during the clip. One action per shot: she turns, the steam rises, the car pulls away.
- Camera. One move: a slow push in, a lateral track, a static frame. Two camera moves in one prompt usually produce neither.
- Light. The quality and direction: "soft window light from the left", "low sun behind the subject".
Put the subject and action first. Models tend to weight the opening of a prompt more heavily, and a prompt that starts with a style note can produce a beautiful frame with the wrong content. If a result ignores half the prompt, cut the prompt in half and keep the strongest idea.
Describe motion, not a still. "A red car parked on a street" gives the model nothing to animate; "the red car pulls away from the kerb, camera static, late afternoon light" gives it a beginning and an end. The guide to making AI videos look real covers the details that sell a generated shot.
Pick the model by budget
Models differ in length, resolution, audio and price, and the right pick depends on the shot rather than on a ranking. Start with what the clip needs to do.
| Need | Model to start with | Why |
|---|---|---|
| A cheap draft | Seedance 2 Fast | Quick iteration at a low cost, shown before you generate |
| Native audio and realism | Veo 3.1 | Sound arrives with the picture, in Fast and Quality tiers |
| Faces and controlled motion | Kling 3.0 | Strong face handling and first and last frame control |
| A long single take | Seedance 2.5 or Wan 3.0 | Clips up to 30 seconds |
Seedance 2 is a reliable default for general work: 4 to 15 seconds, six aspect ratios including 9:16 and 16:9, and native audio included. Its per-second cost is about 7.75 credits at 480p, 16.5 at 720p, 40.8 at 1080p and 83.2 at 4K, so a 4-second 4K clip is 333 credits. The same math on a draft resolution keeps a first test cheap.
Draft on the cheapest model that answers the question, then move the winner to the model that suits delivery. The price is shown before every generation, and failed generations are refunded. If you are buying credits, one-time packs never expire, which makes a multi-day project easier to budget.
Add voice, captions and music
A generated clip is often silent or has only the ambience the model invented. Three additions turn it into a finished video.
- Voice. Generate a voice-over on the text to speech tab, or use the native audio on a model such as Veo 3.1 or Seedance 2 for short lines.
- Captions. Burn in word-timed captions, especially for vertical video. Most feed viewers start muted.
- Music. Add one track under the whole edit rather than a different track per clip. The music tab generates original tracks.
Keep the voice above the music and check the mix on a phone speaker. If a word is hard to hear, the music is too loud. For longer speech, a lip-sync avatar is the right tool; a native-audio model handles a sentence, not a script.
Fix common failures
Most disappointing AI video falls into a few recognizable patterns, and each has a fix that does not require a new model.
- Flicker. Small details shimmer between frames. Shorten the clip, simplify the scene and avoid fine patterns such as stripes or text.
- Slideshow feel. The clip looks like a still with a slow zoom. Add a real action with a beginning and an end, and name one camera move. Our guide to avoiding the slideshow look covers this in depth.
- Warping faces. Use a sharper, front-on source photo, keep the head still and avoid extreme angles.
- Hands and text. Keep them out of the frame or small. If the shot depends on readable text, add it in an editor rather than generating it.
- Drift over a long clip. Split the shot into shorter takes and cut on the action.
Change one thing per retry. If you change the prompt, the model and the length together, you learn nothing about which change worked. Two or three disciplined retries usually beat ten random ones.
Export for each platform
Set the aspect ratio at generation, not in the export. A vertical frame composes differently from a landscape one, and cropping after the fact can cut the subject or the text.
- 9:16 for Shorts, Reels and TikTok. Keep captions out of the top and bottom interface areas.
- 1:1 or 4:5 for feed posts. Centre the subject and keep the product readable at thumbnail size.
- 16:9 for YouTube, a website hero or an embedded ad.
Export the master at the highest resolution you paid for, then resize rather than regenerate. Regenerating for each platform costs credits and produces a different clip, which defeats the point of a consistent campaign. The aspect ratio guide has the platform-by-platform details, and the Seedance effects tab is a fast way to try a known look before you write a custom prompt.
Frequently asked questions
Do I need a photo to make an AI video?
No. A text prompt works from nothing. A photo gives you more control over the subject and is the better choice when the product or person must be recognisable.
What is the cheapest way to start?
Render a short draft at 480p on a fast model, watch it, and re-render only the version you keep. New accounts get 10 free credits with no card.
Why does my AI video look like a slideshow?
The prompt described appearance rather than motion. Name one action and one camera move, and give the action a beginning and an end.
How long can an AI video be?
Single clips are short: Seedance 2 runs 4 to 15 seconds, and Seedance 2.5 and Wan 3.0 go to 30 seconds. Longer videos are several shots joined in an editor.
Which model should I use first?
Start with Seedance 2 Fast for a cheap draft, then move to Veo 3.1 for realism and audio or Kling 3.0 for faces and controlled motion.
Open the text to video tab to start from a prompt, or image to video to animate a photo you already have. There is no subscription needed, credits never expire, and new accounts get 10 free credits with no card, enough for a first draft before you spend on a finished render.
Ready to try it?
Create your first AI video in minutes — no credit card required.
Start Creating Free →