brainax

How to make AI videos: a step-by-step guide for 2026

Typing "a girl walks through a city" into a generator gets you a pretty clip, and the comments instantly say "AI slop". Good AI video is not one lucky prompt. It is a short pipeline: know where models fail, pick the right model for the shot, write the prompt in blocks, keep the face consistent, fix instead of rerolling, count the budget, cut for the feed and label it. Here is that pipeline.

Free course on this topicAI Video, No Flops: free courseStart the course →

Step 1: know what AI video gets wrong

Viewers catch the same mistakes every time: feet sliding over the floor, broken letters and logos, objects that morph into something else mid-shot. Even big productions hit this: the 2024 Coca-Cola AI ad was criticized for exactly these glitches, and the team behind the short film Air Head said they got roughly one usable second for every 300 generated.

Before you publish, do a 10-second check: play the clip at 0.25× speed and watch feet, hands, text and edges of objects. Lesson: Why AI video glitches

Step 2: pick the model for the shot, not the brand

There is no single best AI video generator. As of October 2026 each strong model has its own job:

Models also disappear: OpenAI switched off the Sora app on 26 April 2026 and the API on 24 September 2026. So keep your prompts and reference images on your side, not only inside one service. Lesson: The 2026 model map

Step 3: write the prompt in blocks

A video prompt is not a wish, it is a shot description. Start with a fixed style line you reuse in every shot (for example "35mm, light film grain, 9:16, no text on screen"), then describe the first frame and where everyone stands, then one action. One shot is one action: "she turns and smiles" and "she runs to the car" are two generations, not one.

Choose the mode on purpose. Text-to-video gives the model freedom; image-to-video starts from a frame you already approved, so the look stays under control. Lesson: Anatomy of a prompt

Step 4: keep one face across the whole video

The classic flop is a hero whose face changes every shot. Make a character sheet first: the same person from 6 angles. Then write a short "character passport" (age, hair, clothes, details) and paste it into every prompt word for word. Add reference images: Veo 3.1 accepts up to 3, Seedance 2.5 up to 50.

Don't ask the model to draw your logo: it breaks letters. Add the logo in the edit. Then put frames from all shots in one grid and compare the faces side by side. Lesson: One face for the whole video

Step 5: fix, don't regenerate

When one detail is wrong, a reroll is a lottery: along with the problem you lose everything that turned out well. Diagnose first, then use the right tool: region edit for one wrong spot, restyle for the look, extend or interpolate for length, and upscale only at the very end, only for final shots. Finish the edit in CapCut or the free DaVinci Resolve. Lesson: Edits instead of regenerating

Step 6: count the budget before you press generate

One good shot often takes 20 or more attempts, so the price of an attempt decides everything. Veo 3.1 at 1080p costs $0.40 per second, so one 8-second clip is $3.20. The Fast tier is $0.12 per second and Lite is $0.08.

Real math from the lesson: 300 attempts on the top tier cost 300 × $3.20 = $960. Drafting the same 300 attempts on Lite costs $192, and rendering the 15 chosen shots on the top tier adds $48: $240 in total, four times cheaper. Draft cheap, render final shots expensive. Lesson: Budget and credits

Step 7: assemble for TikTok, Reels and YouTube Shorts

Viewers decide in the first 3 seconds, so the hook goes first. A simple 15-second skeleton: 0–3 s hook, 3–8 s story, 8–12 s the product or the point, last 3 s the call to action.

Generate in 9:16 from the start. If you crop a 1080p 16:9 clip to vertical, only about 608 pixels of width are left. Export MP4 (H.264 video, AAC audio) at 1080×1920, and watch it on a phone with the sound off before posting: many people will. Lesson: Assembling a video people watch

Step 8: label AI video

An "AI-made" label is no longer a matter of taste. From 2 August 2026, Article 50 of the EU AI Act requires labelling realistic AI content, with fines up to €15 million or 3% of worldwide turnover. China has had labelling rules since 1 September 2025. TikTok, YouTube and Meta have their own AI labels, and tools like C2PA content credentials and Google's SynthID mark files invisibly. The simple rule: if it looks real and it was generated, label it. Lesson: Labelling AI

FAQ

How can I make AI videos for free?

Generation itself costs credits on almost every serious model, so "free" means spending as little as possible: write and test prompts carefully, draft on the cheapest tier (Veo 3.1 Lite is $0.08 per second), render only final shots on the top tier, and edit in the free DaVinci Resolve.

Which AI is best for making videos?

It depends on the shot. Kling 3.0 is strong at body motion, Seedance 2.5 at long clips with many references, Veo 3.1 and Gemini Omni at speech and sound, Runway at editing footage. Pick the model for the task, not one brand for everything.

Do I have to label AI videos?

For realistic AI content, yes in more and more places: the EU AI Act requires it from 2 August 2026, China since 1 September 2025, and TikTok, YouTube and Meta have AI labels. Labelling also protects trust: viewers who find out later feel tricked.

Free course on this topicAI Video, No Flops: free courseStart the course →