How to Make AI Videos in 2026: Idea to Export, Step by Step

Three inputs cover almost every job. A text prompt when the scene does not exist yet. A single image when the subject already exists and only needs motion, which is the usual route for products and portraits. A start frame and an end frame when the clip has to travel from one exact composition to another. Inside the AI video generator these are the mode tabs at the top of the left panel, and everything below them changes with the mode.

Which generation mode fits your shot?

• Text-to-video: establishing shots, B-roll and scenes you cannot photograph. Weakest on exact faces and product labels.
• Image-to-video: the frame you upload is the frame you get, so identity, packaging and typography hold. Motion is the only variable.
• Start-and-end frames: two stills, one transition. Use it for reveals, before-and-after cuts and camera moves that must land on a specific composition.
• Templates: a fixed prompt and settings that only need your upload. The fastest route to a first clip and the cheapest way to learn what a model does.

What does a credit buy?

Every new account starts on the Free plan, which the pricing page lists at 40 credits a month with a monthly refresh and no card required; paid plans begin with the Starter tier at 4,500 credits a month. The same page states the rule that matters for budgeting: credits are charged by model, duration, resolution and output count. The generate button shows the exact cost of the run you have configured before you click, so read it every time you change a setting. On 2026-09-05 it showed 829 credits for a 5-second 720p clip on Kling v3 Omni and 410 credits for the same length on the default Seedance model, which is why the 40 free credits are for learning the panel and running image tools, not for a monthly video habit.

A generator gives you shots, not a finished video. Cutting, captions, music, voiceover and colour matching still happen in an editor. Some models add sound, but the Veo 3.1 Lite tier in LazyKiwi does not, so plan on adding audio yourself unless the model page says otherwise.

This is the loop I run for every clip, whether it ends up in a product ad or a vertical story. Each step holds one decision. When a clip comes back wrong, go back to the step that owns the problem instead of rewriting everything.

1. 1

   Write a one-line brief

   Say what happens, to whom, for how many seconds, on which platform. 'A ceramic mug on a wooden desk, steam rising, slow push-in, 5 seconds, 9:16' is a brief. 'Cosy coffee vibe' is not.

2. 2

   Pick the mode

   Existing subject: image-to-video. New scene: text-to-video. Must land on a specific frame: start-and-end. Never use text mode for a real face or a real label.

3. 3

   Pick the model

   Match the model to the job with the table in the next section. For a first attempt, Kling v3 Omni handles all three modes and is the safest default for motion.

4. 4

   Write the prompt in five parts

   Subject, action, setting, camera, light, in that order, with one action per clip. OpenAI's video guide describes the same split: the prompt carries subjects, camera, lighting and motion, while size and seconds are parameters, per OpenAI Platform docs (2026).

5. 5

   Set ratio, duration and resolution

   Choose the ratio the platform needs (9:16 for feeds, 16:9 for YouTube), the shortest duration that fits the action, and the lowest resolution you can judge motion at. Longer and sharper both raise the credit cost shown on the button.

6. 6

   Generate one clip, not four

   Run a single output and look at it before spending on variants. Most first renders fail on a prompt problem that four copies would all share.

7. 7

   Review at full size

   Play it at 100% and watch hands, text, and anything that crosses the frame edge. Note what broke, change one variable, run again.

8. 8

   Export and finish in an editor

   Download the MP4, trim the dead frames at the start and end, add captions and sound, then export with the platform's recommended settings. YouTube lists MP4 with H.264 and an 8 Mbps target for standard-frame-rate 1080p, per YouTube Help (2026).

The workbench lists six video models, and they differ more in limits than in style. Duration ceilings, resolution, aspect ratios and how much prompt they read decide which one fits a job, so the table below is taken from each model page as of 2026-09-05.

Which model for portraits, products and B-roll?

Portraits and anything with a face: image-to-video on Kling v3 Omni, which keeps the uploaded frame and adds motion, or Hailuo 2.3 Fast for a quick 768p pass. Products with labels: Seedance 2.5 at 720p with the packshot as the reference, because its nine reference slots let you pin the label from more than one angle. Wide B-roll and establishing shots: Seedance 2.5 or MiniMax H3 in 21:9, then crop. Anything delivered at 2K: MiniMax H3 is the only option in the lineup.

For a second opinion beyond your own tests, the Artificial Analysis Video Arena (2026) ranks text-to-video and image-to-video models from blind votes, and ByteDance's Seed page (2026) describes Seedance as a multi-shot model that works from both text and image. Sora 2 is not in the LazyKiwi workbench, so do not plan a workflow around it here.

Weird motion almost always comes from asking for too much movement at once. The fix is a prompt that gives the model one subject motion and one camera motion, both slow, with the physics implied by the setting. This is the part of how to make realistic AI videos that no setting can do for you.

[Shot type] of [one subject] [one action with a clear start and end] in [setting with a named light source]. Camera: [one movement, slow]. Lighting: [one description]. Style: [film stock or lens]. Keep [the thing that must not change] unchanged.

A prompt that works with this shape: 'Medium shot of a barista pouring latte art into a white cup on a marble counter, morning window light from the left. Camera: slow push-in. Lighting: soft, warm. Style: 35mm, shallow depth of field. Keep the cup and logo unchanged.' We ran it through Kling v3 Omni and Seedance 2.5; Kling kept the pour continuous, Seedance held the marble texture better.

Which camera words do the models understand?

Use the vocabulary an operator would use. Pan, tilt, push in, pull out, dolly, tracking, arc, boom and handheld are each defined with film examples in StudioBinder's camera movement guide (2026), and models trained on captioned footage respond to those words far better than to 'cinematic camera'. Pick one per clip. Two movements in one prompt is the most common reason a clip warps halfway through.

• Say what should not move: 'the background stays static', 'no camera shake', 'hands stay on the cup'.
• Give actions an end state: 'raises the cup to chest height', not 'drinks coffee'.
• Skip crowds and mirrors on a first pass; both multiply the things that can break.
• Match duration to the action: a pour takes about four seconds, so do not ask for ten.

Prompt length matters less than order. Seedance 2.5 reads up to 5,000 characters and MiniMax H3 up to 7,000, per their model pages, but a short prompt with the structure above beats a paragraph of adjectives. For animating a real photo, where identity drift is the main risk, the photo-to-video guide adds the rules for faces.

The honest answer to how to make AI videos for free is that the Free plan buys you the workflow, not the volume. The pricing page lists it at 40 credits a month with the template library and both video modes included, and at the run prices shown on 2026-09-05 (190 credits for 6 seconds of Hailuo 2.3 Fast at 768p, 473 for 5 seconds of Wan 2.7, 829 for Kling v3 Omni) a single video-model run costs more than a month of free credits. Free, in practice, means template pages marked free, promo-tier models and image tools; the first paid step is the Starter plan at 4,500 credits a month, roughly five Kling clips or more than twenty Hailuo clips at those prices.

Is there a free AI video template?

Some are. Templates mode inside the video workbench fixes the prompt and the settings, so your upload is the only variable, and several template landing pages on lazykiwi.ai carry a free label (the photo slideshow template page, for one, lists 720p free with no watermark). Each template run is priced by the model behind it, and that price appears on the button before you confirm, so check it the same way you would for a plain generation.

• Test at the lowest resolution the model offers (480p on Seedance 2.5) and rerun only the keeper at 720p or 1080p; the button price drops with resolution.
• Use the shortest duration that contains the action; a 5-second clip that works beats a 10-second clip with a good half.
• Generate one output per run. Variants cost the same as the original.
• Sharpen a keeper with the AI video enhancer instead of re-rendering at a higher resolution when the motion is already right.
• Storyboard with stills first; animating a frame you have not approved is the fastest way to lose credits.

How do free tiers compare across AI video tools?

For context, the free tiers of the main generators were read from their pricing pages on 2026-09-05. Runway's Free plan is a one-time 125 credits with no monthly refresh, per Runway pricing (2026). Pika's Basic plan is free with 80 monthly video credits, per Pika pricing (2026). Luma's pricing page starts its listed plans at the paid Plus tier and offers a free trial without itemising it, per Luma Labs (2026). CapCut is an editor with free AI tools rather than a generator, per CapCut (2026).

Short-form is the same workflow with three constraints: 9:16, movement in the first second, and captions that carry the message because most feeds start muted. Generate in 9:16 rather than cropping 16:9 later, since the models compose for the ratio you set; Kling v3 Omni, Veo 3.1 Lite and Seedance 2.5 all offer 9:16 natively.

• TikTok: open on the action, not the establishing shot. A clip whose subject is already moving beats a slow fade in. The TikTok guide covers hooks, trend timing and captions.
• YouTube Shorts: export at 1080 by 1920 and switch on the AI-use setting in YouTube Studio when the clip looks realistic.
• Reels: Instagram compresses hard, so run a 720p render through the AI video enhancer before uploading rather than paying for a 1080p regeneration.

One habit that helps on all three: generate a 16:9 hero clip for YouTube and a separate 9:16 run of the same prompt for feeds. The vertical version composes the subject for the frame instead of cutting it off, and you learn which prompts survive both ratios.

• Over-long prompts. Adjectives do not add control; structure does. Cut to subject, action, setting, camera, light.
• Wrong ratio. Cropping a 16:9 render to 9:16 loses the subject; set the ratio before generating.
• Generating before storyboarding. Approve the still first, then animate it in image mode.
• Ignoring audio. Veo 3.1 Lite in LazyKiwi has no sound; every clip needs audio added in the edit.
• Publishing without a label. Realistic AI content needs a disclosure on YouTube and TikTok.
• Trusting the thumbnail. Watch the full clip at 100%; artefacts hide in the last second of a render.

Do AI videos need a disclosure label?

On YouTube, yes when the content looks realistic: creators must disclose AI that meaningfully alters or generates photorealistic content, and the video then carries an altered-or-synthetic label, while non-realistic content and minor edits are exempt, per YouTube Help (2026). TikTok keeps its own AI-generated content label and the rules for applying it in its help centre, per TikTok Support (2026). Turn the label on at upload; removing a video later costs the reach you built.

How to create AI videos well is mostly discipline: one action per clip, one camera move, the right ratio from the start, and a label when the result looks real. Everything else is a setting you can read off the generate button.