AI Video: What It Can and Cannot Do Today
By William, Partner, technology · · 6 min read
Today AI handles short single-shot clips with simple motion well. It struggles with long durations, characters that must look identical from shot to shot, readable text, hands and complex physics. You get more by writing one concrete shot and starting from a good image.
What does an AI-generated video clip do well today?
It does short shots with one subject and simple motion well: a camera creeping toward a cup, steam rising, leaves moving in the wind, a product turning on a surface, a landscape with drifting clouds. None of these ask anything to stay identical for long.
That is the useful rule: the less the model has to remember, the better it turns out. A clip of a few seconds with one clear action is often usable. A long scene with three characters, dialogue and a change of location is not.
It is especially handy for:
- Backgrounds and moods for a piece (a kitchen, a street, an office) with a message laid on top.
- Cutaway footage (b-roll) that supports a caption or a voiceover.
- Bringing a good image to life so it is not static in short-video formats.
- Visual ideas you cannot film: an impossible camera move, an environment that does not exist.
What are the real limits of AI video?
There are five, and it pays to know them before spending credits.
Duration. Models generate short clips, from a few seconds to a bit over ten depending on the model. The exact figure differs between models and changes month to month, so check what your tool offers at the moment you generate. A longer video is built by joining clips, not by asking for one long clip.
Consistency of characters and objects. The same person, garment or package tends to drift subtly between clips: the face, the hairstyle, the colour, the shape of a logo. Inside a single clip it usually holds up better; across separate clips, worse.
Text. Letters and signs inside a video warp or change as they move. If the message must be readable, add it afterwards, on top of the video.
Hands and fine detail. Extra fingers, objects passing through each other, cutlery holding nothing. The closer to the camera and the longer on screen, the more it shows.
Physics and complex actions. Pouring liquids, fabric in motion, objects colliding, people running or dancing: any of these can look convincing or give the trick away. Smooth, continuous movement is the safer bet.
Nobody can promise what quality you will get in a specific case. The honest approach is to run a cheap test and decide with the result in front of you.
Is it better to generate video from text or from an image?
For most brand work, start from an image. With a starting image you decide the look and the model only has to animate it; with text, the model decides look and motion at the same time.
| Text to video | Image to video | |
|---|---|---|
| Control over the look | Low: the model interprets | High: it comes from your image |
| Best for | Exploring ideas, moods, standalone shots | Products, brand scenes, coherent series |
| Main risk | Very different results between attempts | Motion that distorts the image |
| Cost of testing | Several attempts until it lands | Fewer attempts if the image is already good |
A reasonable flow: get the image right first (see how to write AI image prompts), then animate it with a short motion instruction.
How do you write the shot so the model understands it?
Write it as if directing a single shot, with three elements: what we see, how the camera moves, and what moves inside the frame.
- What we see: subject, place, light. "A coffee cup on a wooden table by a window, morning light."
- Camera: static, slow push-in, gentle pan, overhead. One camera instruction per clip.
- Internal motion: "steam rises from the cup", "the curtains sway slightly". One action, and a gentle one.
Example with a coffee shop called "North Bean": "Medium shot of a white cup on a light wooden table by a window, soft morning light. The camera moves very slowly toward the cup. A thread of steam rises. No people, no text, no logos. Vertical 9:16."
Three practical tips:
- One idea per clip. If you need two actions, make two clips.
- Ask for slow motion. Fast movement is where models fail most.
- Remove what you do not need: people, text, hands. Fewer elements, fewer errors.
How do you keep the cost of generated video under control?
Test small and spend the expensive model only at the end. Video is among the most expensive things you can generate, and cost depends on the model, the duration and whether it includes sound. Always check the estimated cost before launching.
In igroui you see that cost before generating, credits are reserved while it runs and the final charge is the real cost, which can vary slightly; if generation fails, the reservation is released. You can choose model, duration and ratio, and some models generate with sound while others do not.
Habits that save credits:
- Start with the shortest duration that tells the idea.
- Test on a cheaper model; move up only for the version you will publish.
- Generate without sound if you will add your own music or voice later.
- Before asking for three variations of a shot, fix the description: a vague prompt times three costs three times as much.
- Keep the clips that work and reuse them instead of regenerating.
To see how a clip is generated from text or an image, choosing model, duration and ratio, there is the AI short videos feature.
What about long videos with several scenes and the same character?
A single clip still cannot handle them well; the tool has to help you keep style and character across scenes. In igroui you can generate individual clips or work with scene-based video, with a script and a consistent character, from the video editor.
And if you prefer to do it by hand with short clips, here is what you can do:
- Think of the video as a sequence of 3-6 second shots and generate each one separately.
- Write the same style into every shot (light, palette, camera type).
- Join them outside, in your usual editor, with your own music or voice.
- Choose shots where the character is not seen up close: hands out of frame, silhouettes, product details.
What should you check before publishing an AI-made video?
Check four things: the whole clip, fit with your brand, rights and transparency.
- The whole clip, at normal speed and in slow motion. Failures often show up in just a few frames: an odd hand, an object that appears and vanishes.
- Fit with your brand: colours, style, tone. If a shot does not look like it belongs to your brand, leave it out.
- Rights: the terms of the model and of the network you publish on, and that you are entitled to use any images you uploaded as a base.
- Transparency: say the video was made with AI. That is the recommended practice and, depending on the network and the country, may be mandatory. Whoever generates and publishes is responsible for the output.
And in igroui nothing is published without human approval: every piece waits for a yes, and the approval workflow lets you request changes with a comment on the exact version. That is the moment to watch the clip as an audience member, not as the person who asked for it.