Generative AI Video: What Is Generated, and What You Control
Try it: pick one thing you care about in a shot, a face, a coat, where someone stands, the weather, and decide whether it should be set with a picture, with words or with a setting. The table below says which one holds.
Generated or assembled: a quick way to tell
A template or stock editor assembles: it takes clips, text and music that already exist and places them. Nothing in the result was made for your sentence, so it is repeatable, cheap per change and limited to what exists. A generative tool makes: the picture, the speech and the sound are produced for your request. You can ask for something no library holds, a character with your description in a room you described, but each request is a new draw.
The test is easy. Ask for something that cannot be in stock footage, such as a named character in a specific outfit saying a specific line. If the tool gives you back a close match of existing material, it is assembling. Here, a segment is produced from your cast pictures, your shot text and your voices; nothing in it is retrieved from a library of clips. The join at the end is the one non-generative step: it stitches finished segments together.
Pictures bind
When a shot is generated, the model receives the sheets of the characters, the scene and the props named in it as pictures, and it reproduces them closely: face, clothes, materials, and whatever light and background are in the picture. That is the strongest control there is, and it cuts both ways. A picture binds what it shows, and no sentence will undo it. So the way to change a face or a coat is to change the picture: edit the description and regenerate the sheet, replace it with your own image on a paid plan, or add a state picture for a changed look.
It is also why sheets are neutral: a character drawn under noir lighting brings that lighting into a sunny picnic. Mood is written per shot in the words, where it belongs. Why sheets look plain · Using your own reference images.
Words steer
The shot text decides what the picture does that the pictures do not: where people stand, what they do, where the light comes from, the weather, the time of day, how the camera moves and, word for word, what is said. Words work on the layer above the pictures, so rain, dusk, a person sitting down and who holds a cup all ride on top of whatever the sheets show. They steer rather than fix. Nouns weigh more than function words, so a negative often produces the thing it names, and a list of adjectives pushes the important words down. Write what is in front of the camera, with the cause of the light, not the verdict you want on the result.
Describe the cause, not the result · No magic words · Write camera moves.
Settings frame
Settings set the box that pictures and words work inside. The visual style is the medium, from photographic to painted to 3D, and applies to everything. Scene style and costume style are applied once, when the script is read, and flow into every description. The shape, 9:16 or 16:9, is fixed after Step 1. The video tier sets the resolution, from 480P to 2K, and the longest segment, 15 or 30 seconds; the seconds in each shot add up inside that ceiling. Settings never say what happens in a shot, and the system keeps them from fighting the words: it does not tell the model where to put the subject, because that is the storyboard’s job.
Choosing your style settings · 15 or 30 second segments · Vertical or landscape.
What you want to control, and which control holds it
| You want to set | Use | How firmly |
|---|---|---|
| A face, age, build | The character sheet | Binds: copied closely |
| An outfit | The sheet, or a state picture for a changed outfit | Binds; words cannot override it |
| The layout of a room | The scene sheet and the scene’s 3D model | Binds the space |
| Where people stand | Words, as room anchors | Steers; holds across cameras |
| Light, weather, time of day | Words in the shot | Steers; works on top of the pictures |
| What is said | A line with speaker and voice tags | Spoken word for word |
| How a voice sounds | A written voice or a preset from the library | Steers, or binds once a preset is attached |
| Length and detail | Seconds per shot, the tier | Frames: within the ceiling |
| Exact framing, small gestures, hands | Nothing reliably | Left to the model; regenerate |
What stays random, and how to work with it
Two identical requests can give two different clips; the terms say so and the honest version of any generative product is one where that is expected. The parts that vary are the exact framing, the performance, small details and anything you left unsaid. Work with that, not against it: read the storyboard before you pay for video, shoot one segment first, fix single shots by editing their text and regenerating only those, and keep earlier takes, since every generation of a segment is kept in Version History and switching between them is free. Edit one shot and regenerate · Segment versions.
Because every clip is a new draw, cost is per attempt, not per result: 10 credits a second on the default tier. The cheapest control you have is the one before generation, the text, which is free to edit as often as you like.
| Kind of control | Examples | Holds by |
|---|---|---|
| Pictures | Character, scene and prop sheets, state pictures, your own photo | Being copied closely |
| Words | Shot text, positions, light, weather, dialogue lines | Steering what the pictures leave open |
| Settings | Visual, scene and costume style, shape, tier, seconds | Setting the box everything happens in |
| None | Exact framing, gestures, hands | Regenerate and choose |
FAQ
What is generative AI video?
Video whose pictures, speech and sound are produced by a model from a description, rather than assembled from existing footage or a template.
How is it different from a template video editor?
A template editor places existing clips, text and music, so it is repeatable and limited to what exists. A generative tool makes new picture and sound for each request, which allows new scenes but not identical repeats.
What can I control in a generative video?
Pictures bind a face, outfit and room; words steer positions, light, weather and dialogue; settings frame style, shape, resolution and length. Fine framing and gestures are left to the model.
Why does my text not change the character's clothes?
A reference picture is copied with its clothes, and text cannot override it. Change the picture, edit the description and regenerate the sheet, or add a state picture.
Can I get exactly the same video twice?
No. Generation is probabilistic. You can keep a take you like, switch between takes for free, and regenerate only the shots that miss.
Is any part of the video not generated?
The join that stitches finished segments into an episode is ordinary video processing. Voices come from a library of presets or a written description, and the score is chosen for the episode.
Does generative mean random?
Partly. The pictures and words you supply hold the identity, room and story; what varies is the exact framing, performance and small detail.
What is the cheapest control?
The text. Editing the storyboard is free, and a sentence fixed before generation saves a regeneration at the per-second price.
Set the pictures first, then steer with words
Approve the cast, write the shot, and generate one segment before the rest.
Updated November 2, 2026