AI Video Generation: How It Actually Happens
Try it: generate one segment twice on the cheapest tier without changing a word, then change one sentence and generate once more. Two identical requests can give two different clips; the third shows how much a single sentence of text matters.
One generation is one short clip
A video model takes a description and some reference pictures and returns a clip. It does that once per call, and the tier you pick decides the longest clip a call may return. On SceneMixer that ceiling is 15 seconds on some tiers and 30 seconds on the long-take tiers, and the storyboard is written to fit under it. One such clip is a segment. A minute of finished video is therefore at least four segments on a 15-second tier and at least two on a long-take tier, each made by its own call, never by one long one.
Inside a segment the story is cut into three to five numbered shots, and the seconds of the shots add up to the segment’s length. The ceiling is a maximum, not a target: a segment with two short shots may run only a few seconds. When the shots add up to more than the limit, the silent ones are tightened first and then the segment is split at a shot boundary. Nothing is cut off the end. Price follows seconds, so an episode costs the same cut into short or long pieces; what changes is how much you pay to redo one. 15 or 30 second segments.
What goes into one call
A call is made of three things:
- The segment’s own text. A header that names the one place and the time, then numbered shots with their seconds, the people and props named with @ chips, and the lines with a voice.
- Reference pictures. One for each person in the shot, one for the place, one for each prop that matters, and a state picture for anyone in a different outfit or condition.
- The tier. It fixes the resolution, the longest clip and the price per second.
The clip before it is not shown to the model. That is why a segment must stay in one place and one timeline: it is drawn from one set of scene pictures, so a shot that jumps to another room or another day is refused before you pay and has to become its own segment. A segment stays in one place and one timeline.
Why faces change, and what holds them still
Because each call starts from nothing, anything that is only described in words is invented again every time. Most of what looks like “AI inconsistency” is that, and each case has a different supply.
| What drifts | Why it happens | What the site supplies | What you do |
|---|---|---|---|
| A face between shots | No memory of any earlier call, so the face is invented again | One approved picture per character, attached to every shot that names them | Name the character with the @ chip; fix the picture once in step 2 |
| A room between shots | Same reason, for places | One picture per location, and an automatic AI 3D model of each scene so who sits where is settled before the shots are made | Keep one place per segment |
| An outfit or an injury | Nothing carries a wound from one call to the next unless it is written again | A running record kept in text, and state pictures for looks that last | Write the change in the shot where it happens |
| A voice | No call remembers how a character sounded in an earlier one | One voice per character, chosen once and used for every line | Pick or upload the voice in the cast |
That is also why a higher tier does not fix a wrong face: resolution changes how sharp the picture is, not who is drawn. My character’s face changed: a checklist and why a whole series stays consistent go through the fixes.
What carries across a join
In storyboards written or regenerated since late October 2026, the story is carried from one segment to the next by a running record that the site keeps while it reads your text: who is in which scene, whether they sit, stand, lie, kneel or crouch, what is in their hands, which look they wear, wounds and mud on the body, wet or dry, unconscious or not; and for each scene the time of day, the weather, the light, whether doors and windows are open, and whether a crowd is there. Each new segment starts from that record plus its own text and pictures.
| You write, in segment 3 | Segment 4 starts with |
|---|---|
| @Mara takes off her coat and lays it over the back of the chair. | Mara without the coat; the coat on the chair |
| @Leo is hit across the face; blood runs from his left eyebrow. | A cut on Leo’s left eyebrow, still there |
| @Leo washes his face at the sink. | The blood washed away |
The record holds what is still true when a shot ends, not what happens during it. It is not a picture of the previous clip: no last frame is attached, and regenerating one segment never changes the next. It also invents nothing you did not write. Storyboards from before late October 2026 work the older way, each segment from its own text and pictures. What the system keeps track of from one segment to the next.
From the button to the file: where the minutes go
- Credits are held. Pressing Generate holds the segment’s price; it is spent only if the clip arrives.
- Your own limit. An account runs a fixed number of generations at once: 3 on Free, 5 on the first paid plan, more above that. The rest wait, credits held, and start by themselves.
- The shared limit. The whole site can have only so many jobs in flight at a model service, so when it is busy yours waits behind other people’s. The top plans put their jobs ahead in that line.
- The model runs. This is almost all of the time. An internal measurement in August 2026 averaged about a minute and a half for a setup picture and about nine and a half minutes for a segment; the card on screen shows a typical ten to twenty minutes for video.
- Delivery. The clip is stored, the AI badge is added unless your plan switched it off, and the card turns Done.
A cheaper, lower-resolution tier is usually also a faster one, which makes it the quickest way to find out whether a shot works. Why a generation takes time.
When a call fails
Models fail in different ways and the site treats the ways differently. If the service says it is busy or rate-limited, the call is tried a second time on another provider account for the same model, with no change to what you asked for or the tier you picked. A moderation refusal and a parameter error are not tried again: a refusal is a judgement, and asking again gets the same answer.
What happens to the credits depends on what is known. If the model never received the request, or answered that it was busy, the credits are released back to you. If the call timed out or the connection broke in flight, nobody can tell whether the model ran and billed, so the credits stay frozen, tagged Under review in your credit history, until a person settles it. A balance that looks short after a failure is usually held, not gone. When a generation fails · Provider busy, not charged.
Steady a generation before you pay for it
- Shoot one segment first. Check the faces, the room and whether the line is spoken before you start the rest.
- Name who is in frame. A shot that says “a woman turns to the window” gives the model nothing to hold; @Mara turns to the window does. Who is in frame.
- Say what is there, not what is not. “An empty street” works; “no people” invites people. Say what is there.
- Fix text before you re-render. Editing a storyboard is free; a regeneration is a full charge.
- Draft cheap, finish high. Draft at the lowest tier, finish at the highest.
Make it in SceneMixer, start to finish
- Script: paste your own on Upload Script, or write a one-line brief in AI Create and let it draft the episodes.
- Step 1: Type Short Drama; Scene Style Street Life; Costume Style Streetwear; Visual Style Realistic HD; Target Region; dialogue language; 9:16 Portrait. Then read the script.
- Step 2: the two leads with a description each, then the alley and the rooftop as locations. The first 15 setup images on an account are free.
- Storyboard: generate episode 1 first (free on your first drama) and read it before you render anything. One place per segment.
- Video: render on the default tier at 10 credits a second, fix single shots instead of whole episodes, then put the episode together and post.
| Stage | What happens | Typical time | What you control |
|---|---|---|---|
| Read | Your text becomes a summary, an episode list and a cast, places and props in words | Minutes | The text; the three styles |
| Setup pictures | Each description becomes a reference picture | About a minute and a half each | The description; regenerate one |
| Storyboard | Each episode becomes segments, and each segment becomes numbered shots | A few minutes | The text, which is free to edit |
| Segment | One model call returns one clip | Typically ten to twenty minutes | The tier, the seconds, the shot text |
| Join | Finished segments become one file per episode | After the last segment is done | Using one tier for every segment avoids a re-encode |
FAQ
How does AI video generation work?
A model takes a text description and some reference pictures and returns a short clip in one call. Longer videos are made by joining many clips, with the system supplying the characters, places and story so far from outside.
Why are AI video clips so short?
Each call has a ceiling set by the model tier, 15 or 30 seconds here. A longer video is several segments joined, not one long generation.
Why does a character's face change between clips?
Each call starts with no memory of the others. A face that is only described in words is invented again; a face supplied as the same reference picture, with the character named in the shot, stays the same.
Does the model remember the previous clip?
No. A segment is generated from its own text and pictures. A running record of who is where and what they wear carries the story across the join, and no last frame is attached.
How long does AI video generation take?
A segment typically takes ten to twenty minutes, depending on the tier and the queue. You can leave the page; the work continues on the server.
What happens if a generation fails?
If the model never received the request, the credits are released. If the outcome is unknown, such as a timeout, they stay frozen until a person settles it. Delivered failures are not charged.
Why does the same prompt give different results?
Generation is probabilistic: two identical requests can give two different clips. That is why the cheap tier is for finding out, and why editing one sentence is worth testing.
Will a higher tier fix a wrong shot?
Rarely. The tier sets resolution and the longest segment, not who is drawn or what happens. The shot text and the character's picture decide that.
Watch one request go from text to clip
Paste a page, read the storyboard, and shoot a single segment first.
Updated November 2, 2026