Video AI: The Five Jobs Inside One AI-Made Video
Try it: open any sample project and name the job behind five things: the summary in Step 1, a character picture in Step 2, a shot list in the editor, the Music switch beside the player, and the Composite Episode button.
Job 1: Reading, done by a text model
In: your script, or the one line that AI Create expands. Out: a summary, characters with an appearance and a voice described in words, locations, props, an episode split, then for each episode a storyboard of numbered shots with seconds, names and dialogue lines. See it: Step 1 shows the summary; the storyboard is in Step 3 and in the editor as a Shots list or plain Text. Change it: edit any of it by hand for free, ask an AI edit in plain words at 1 credit per revision, or restore an earlier version from text history. Cost: the read is 4 credits per 3,333 English characters, once; a storyboard is 1 credit per 100 counted characters of its episode.
The text model draws nothing and shoots nothing. A storyboard is a format the rest of the system reads literally, shot markers, seconds and references to the cast, so its quality is the ceiling for everything after it. What is inside a project.
Job 2: Drawing, done by an image model
In: each description plus the look you chose. Out: a sheet per character (always 9:16), per location and per prop (always 16:9), optional state images, and cover posters. See it: Step 2, one card per entry. Change it: edit the description and regenerate, or replace the sheet with your own image on a paid plan. Cost: the first 15 on an account are free, then 7 credits each on the default image tier.
These pictures are not the video. They are references that every later shot is made from, which is why sheets are neutral on purpose: a plain background and plain light, because the model copies the light and mood of a reference as faithfully as the face. Mood is written per shot in the storyboard. Why sheets look plain.
Job 3: Shooting, done by a video model
In: the sheets of everyone and everything in the shot, the shot text with its seconds, and each speaker’s voice. Out: one segment of up to 15 seconds, or 30 on the long-take tiers, with the picture, the spoken lines, the speaker’s lip movement and the ambience of the place generated together as one clip. See it: the preview player, with every take kept in Version History. Change it: edit the shot text and regenerate that segment, or use Edit this clip for a trim, a picture change, a re-voice or an enhance. Cost: 10 credits a second on the default tier, more on higher tiers; the price is on the button.
This is the job that costs real money and takes real time, typically ten to twenty minutes per segment. It is also probabilistic: two identical requests can give two different clips. The tier picker in the editor’s top bar sets the resolution, from 480P to 2K, and the longest segment it takes. Dialogue and sound in one generation.
Job 4: Scoring
In: the finished storyboard, its scenes, the mood of each stretch and where the lines fall. Out: a plan for the episode’s music, pieces of at least half a minute that change at scene boundaries with a short crossfade, mixed in under the voices. See it: the Music switch beside the player. Change it: only on or off; there is no track picker, volume control or button to compose again. Cost: included in the video price. The built-in score.
Job 5: Joining, which is not a model
In: the active take of every segment, in order. Out: the episode file. See it: the Composite Episode button, available once every segment is done. Change it: switch a segment to another take or regenerate it, then join again; the earlier file stays until the new one replaces it. Cost: 1 credit a second.
The join is plain video processing. It joins without re-encoding wherever it can, so the picture you approved is the picture that ships, and it cannot repair a clip that is wrong. Clips are joined as they are, so a mark that was on some segments and not others stays that way. Composite and share an episode.
Which job caused this?
| What you see | Which job | Where to fix it |
|---|---|---|
| A shot shows the wrong person or action | Reading | Edit the shot text, then regenerate that segment |
| The same face is wrong in every shot | Drawing | Edit the description in Step 2 and regenerate the sheet |
| One shot is odd, the others are fine | Shooting | Regenerate that segment, or Edit this clip |
| A line is spoken in the wrong voice | Shooting, from the voice tag | Check the speaker and voice tags in the shot, then the character’s voice |
| The music is not wanted | Scoring | Turn the Music switch off and download |
| Some parts of the episode carry the badge, some do not | Joining | Set the badge switch, regenerate the odd segments, join again |
Which of these you can choose a model for
Text, image and video each work through a tier: a row of configuration that says which service, which model, what resolution and how long a segment may be. The video tier is the one worth choosing, because that is where the money and the visible difference are; its picker is in the editor’s top bar. The text and image tiers live in Settings under Model Preferences, and the defaults are the ones the pipeline was tuned around. This page names no models because tiers are configuration and change; the picker and the pricing page show the live list. What runs underneath.
| Job | Input | Output | You can change | Cost |
|---|---|---|---|---|
| Reading | Your words | Cast, locations, props, shots | Everything, by hand or by AI edit | By length; hand edits free |
| Drawing | Descriptions and the look | Sheets and state images | Description, regenerate, upload (paid plan) | 15 free, then per image |
| Shooting | Sheets, shot text, voices | One clip with sound | Shot text, tier, clip fixes | Per second |
| Scoring | The storyboard | Music under the voices | On or off | Included |
| Joining | Finished segments | The episode file | Which take, then join again | 1 credit a second |
FAQ
What is video AI?
A set of models and tools that read a text, draw reference pictures, shoot clips with sound, score them and join them. On SceneMixer it is five jobs, and three of them use models you can pick.
Is one AI model making the whole video?
No. A text model, an image model and a video model do different jobs, and the join is ordinary video processing. The music is planned and mixed by the site.
Which part costs the most?
Shooting. Video is charged per second, and every other step is small by comparison.
Can I choose the model?
You can choose the video tier in the editor and the text and image tiers in Settings. The pricing page shows the live list; defaults are what the pipeline was tuned on.
Why can I not edit the sound separately?
Voices, lip movement and ambience are generated with the picture as one clip. You can re-voice a finished clip, turn the score off, or regenerate the segment.
Does the AI read my script literally?
It reads it into a cast and shots, adding what a storyboard needs. You can read and edit the result before any video is paid for.
Is there a human in the loop?
You are. The two points where you check are the cast in Step 2 and the storyboard before video. Both are free to edit.
Why do two identical requests give different clips?
Video generation is probabilistic. That is why you read the storyboard first, shoot one segment before the rest, and regenerate single shots.
See the five jobs on one short script
Read it, check the cast, shoot one segment, and notice which job each screen belongs to.
Updated November 2, 2026