Generate Videos With AI: What Happens to Your Script
Try it: paste one scene written twice, once as lines (Mara (quietly): I saw him.) and once as prose (She told him quietly that she had seen him.). Compare the two episode outlines. The lines survive word for word; the prose becomes a summary.
You write a script, and the read is the first thing that happens
There is no single prompt box that produces a video. Text goes in, a read runs, and three layers of text come out for you to steer: the summary and the cast, then each episode’s storyboard, then each segment’s shots. The video model only ever sees the last layer and the pictures made from the cast. So the useful question is not “what do I type to get this shot” but “what do I write so that the read, the cast and the storyboard all agree on this shot”. Uploading a script and parsing it follows the buttons.
What the read takes in
| Input | The rule |
|---|---|
| Format | Pasted text, or TXT, DOCX or Markdown. A DOCX is unpacked in your browser first. PDFs and links are refused: export to text or Word, and paste the text itself |
| Minimum | About a paragraph, roughly 330 characters of English. Below that there is nothing to split into episodes |
| Maximum | A hard stop, not a silent cut. Under about 10,000 counted characters the script is read in one pass; above that it is read in two, up to 100,000 counted characters, or 300,000 raw characters of English |
| Counting | Every 3,333 characters of an alphabetic language count as 1,000 |
| Price | 4 credits per 1,000 counted characters, charged once however the script is read, and refunded if the read fails |
| Language | Any. Dialogue in another language than the project’s dialogue language is translated into it at the read |
The ceiling refuses rather than truncates for a reason: the failure measured on over-long input was a script that came back looking complete and was not, such as a fifty-episode book returned as ten. What you can feed in.
What the read hands back
The reading panel names its stages: finding characters, finding scenes, finding props, splitting episodes, shaping the visual style, preparing episodes. When it ends, Script Parsed shows:
- the title, the protagonist, the era and how the story was adapted;
- the episode list, cut according to the project’s Type (short episodes with hooks for a Short Drama, longer ones for a TV Series);
- the counts of characters, scenes and items, each with a description and no picture yet.
For a long script the first pass makes the cast, the second makes the episode list with a short outline each. The first three episodes get their full outline in the background straight away; the others are prepared when you press Generate storyboard on them, and the button waits and continues by itself. If the split is wrong, this is the moment to fix the text and read again. Long scripts are read in two passes · Checking the summary.
Write dialogue so it survives
| You write | The storyboard gets |
|---|---|
| Mei (quietly): I saw him. | The line, spoken quietly, on Mei’s shot |
| She told him quietly that she had seen him. | A summary. Nothing is spoken |
| Father (on the phone, hoarse): Come home. | The line, heard, without showing the speaker |
| (caption: Three years later) | Kept as a caption line; not drawn into the clip at the moment |
| Crowd: Run! | A group line, spoken together |
One sentence per line, speaker first, delivery in brackets before the colon. Off-screen speech is a note before the line: (on the phone), (voice-over), (over the radio). A standard screenplay layout also works. Write dialogue so it is picked up · Using a standard screenplay as input.
One spelling for every name and place
The read matches names literally. A character called MARA in one scene and MRS. CHEN in the next becomes two characters. A place written as MARA’S KITCHEN in one scene and KITCHEN in the next can become two locations, and each one gets its own picture. Pick one spelling, keep scene headings such as INT. KITCHEN – NIGHT, and the read will find the places and the time of day from them.
What to leave out of the script
- Camera directions and shot lists. The storyboard plans its own shots, and you can edit them afterwards.
- Tag lists such as “masterpiece, 8K, cinematic”. The models read sentences; the tags add nothing and can be painted literally. No magic words.
- Negatives. “No crowd” puts a crowd on the list. Write the empty street. Say what is there.
- Several stories at once, and long backstory. One script, one story; the read invents what it needs.
- Real people and brands. Keep to invented names.
The second place to steer: the storyboard
After the read each episode is turned into a storyboard on demand: segments of three to five numbered shots, each with its seconds, the people and props as @ chips, and the lines with a voice. This is the document the video model actually sees, and editing it is free. A shot has the shape below, which is an example of the form, not a recording of output.
[Shot 2] 3s: Close-up on @Mara, lit from a window on the left. She turns the key over in her palm. @Mara (quietly): It still fits.
One action per shot, the light by its source, who is in frame by name. Generate the storyboard for an episode · How to write one shot.
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 Parisian Streets; Costume Style Smart Casual; Visual Style Filmic Live-Action; Target Region; dialogue language; 16:9 Landscape. Then read the script.
- Step 2: the translator and the diplomat as the two leads; the bookshop and the quay as locations; the key as a prop. 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.
| Layer | Takes in | Hands back | Where you steer |
|---|---|---|---|
| The read | Your script or story | Summary, episode list, cast, places and props as words | The text; one name per person; dialogue as lines |
| The cast | Those descriptions | A picture for each character, location and prop | Edit a description; replace a picture |
| The storyboard | An episode outline and the cast | Segments of numbered shots with seconds and lines | Edit the shot text; free |
| The segment | One segment’s text and pictures | One clip | The tier, the seconds, a regeneration |
FAQ
Is there a prompt box for generating videos with AI?
No single one. You write or paste a script, a read turns it into a cast and episodes, and each episode becomes a storyboard whose shots are what the video model sees.
What should I write to generate a video with AI?
A script: scenes with places, characters named the same way each time, and dialogue as Name: line. Leave out camera directions, tag lists and negatives.
How long can the script be?
From about a paragraph up to a hard ceiling. Above about 10,000 counted characters it is read in two passes, up to 100,000 counted characters.
What does the read cost?
4 credits per 1,000 counted characters, charged once, and refunded if the read fails. The first drama's episode-1 storyboard is free.
Why was my dialogue turned into a summary?
Speech folded into narration is paraphrased. Write each spoken sentence on its own line as Name: words.
Why did one character become two?
The read matches names literally. Use one spelling for each character and each place.
Can I paste a link or a PDF?
No. Links are not fetched and PDFs are refused. Paste the text, or export to text or Word.
Can I edit what the read returns?
Yes. The summary, descriptions and storyboard text can all be edited for free; if the episode split is wrong, fix the text and read again.
Paste one scene and read what comes back
The read costs a few credits for a page; the storyboard of your first episode is free.
Updated November 2, 2026