# Turn a script scene into a Blender previs with real cameras and sound

Go from a written scene to a slated, cut and voiced previs in Blender: a cinematographer's shot list, matched coverage, a jump-cut check and a measured mix.

Author: Robert Torres

Canonical: https://voyager.so/blog/script-scene-to-blender-previs

Published: 2026-10-19T09:52:00-07:00
Updated: 2026-10-19T09:52:00-07:00

A previs is only useful if a director could shoot from it. If the coverage doesn't cut together, the voices sound wrong or the café crowd drowns the dialogue, it is a grey cartoon, not a plan.

For this example, we staged the café scene from chapter 1 of Hemingway's *The Sun Also Rises*: Jake, Robert Cohn and Frances at a terrace table at night, a kick under the table, then the walk to the corner kiosk. The result is 23 shots and 2 minutes 20 seconds, with every frame slated like a real setup.

![A grey Blender previs of a Paris café terrace at night: a crane down to three mannequins at a table, matched over-the-shoulder shots, close-ups, then two men walking to a corner kiosk, each frame slated with its setup, lens and rig.](https://cdn.moda.app/voyager/blog/script-scene-to-blender-previs/hero-81d5207d14d825bb.webp)

[Video: A grey Blender previs of a Paris café terrace at night: a crane down to three mannequins at a table, matched over-the-shoulder shots, close-ups, then two men walking to a corner kiosk, each frame slated with its setup, lens and rig.](https://cdn.moda.app/voyager/blog/script-scene-to-blender-previs/film-8abbe3d5c15464c4.mp4)

23 shots, 960×540 at 24 fps, rendered in Blender Workbench. The dialogue is the book's, voiced with generated voices. Sound on.

Dialogue and voice-over from the novel's café scene with Jake, Robert Cohn and Frances: Jake mentions a girl he knows in Strasbourg, Frances kicks him under the table, and Jake suggests Bruges or the Ardennes instead. Jake and Cohn walk to the corner kiosk, argue about Frances and plan a trip to Senlis, and Cohn walks back to the café. Under the dialogue: café crowd and street sound.

**Write the shot list the way a cinematographer would, then check every cut against it.** Our first cuts went from a wide crane shot to a wide of the same table from nearly the same angle. That is a jump cut, and nothing in a loose shot list stops it. Real setups, matched coverage and a numeric cut check do.

## What you need

* **Skills:** `edit-in-blender`, `voiceover`, `make-sound` and `gpu-machines`.
* **Plan:** Starter or higher, for the voices and sound. Drafts render locally. Voyager GPUs allow one GPU on Starter, two on Creator and four on Ultra; our four-GPU final render needs Ultra or higher.
* **Tools:** Blender with the Workbench engine, and FFmpeg for the slate and the mix.
* **Made in Voyager:** this example is a real Voyager run of the workflow, from the shot list to the slated, mixed cut.

## Pick one scene and its geography

Start with one scene of a few pages, not the whole script, and a simple one first: two or three people in one location. Our crowd-and-animals scene took far more rounds than the dialogue scenes. Read it for the place as well as the lines: who sits where, which way people walk and where the scene ends up. Our text came from [Project Gutenberg](https://www.gutenberg.org/ebooks/67138), where the novel is public domain in the US.

Fix the geography before any camera exists. Here Jake sits on the west side of the table and Cohn on the east, with the street side as the line: Jake stays screen-left and Cohn screen-right throughout, and they walk east to the kiosk. Draw it as an overhead plan with every camera marked, the diagram a DP keeps with the shot list.

## Write a real shot list

Decide what the scene is about before choosing a single angle: whose eyes we watch it through, and the two or three moments it turns on. A list with one shot per line of dialogue comes out wooden; ours only got good once it was built around a few moments and one point of view.

Choose a camera package, then number setups by scene and letter, the way a production does: 1, 1A, 1B. Give each one a lens, height, support, what it covers and why it is there. We shot on an ALEXA Mini with Cooke S4/i spherical primes, mostly on sticks with a few Technocrane moves, for a classic Hollywood look.

```text
Write a shot list for [scene] the way a cinematographer would.
Camera: [body and lenses]. Mostly [support], with [moves] where they
earn it. Number setups by scene and letter. For each: shot size,
lens, camera height, support and move, what it covers and why it
cuts here. Start from whose scene it is and the moments it turns
on; don't give every line its own shot. Keep the 180-degree line:
[character] screen-left, [character] screen-right. Match the paired
over-the-shoulders. Judge honestly whether it's good storytelling,
then show me the list before you build anything.
```

Match paired coverage exactly. The two over-the-shoulders at the table use the same lens, distance and height, so they cut together as one conversation.

![Two slated previs frames side by side: an over-the-shoulder close-up on Jake and the matching reverse on Cohn, both on a 40 mm lens at 1.2 m.](https://cdn.moda.app/voyager/blog/script-scene-to-blender-previs/matched-ots-8ef599278f2326a2.webp)

Setups 1C and 1E: matched over-the-shoulders on a 40 mm lens at 1.2 m. The slate on every frame carries the setup, lens, support and camera height.

Leave handheld out unless the moment needs it. Our handheld inserts read as jarring next to the locked-off coverage, so they went on sticks or a slow dolly.

## Audit the list against stills

Before building the whole film, render one still from every camera and audit the list against those frames with a written checklist. Ours came from research passes on film grammar, classic scene breakdowns, framing rules and previs practice, and it is what turned a passable list into usable coverage. It caught a 28 mm lens the Cooke set doesn't make, every table camera sitting slightly above seated eye level, a reverse pair that didn't match, and the scene's turning point shot in the same loose over-the-shoulder as the small talk.

```text
Before building the full previs, render one still per camera and
audit the shot list against them with this checklist. Fix every
finding before you build.
Grammar: one line of action per beat, every camera on one side.
Establish each location once and reuse the master. Easy talk in
matched over-the-shoulders, rising tension in clean singles, the
tightest frame saved for the peak. Don't hold a two-shot through the
dialogue: cut into close-ups on the lines. Reverse pairs match lens,
distance, height and subject size. A silent character gets a clean
single, cut to on beats, not lines. Inserts come from a look: look,
see, react. Every cut changes angle by 30 degrees or more or clearly
changes size. Moves only when motivated. Walk-and-talks lead
head-on, and the pair never swaps sides. End on the watcher's look.
Cameras only where a crew could stand. A dialogue close-up sits near
the other character's eyeline, about 30 degrees off, with the actors
cheated toward camera.
Framing: eyes 28 to 38 percent from the top of frame. Over-the-
shoulder foreground 20 to 33 percent of the width, never over eyes or
mouth. Camera at eye height (seated 1.2 m, standing 1.6 m) unless a
high or low angle is labelled. No lens under 32 mm on a close shot.
No other head overlapping the subject's. Only focal lengths the lens
set actually has.
```

Treat the checklist as a default. A production with its own house style should swap in its rules, but keep the step: audit stills against written rules before anything is built in full. Build the mannequins at real head and body proportions too; oversized stand-ins throw off every framing number.

## Build from one spec, with cameras placed by rule

Put the shot list, camera plan, slate text and audio timing in one spec file that drives the whole build. Place each camera by rule from the characters' marks and the shot size, and fail the build before rendering if a camera is outside the set, inside a wall or across the line. Hand-placed cameras were the biggest time cost in our run; with the rules in place, a later scene needed two rounds of fixes instead of four or five.

Keep the checklist, slate layout and gates as files in the project, and start every new scene from them. Our second scene repeated the first one's mistakes because the lessons stayed in the first scene's folder.

## Check every cut and every frame

Measure each cut. When the same character appears on both sides, the camera angle or the shot size has to change enough to read as a new shot; otherwise flag it as a jump cut. Then check framing in each shot: eye line, headroom and how much of the frame the foreground shoulder takes.

```text
For every cut, list the characters on both sides. Where one appears
in both, report the change in camera angle and shot size, and flag
any cut that changes too little as a jump cut. Then check each shot's
framing: eyes near the top third, headroom, over-the-shoulder
foreground width, and that the subject faces camera, is not blocked
and is not cut off by the frame. Fail on any flag, and on any scene
running more than 10 percent over the shot list. Fix every flag
before you show me the cut.
```

Make the check fail, not just measure. A table of framing numbers doesn't stop a close-up of the back of a head or an actor hidden behind a car; a failed check does.

Burn a slate into every frame. It turns review notes into specific fixes and makes the previs readable to a crew. This is our default layout; swap in your production's own if it has one.

```text
Slate every frame in two bars. Top bar: title, scene name, previs
version and date on the left; timecode in the center; shot code
(SQ0100_SH0010), scene and setup on the right. Bottom bar: shot size
with a short description, camera body and lens, support and move,
camera height, and the frame counter for the shot (0030 / 0311) on
the right.
```

## Cast voices from descriptions

Write a short description of each character's voice and generate a few auditions per character. Pick one each, then use those takes as references for the dialogue, so every line keeps the same voice. Start every voice from your description rather than a preset: preset voices, and voices without references, drifted between lines and didn't fit the characters.

```text
Cast the voices from written descriptions. For each character,
write a voice description (age, accent, pitch, manner) and generate
three auditions. Let me pick, then use the picks as references for
every line, recording each scene as one multi-voice dialogue take.
Don't use preset voices.
```

Reuse each character's reference voice in every scene they appear in. Then mix to a measured target, not by ear: our first mix had the café crowd far louder than the dialogue, and measured loudness put it back underneath. Transcribe the final mix and check every line is there; one pass had dropped most of a line.

## Render fast and keep it grey

A Workbench render keeps iteration cheap: the whole 2-minute-20 cut, 3,354 frames at 960×540, rendered in under two minutes. That made it practical to go through twelve versions.

For the final look, later scenes moved to grey clay in Cycles, with a sun, shadows and lens depth of field, rendered on Voyager GPUs. A 2-minute-16 scene took 19 minutes on four GPUs and about 330 credits.

We also tried restyling this multi-shot dialogue scene into drawn or photoreal footage with video models. Faces and costumes drifted between shots and figures grew extra limbs; a photoreal pass of 21 shots ran about 5,300 credits. For a dialogue scene like this one, the clean grey previs answered the brief better.

Start with one short scene, and **cut it on paper like a cinematographer before Blender opens.**


## Voyager is a desktop AI agent for creative work

Describe what you want to make; Voyager plans it, uses your apps and tools, and keeps everything editable.

[Sign up](/sign-up)