Text2Movie

The film language each AI video model's docs let you write

Directing notes: writing shots for AI video

The notes cover choosing between prompt words and presets, planning cuts, pinning frames and what to do when the docs are silent. As of 2026-09-25.

What this reel holdsShort notes on writing camera language for AI video models, grounded in their docs. The reel holds 12 pages.ReelDirecting notes12 pagesEach pageOne question, answered from the docs.Each quoteDoc page and day read.
Fig. 1 Each page on the reel links back here.

Every note takes its examples from the register, each linked to the doc it comes from.

No model was run; the notes rest on the docs alone.

1On this reel

  • Framing words first — Before any camera move, a shot needs a size and an angle: close-up or wide, low or high. They cost nothing and change what a shot means.
  • How much should move — A motion setting changes the amount of movement without rewriting the prompt. Six models document one.
  • Menus for shot size and angle — Three tools put shot size and camera angle in menus rather than in the prompt: Adobe Firefly, Leonardo Motion 2.0 and LTX Studio. A menu makes the choice repeatable, since the same pick sends the same request every time.
  • Motion scores in open models — Two open models expose a single number for how much the picture moves: Step-Video-TI2V's motion_score, 5 by default, and Open-Sora's motion score, 4 by default. Both are written into the request rather than described in words.
  • Moves from references — A few models take a camera move from something other than words: a reference video, a single image or a drawn path.
  • One word, many models — Close-up each appear in the docs of ten models, more than any other camera words on the site. Terms like the whip pan appear in only two.
  • Pinning the last frame — A last frame fixes where a shot ends, which makes the next shot's start predictable. Most models on the register take one in some form.
  • Planning cuts inside one clip — A few models let one generation hold several shots, numbered with their own durations. That keeps light, cast and place consistent across the cut, because one pass makes every shot.
  • Prompt words or presets — Most models take camera moves as words in the prompt; a few offer a fixed list of commands or presets. Words are open-ended and hard to check; a preset does one thing the maker has tuned.
  • Reading a prompt guide — A prompt guide shows what a maker expects a user to write. The examples are the most useful part, because they show which words the maker trusts.
  • Saying what not to show — Four models document a negative prompt field for video; one maker advises against negative prompts altogether.
  • When the docs are silent — Many points on the register read “None in the docs”. That records the documentation, not the model's ability.

2The film language, written for you

Where “Directing notes” has to be written by hand, SceneMixer's pipeline writes the shots; its guides explain:

The samples page for 15 languages lists one finished series per language, each with the cast speaking that language natively and every segment's storyboard shown next to its video.

SceneMixer, AI Short Drama in 15 Languages (2026): Interface, Scripts and the Dialogue Your Cast Actually Speaks — “Can I watch a sample in my language before I start?” (read 2026-10-13)

It rewrites one episode's storyboard from its outline, discarding that episode's segments and videos.

SceneMixer, Checking the Summary, and What “Re-parse” Means — “What does Re-parse in step 3 do?” (read 2026-10-13)

3Other reels