AI VIDEO QUALITY

Design shots the model can actually make

The cheapest slop to fix is the slop you never ask for. Write the script around what the model does well, and keep every shot small enough to judge.

Figure 1: One character, one line, one small motion, no prop in hand: the shot that works.

1. One character, one line, one small motion

Image-to-video models are good at one thing in particular: bringing a single, well-lit character to life for a few seconds. They are bad at keeping several things true at once. So the shot rule is strict:

  • One character in frame for any talking shot. Two people talking becomes two close-ups, because the model cannot reliably keep the listener’s mouth shut.
  • One spoken line per clip, at most.
  • One small motion: a nod, a lifted chin, eyes turning. A shoe flying off and landing is two shots.
  • No prop in the hands unless the story needs it. If a prop matters, put it in the still and describe it as unchanged.
  • Seats visible under anyone sitting, or they float.
  • Animals as still inserts, not acting.

2. Clip length = the line plus a breath

Idle seconds are where drift lives. A character holding still for six extra seconds starts to melt, float or wander. Make each clip about as long as its line plus one to one and a half seconds. Trim each finished take to its line or action: dialogue shots usually land at two and a half to three and a half seconds, action shots at one and a half to two and a half.

Trimming also fixes runtime without cutting story. In this project the first episodes ran 40% long simply because every take was left at full length.

3. A prompt template that worked

The approved clips used prompts of 85 to 135 words with every sentence kept, every time:

[Character] [one small motion] and says [delivery word]: “[EXACT LINE]”. In [fixed voice description]. Complete exact words; native speech with visibly synchronised mouth; mouth closed before and after the line. Eyes [direction]. Camera locked off on a tripod, completely static. Keep face, hair, costume and background exactly as in the first frame. Only [Character] in frame; nobody else appears; the edges of the frame stay empty. Hands relaxed with five fingers; nothing new appears in anyone’s hands. Sound: dialogue and quiet room tone only, no music.

Three details matter more than they look:

  • Copy the voice description word for word on every clip of that character. A paraphrase is a different voice.
  • Never write a bare “quietly”, “softly” or “whispers”. The model takes it literally and you get a mumble nobody can hear. Write “in a clear, fully audible stage-whisper”.
  • Ask for no music in the clip. Music is laid once, in the edit, where it can be lowered under speech.

4. Never ask the video model to write

Writing on signs, notebooks and posters is where generated video fails most visibly. In this project the model invented a readable chore list (“Feed Sparky, Practice piano”) for a shot whose script only said “nineteen steps, all crossed out”. Put any writing the story needs on the still with an image edit, check it letter by letter, or add it in the edit. Lesson 7 shows how to catch the rest automatically.

Layman rule: If a shot needs two things to stay true at once, it is two shots.