Ideas for a more imaginative world

Cinematic lighting

Why most prompts fail on the third reroll

A prompt that names a look but not the optics will drift every time you run it. Here is what to state instead.

Read the story
Night portrait of a woman in a translucent rain jacket lit by magenta and cyan neon in a wet alley
From the prompt libraryNeon Alley Rain Portrait
Written by
Lokesh Kapoor
Published
Reading time
5 min read / 558 words

Most prompt collections are lists of text nobody has run twice. They read well, they produce something plausible once, and then they drift — because the prompt described a result rather than the conditions that produce it.

The difference between a look and a setup

"Moody food photography" is a result. A model asked for it will pick a lighting setup for you, and it will pick a different one every time.

Single soft window light from the left, deep shadows, no fill is a setup. There is nothing left to choose, so there is nothing left to drift.

Describes a result

Hyper realistic cyberpunk street at night,
neon lights, 8k, trending on artstation,
masterpiece
  • Buzzwords that dilute the latent space
  • No camera, no optics, no light source
  • Fifteen re-rolls before one usable frame

Describes a setup

Cinematic 35mm anamorphic wide shot of
{{location}}, rain-slicked asphalt reflecting
cyan neon signage, low-angle tracking,
shallow DOF, Cooke optics, Kodak 500T grain
  • Exact optics and film stock, stated
  • One variable left open, on purpose
  • Reproduces on run number one

What to state

In order of how much they matter

  1. The light

    How many sources, how hard, and from where. One named source beats any adjective you can reach for.

  2. The optics

    Focal length changes the relationship between subject and background more than any styling instruction will. 85mm and 35mm are different photographs of the same scene.

  3. The surface

    On anything reflective — glass, metal, wet asphalt — the surface is the image. Name it before you name the subject.

  4. What to exclude

    A negative prompt naming the specific failure is worth more than three more adjectives. Blown highlights, distorted hands, HDR halos.

The test

Run your prompt three times. If the three outputs differ in ways you did not choose, the prompt is underspecified.

If you cannot predict what changes between runs, you have not written a prompt. You have written a wish.

Does this apply to video prompts too?

More so. A video prompt has every variable a still has, plus duration, camera motion and continuity — and a loop that drifts is visible on every repeat rather than once.

Do longer prompts always beat shorter ones?

No. Length is a side effect of specificity, not a substitute for it. A short prompt where every clause is a decision beats a long one padded with adjectives.

What about negative prompts?

Name the failure you actually see, not a generic list. "Blown highlights" is useful on a packshot and noise on a landscape.