Motivated lighting: name the source in frame

Two prompts for the same shot. The first asks for "a man at a desk at night, moody lighting". The second asks for "a man at a desk at night, lit by the green glow of the monitor in front of him". The second one comes back with shadows that agree with each other, a face lit from one direction, and a room that falls off believably behind him. The first comes back looking graded rather than lit.

The difference has a name in production: motivated lighting. Every light in the frame is explained by something that exists in the world of the shot — a lamp, a window, a screen, a fire, a passing car. On a set that discipline is about believability. In generation it does something extra: it gives the model a position to compute from, which is the one thing a mood word can never provide.

Why a named source outperforms a described quality

A model deciding how to light a frame has to place shadows, falloff, catchlights in the eyes, and the brightness of every surface. If your prompt supplies a source, all of those follow from one place — the shadow goes opposite the lamp, the near cheek is brighter than the far one, the eyes carry a reflection of the lamp. If your prompt supplies only an adjective, each of those decisions gets made semi-independently, and the results disagree.

That is the mechanism behind the most common lighting complaint: renders that look flat and sourceless. They aren't underlit. They're lit from everywhere, because nothing said where the light lives.

The source doesn't have to be visible in frame — it has to be nameable and locatable. "A window just off frame at camera left" is motivated. "Warm light" is not. The sentence needs an object and a position; everything else resolves from there.

A named source also buys you four properties for free. harsh fluorescent overheads carries size, direction, colour and a whole genre of interior in three words. That is why one concrete source beats a stack of style adjectives — the lighting prompt formula is the longer version of the argument.

The tells that light is unmotivated

Four things to look for in a render before you accept it:

1. Disagreeing shadows. The subject's shadow goes one way, the object on the table shadows the other. Two sources that don't exist.
2. Catchlights pointing nowhere. The reflections in the eyes sit in a position no source in the frame could occupy. Easiest to spot in a close-up and hard to unsee afterwards.
3. A lit background with nothing lighting it. The wall behind the subject is evenly bright in a scene whose only stated source is a small lamp. Falloff has been ignored because no distance was implied.
4. Even light at night. Both sides of a face equally lit in a dark room. The model needed a key, wasn't given one, and invented ambient light from nowhere.

All four have the same cure: name one dominant source, put it somewhere, and let the rest of the frame be fill or fall away.

Interiors: the window is the motivator

For daytime interiors the window does almost all the work, because it is the source viewers already expect and the one the model has the most examples of.

Late afternoon, small kitchen. Light comes from a single window at
camera left, low and warm, throwing a long soft shadow of the window
frame across the floor. The right side of the room falls into shade,
with only bounce from the white wall filling it.

Three things in that block are doing real work. The window is placed (camera left). Its effect is described (a long shadow of the frame across the floor) — a consequence the model can render even where it has no explicit lighting parameter. And the fill is motivated too: bounce off a named white wall, rather than an unexplained ambient lift.

Motivating the fill is the step people skip. If you name a key and stop, the model will supply its own fill and often supply too much, which is what flattens interiors back out. Say where the fill comes from, or say there isn't any: the shadow side goes almost black.

Night: you have to give it something

Night is where unmotivated lighting fails hardest, because a face still has to be legible and the model will find a way to make it legible with or without your help. Left alone, it invents a soft key from an empty corner and you get a scene that is dark in the grade but not dark in the physics.

Give it a practical. The reliable ones, roughly in order of how consistently they land:

Source	Phrasing	What it gives you
Screen	lit only by the cold glow of the laptop screen in front of him	Hard-ish underlit key, cool colour, strong falloff
Street lamp	sodium street lamp outside the window, orange light through the blinds	Directional slats, warm/cool contrast against interior
Practical lamp	a single warm table lamp at frame right, everything beyond it dark	Classic pool of light, steep falloff
Headlights	passing headlights sweeping across the wall behind her	Moving light, motion built into the shot
Fire	firelight from below, flickering, warm orange	Underlit, unstable, high contrast

Two rules make night scenes hold. Name exactly one dominant source and let everything else be spill from it — night with three equal sources reads as a lit set rather than as night. And say explicitly what stays dark. Phrase that as a state, not a prohibition: no harsh shadows and not too dark are weak or counterproductive, because the thing you named is the thing that gets rendered. Write everything beyond the pool of lamplight falls to black and you have described a frame instead of banning one. The longer version of that argument is in what replaced negative prompts.

Product: name the reflection, not the light

Product work inverts the technique in a useful way. On a reflective surface — glass, metal, gloss plastic — the highlight is an image of the source. So the most precise instruction available isn't "a large softbox at camera left"; it's a description of the reflection that softbox makes.

A tall, soft-edged rectangular highlight running down the left edge
of the bottle, fading toward the base. A second, narrower highlight
on the right shoulder of the glass. Background falls to dark grey.

That specifies size, position, quality and falloff in terms of pixels on the product rather than a fixture the model has to simulate. For anything shiny it is the highest-precision lighting prompt there is, and it is usually what separates a product render that reads as real from one that reads as plastic.

For matte products the reflection trick doesn't apply and you go back to a named source plus falloff. Either way, build the still first and animate from it: a well-lit frame used as a reference image constrains a clip's lighting far harder than any sentence will, which is the working pattern behind the product video generator.

The continuity payoff

Here is the return on all of this: a named source is a thing you can copy between shots, and an adjective is not.

Across a four-shot scene, keep the source sentence byte-identical and change only the camera. Rewriting it as "warm side light" in shot three is exactly how a room's lighting changes halfway through a scene, because a fresh description samples a fresh setup.

The one place this needs care is the reverse angle. When the camera turns to face the other way, the source stays in the same place in the world but arrives from the other side of frame. If the window was at camera left in the wide, it is at camera right in the reverse — and if you paste "window at camera left" into the reverse prompt unchanged, you have moved the window. Write it as two locked variants, one per camera side, and pick the right one per shot. It is the single most common continuity error in generated scenes and it is entirely avoidable.

A practical way to settle the look: describe the source once and fan the same prompt across several named models in one request in the agent chat, then keep the one whose shadows behave. Check the assembled sequence with a preview: true pass — 480p, no credit cost, with a short per-user cooldown between requests, and the final export charged once regardless of clip count. Shadow direction and disagreement between shots both survive the downscale, so the low-resolution pass genuinely answers this question. The previews and final export breakdown has the detail.

FAQ

Does the light source have to be visible in the shot?

No. It has to be nameable and placeable, which is a lower bar. "A window just off frame at camera left" motivates the light perfectly well without the window ever appearing. Visible practicals add something — a lamp in shot gives the viewer a reason for the light — but the prompting benefit comes from naming a position, not from the source being on screen.

Why do my night scenes look grey instead of dark?

Because nothing in the prompt told the model where the light was, so it supplied an even ambient key and the grade did the rest. Fix it by naming one dominant practical — a screen, a lamp, a street light through a window — and stating what stays unlit, phrased as a description rather than a prohibition. The shadow side goes almost black works. Not too bright does not.

What happens to a motivated source on the reverse angle?

It stays put in the world and swaps sides of the frame. A window at camera left in the wide is at camera right once the camera turns around, so pasting the source sentence in unchanged silently moves the window to the opposite wall. Write two locked variants up front, one per camera side, and pick the right one per shot rather than editing the sentence fresh each time.

What's the best lighting instruction for a shiny product?

Describe the reflection rather than the fixture. On glass or polished metal the highlight is a direct image of the source, so "a tall soft-edged rectangular highlight down the left edge, fading toward the base" specifies size, position, quality and falloff in terms of what actually appears in frame. For matte products, fall back to a named source with a stated distance and let falloff carry the depth.