How do you describe lighting in AI video prompts to get cinematic, consistent results?

Describe lighting in four ordered layers — source, direction, quality, and color palette — and write that exact block into every shot prompt across the scene. Lock it once at project start (named tonal modes with hex values work best), then carry the same descriptor verbatim through every generation so the look doesn't drift.

Write your lighting block in this fixed order — source → direction → quality → palette — and put it in the same slot of every prompt:

1. Source (the motivated light). Name the actual light in the world: "warm yellow from the practical lamps only," "midday sun through a north-facing window," "single neon sign camera-left." One motivated source per shot. Hridaye, invideo's creative director, calls this out directly: "Lighting corrections should explicitly reference the source material rather than using generic descriptors — specifying 'warm yellow from the lamps only, like all the refs' produces more accurate results than general 'warm lighting' prompts."

2. Direction. Where the light comes from relative to subject: backlit, 45° key from screen-left, top-down overhead, rim from behind. Avoid stacking conflicting directions in one prompt — that's the single biggest cause of flat, muddy AI output.

3. Quality. Hard vs. soft, diffused vs. specular, contrast ratio. This is where the cinematic vocabulary lives — high-signal terms the models respond to: chiaroscuro, soft diffused fill, hard key with deep falloff, volumetric god rays, low-key 85:15 dark-to-light, golden-hour bounce, sodium-vapor wash. Specify the ratio when you want drama: "85:15 dark-to-light, key only on the face, rest in shadow" is more controllable than "moody lighting."

4. Palette anchor. Lock the color in hex or named tonal modes, not adjectives. Encode it once — e.g., "Mode A — split-toned amber (#D89B5A) and emerald (#2E5F4A), 3200K key, 5600K practicals" — and reuse that exact string. Generic "warm tones" drifts every generation; a named mode with hex values reproduces.

The consistency rule that makes this hold across a scene: write the four-layer block ONCE, then paste it verbatim into every shot prompt in that scene. The Arcane-style production proved this — the team wrote one style-and-lighting block and every prompt after that started with it, which is why 41 final clips cut together as one continuous look. Switching wording mid-scene ("warm lamp light" in shot 1, "amber glow" in shot 4) is what causes drift.

For cross-scene consistency, go one level up: write a short lighting bible — your motivated sources per location, your contrast ratio per emotional stage, your palette modes with hex — and load it once into the invideo agent at project start. invideo is an agentic video creation platform that holds project context across every shot, so the lighting grammar stays locked without you re-pasting it. A documented horror short used exactly this — an 85:15 dark-to-light ratio written into the loaded doc, and the agent flagged a scene where "shadows were leaning blue-green instead of neutral gray" against the locked rule and offered a corrected warmer pass without being asked.

Which model handles your lighting cue depends on the look. Veo holds soft naturalistic daylight and skin tones cleanly; Kling renders hard contrast and stylized neon well; Seedance 2.0 carries lighting context across reference-to-video segments, which is what you want for matched coverage in the same scene. You don't pick a platform per model — every roster model is available inside invideo and the invideo agent routes the shot to whichever one fits the lighting brief.

Before/after the framework. Flat prompt: "woman sitting at a table, warm lighting, cinematic." Four-layer prompt: "woman at table, single tungsten practical lamp camera-right as motivated key, 45° from subject, hard quality with deep falloff into shadow, 85:15 dark-to-light ratio, palette locked to amber #D89B5A key and teal #1F3A44 shadow, 3200K, no fill." The second one reproduces; the first one rolls dice.

Pitfalls to cut from your prompts:

• Conflicting sources ("sunset window light AND overhead fluorescent") — pick one motivated source.
• Missing motivation — "moody blue light" with no in-world source reads as a filter, not lighting.
• Over-specifying Kelvin without specifying mood — "5600K" alone is colder than you think; pair it with the contrast ratio and palette.
• Using illustrated or animated frames as direct reference attachments — those don't transfer as lighting. Instead, ask the invideo agent to read the colors and textures off the reference and prompt for those, which a documented production confirmed: "the gens came back hyper-realistic with the exact colour temperature I was looking for."
• Changing the descriptor wording shot-to-shot in the same scene — that's drift you're building in yourself.

These are the levers that consistently produce cinematic, on-model lighting — what exact values you set depends on the film.

Watch some of these to see what works for you:

Lighting corrections should explicitly reference the source material rather than using generic descriptors — specifying 'warm yellow from the lamps only, like all the refs' produces more accurate results than general 'warm lighting' prompts.

— Hridaye, invideo's creative director