How do you describe lighting in an AI video prompt to get cinematic results?

Write lighting as four ordered attributes: quality + direction + color temperature + motivated source — e.g. "soft diffused key from upper-left, warm 3200K, motivated by a low window". Pair it with a hard key for contrast ("single hard warm key, side-raked, skin rim-lit, environment falling into shadow"), name a real-world reference (golden hour, overcast softbox, neon rim, practical lamp), and lock it on a still before spending video credits.

Start by writing the lighting line as a structured four-part description before you touch any other prompt element: quality (soft diffused / hard / raked / bounced), direction (key from upper-left, side-raked, backlit at 45°), color temperature (warm 3200K / neutral 5600K / cool 6500K, or named — "golden hour", "blue-hour overcast"), and motivated source (window, practical lamp, streetlight, fire, monitor glow). Generic prompts like "cinematic lighting" produce flat output; craft-specific direction like "single hard warm key light, side-raked, skin rim-lit, environment falling into shadow" gives the model something to actually render. invideo's creative director Hridaye puts the principle directly: "Lighting direction specified as a single hard warm key light, side-raked, with skin rim-lit and environment falling into shadow produces editorial-grade separation."

A lighting lexicon worth memorising. Pull from these high-signal phrases in your prompts — each one carries an entire visual setup the model already understands: golden hour (low warm sun, long shadows, hazy backlight), blue hour (cool ambient, no direct sun, soft falloff), overcast softbox (flat shadowless daylight, neutral temperature), low-key chiaroscuro (single hard source, 80% shadow, deep blacks), neon rim light (cyan/magenta edge-light from behind, dark midground), practical lamp motivation (warm tungsten pools, falloff into shadow), high-key fashion (multi-source soft fill, no harsh shadows, even skin), raked window light (hard side light through blinds, slatted shadows), firelight flicker (warm 1800K, animated bounce, low ambient), moonlit cool (top-down soft blue, low saturation).

Always state motivation. A light source needs a reason to exist in frame. "Soft daylight from the left" works; "soft daylight from the left, motivated by an unseen window, with a warm bounce card filling the shadow side" gives you cinematic depth. Hridaye's own direction logs read this way — "a sunny day, slightly backlit to the characters, not harsh sunlight" — the qualifier (not harsh) does as much work as the noun (sunny).

Use negatives to kill the AI look. In the same prompt or in the agent's standing don'ts, exclude the failure modes: flat lighting, blown highlights, mixed color temperatures, evenly-lit faces, generic studio softbox, plastic skin sheen. These five exclusions remove roughly the same defects that make AI footage read as fake.

Lock lighting on a still before you animate. Generate the lighting setup as a single keyframe image first — cheap to iterate, and the still becomes the visual anchor for every video clip in that scene. Once the still reads cinematic, pass it as the reference frame into video generation; the model carries the lighting forward instead of re-rolling it per clip. Across documented productions, video credits should land on locked frames, never on exploratory lighting iterations.

Route to the model that lights best. Lighting quality varies sharply by model: for image keyframes, Nano Banana renders light with the most physical accuracy (soft falloff, accurate bounce, believable shadow density); GPT-Image-2 is stronger for designed/graphic lighting and on-frame text. For motion, Seedance 2.0 holds lighting continuity across a multi-shot single-pass generation, and Kling preserves rim and practical sources cleanly across cuts. Inside invideo, the agent routes each shot to the right model automatically — every roster model is available in one project, so you don't pick a platform per lighting style. Hridaye documents the exact pairing: "build the base image with GPT-Image-2 to get the aesthetic, then run Nano Banana to lock" the lighting and product details.

Store lighting in agent context, not in every prompt. Once a lighting direction is locked for a project, drop it into the invideo agent's context tab as a standing rule — "single hard warm key, side-raked, 3200K, motivated practicals; no flat fill, no mixed temperatures." Every subsequent shot inherits it without re-prompting. In documented productions, this is what stops lighting from drifting between shot 1 and shot 12.

A copy-pasteable template: "[Subject + action]. Lighting: [quality] [direction] [color temperature], motivated by [source]. [One mood/contrast qualifier — e.g. skin rim-lit, environment in shadow]. Camera: [framing + movement]. Avoid: flat lighting, blown highlights, mixed color temperatures."

If the output still reads generic after two passes, don't keep iterating on the same model — render the identical lighting prompt across Nano Banana, GPT-Image-2 and your video model, compare side by side, and pick the model that lit the shot correctly. Iterating one model when the model is the problem wastes credits.

Lighting direction specified as 'single hard warm key light, side-raked' with skin rim-lit and environment falling into shadow produces editorial-grade separation in AI-generated fashion images.

— Hridaye, invideo's creative director