What cinematography terms work best in AI video prompts for Runway, Kling, and Sora?

The cinematography terms that consistently land across Runway, Kling, and Seedance 2.0 fall into six buckets: shot size (ECU, MCU, wide), camera movement (slow dolly in, tracking, crane up, whip pan, handheld), angle (Dutch, low, bird's eye), lens (35mm, anamorphic, shallow depth of field, rack focus), lighting (golden hour, rim-lit, overcast soft, motivated practicals), and grade (teal-and-orange, desaturated, high contrast). Specificity beats the word 'cinematic' every time.

Write prompts as a director writes a shot card — subject, action, then the six cinematography layers in this order: shot size → camera movement → angle → lens → lighting → grade. Replace vague modifiers like 'cinematic' or 'epic' with the precise term. Before: 'cinematic shot of a woman walking through a market.' After: 'medium close-up, slow dolly in on a woman walking, eye-level, 35mm shallow depth of field, golden-hour backlight with rim on her hair, desaturated teal-and-orange grade.' The model now has something to render.

Shot size — extreme close-up (ECU), close-up (CU), medium close-up (MCU), medium shot (MS), wide shot (WS), extreme wide / establishing shot (EWS). Name the size; don't make the model guess from subject distance.

Camera movement — slow dolly in, dolly out, tracking shot, crane up, crane down, whip pan, tilt up/down, pan left/right, handheld, gimbal-smooth, locked-off static. The most common beginner mistake here is using 'zoom in' when you mean 'dolly in.' Dolly moves the camera through space and changes parallax; zoom changes focal length without moving. Models read these as different physics — say which one you actually want.

Angle — eye-level, low angle, high angle, Dutch angle, bird's eye / overhead, worm's eye, over-the-shoulder. Pair the angle with intent in your head, but write only the term.

Lens behavior — 35mm, 50mm, 85mm, anamorphic, spherical, shallow depth of field, deep focus, rack focus, lens flare, Black Pro-Mist diffusion, circular bokeh. Lens vocabulary is the highest-leverage layer most users skip; '85mm shallow depth of field' restructures the whole frame.

Lighting — golden hour, blue hour, overcast softbox, harsh midday sun, neon-lit, rim-lit silhouette, backlight, motivated practicals, hard side light, 85:15 dark-to-light ratio for low-key horror. Name the source and the quality, not just 'moody.'

Color and grade — teal-and-orange, desaturated, high contrast, split-tone amber and emerald, milky lifted blacks, crushed blacks. Naming a palette mode locks the look.

Director and DP name-drops as shorthand — 'shot in the style of Roger Deakins' or 'Wong Kar-wai color grammar' compresses dozens of parameters into one phrase. Use sparingly and with one other concrete term, or the model defaults to surface mimicry.

How the three models react to the same vocabulary. Runway responds strongly to motion primitives and physics verbs — lead with the camera move and the force in the frame. Kling 3.0 handles structured shot descriptions and multi-shot sequences natively, so longer scene-level prompts with subject + action + lens + light hold together across cuts. Seedance 2.0 reference-to-video carries character and location context across clips, so cinematography terms work best when paired with a reference image and a short, declarative shot description rather than a long brief. invideo holds all three models behind one agent, so you write the shot in director language once and the invideo agent routes it to the model whose strengths match the shot — Runway for force-driven movement, Kling for multi-shot scenes, Seedance 2.0 for reference-continuous takes.

The on-set language angle. Treating these terms as how you'd talk to a DP on set — not as keywords — is what makes them work across a pipeline. When the invideo agent holds your shot list, the same cinematography vocabulary you wrote for the storyboard agent flows into a DOP agent per scene and out to the generation model without translation, which is why naming a 'creative producer agent' and a 'DOP agent' with this shared language travels cleanly between roles. As Hridaye, invideo's creative director, puts it: "Pretty much exactly like how I would talk to my DOP on set or how I would talk to my DA on set."

One sanity check before you generate — read your prompt and confirm you've answered four things: what is the subject, how does the camera move, what is the light doing, and what motion is happening in the frame. If any layer is missing, the model fills it with its average — which is what produces the generic 'AI look.'

Watch some of these to see what works for you:

Pretty much exactly like how I would talk to my DOP on set or how I would talk to my DA on set.

— Hridaye, invideo's creative director