Light and colour
Lesson 6 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Hard and soft light
The first thing that tells the brain “this was filmed, not drawn” is light. Light has a character. A small source (midday sun, a bare bulb) gives hard light: crisp shadows, strong contrast, a dramatic look. A big source (overcast sky, a curtained window, a softbox) gives soft light: gentle shadows and flattering skin. By default, models often light everything evenly from everywhere, and that’s the classic “AI look”. So always describe the light in words: “soft diffused window light” or “hard midday sun, crisp shadows”.
Time of day
Photographers plan shoots around golden hour: roughly the hour after sunrise and before sunset. The sun is low, the light is warm and soft, shadows are long and skin glows. Terrence Malick shot much of “Days of Heaven” (1978) at exactly that time, and the film won the Oscar for cinematography. Blue hour is the twilight right after sunset: the sky turns cool and windows light up. Noon means light from above, short harsh shadows and eyes lost in shadow. A real crew gets maybe forty minutes of golden hour; in AI it’s free: just write “golden hour”.
Three-point lighting
This is the basic setup for any interview or ad. The key light is the main one, usually 45° to the side: it shapes the face. The fill shines from the other side, weaker, and decides how dark the shadows are: little fill means drama, lots of fill means a bright ad. The back light hits from behind and draws a bright edge on the hair and shoulders so the hero doesn’t blend into the background. The second trick is motivated light: the light in the frame has an obvious source. Kubrick shot scenes of “Barry Lyndon” by real candlelight on an ultra-fast lens, and it still looks alive. In the prompt, name the source: a window, a neon sign, a phone screen.
Skin without plastic
The biggest giveaway of AI people is skin: too smooth, even and shiny, like a doll. Models learned from tons of retouched photos, so words like “perfect”, “flawless”, “8k” and “masterpiece” pull the image toward gloss and a 3D render. People who make realistic generations do the opposite: they ask for imperfection. Write “natural skin texture, visible pores”, add small details like freckles or slight redness, and top it with “subtle film grain” or “shot on 35mm film”. Film grain works like glue: it hides small artifacts and makes the frame feel filmed.
Colour grade and palette
Colour grading is the overall tint of the whole video, and it sets the mood before the viewer even gets the plot. Teal & orange (warm skin, teal shadows) is the language of blockbusters: skin pops against the background. Warm film means nostalgia and summer. Neon night means energy and the city. Muted colours mean calm and a premium brand. The AI rule: pick 2–3 colours and write them the same way in every shot’s prompt. Otherwise one shot comes out blue, another yellow, and the ad looks like a collage of other people’s videos.
What to keep out of the frame
The model sees words, not negations. Write “no neon signs” and neon may appear, because the word “neon” is already in. Google’s prompt guide for Veo advises against writing “no” and “don’t”: you simply list unwanted things in the negative prompt. So in the main prompt, describe what you need: not “no signs” but “a dark residential street with warm streetlamps”. For junk there’s a separate negative prompt field, which Kling and Wan have, for example: that’s where plastic skin, scribbled letters, extra fingers, blown highlights and a cartoon look go.
A mood board instead of words
Words like “cosy” and “stylish” mean something different to everyone, but a picture is clear to all. Pros start with a mood board: 4–6 reference frames, a palette of five colours and a couple of words about texture. Mira made hers for the hoodie drop: concrete, wet asphalt, a low sunset from the side, film grain. Then a simple trick: first generate one perfect still in an image model, then bring it to life with image-to-video. The light and colour are already locked in the frame, and the video keeps them. Use your own references or ones you have rights to, so you don’t copy someone else’s ad shot for shot.
Check yourself
- Mira is shooting the hoodie on a rooftop, but the video looks flat and “too AI”. What should she add to the prompt first?
Correct answer: “golden hour, low warm sun from the side, long shadows”. Flatness comes from light coming from everywhere. Low directional light instantly adds depth, shadows and warm skin. - The face close-up looks like a plastic doll. What should change?
Correct answer: Remove “flawless”, add “natural skin texture, visible pores, subtle film grain”. Plastic comes from words about perfection. Asking for pores and grain brings back the texture of real skin. - The ad has three shots: one bluish, one yellow, one neutral. Together it looks like a collage. How to fix it?
Correct answer: Copy one light-and-palette block into all three prompts. A video feels like one piece when every shot shares the same light and palette. The easiest way is to keep that text identical. - Mira writes “no neon signs, no text”, and the frame has even more neon lettering. What should she do?
Correct answer: Describe what she wants: “dark residential street, warm streetlamps”, and put neon and text in the negative field. The model latches onto the words “neon” and “text”. Describing the scene you want plus a separate negative field removes the extras.