Learn AI Filmmaking Day 2 — Light, Lens and Look
Day 2 of 7: a change of light is a change of state
Series note. This is the second of seven articles that teach AI filmmaking from the ground up. Article 1 built the atom, a single shot with a size, an angle and one motivated move. This article keeps the single shot and changes only the light, because light is the fastest way to tell a viewer that something has changed.
The film this article builds: 15 seconds from a single generation. Watch it with sound.
The brief
Before film had lights it had the sun. Thomas Edison’s first studio, the Black Maria, went up in West Orange in 1893 as a tar-paper shack on a pivot, with a hinged roof, so the whole building could be turned through the day to keep sunlight on the stage. Light was the first production problem in the history of the medium, and every crew since has spent most of its day solving it.
For us the problem is different, though. A model does not need a lamp; it needs the words for one. The good news is that a century of cinematography gave us a precise vocabulary, and the models were trained on films that used it.
So here is the brief. We are going to make a fifteen-second film, one shot, camera nearly still, with a single change of light in the middle of it. Its logline: a man throwing the most careful candlelit dinner of his life flips on the string lights and reveals that the guest of honor is his golden retriever, and this is her adoption day.
The joke is built entirely out of light. In the dark the viewer sees a romantic dinner; in the light they see who it is for. Figure 2.1 shows both states side by side, and the rest of this article explains how to write each one.
Figure 2.1: Light is a state. The same room, the same table, two different stories.
Light asks three questions
Cinematographers argue about lighting for hours, but the argument always comes down to three questions. Where does the light come from, how hard is it, and what color is it. Answer those three for every shot and you have described the light well enough for a crew or a model to build it.
Source. Source is where the light comes from and whether the viewer can see why. A light that has a visible or implied reason inside the scene (a window, a lamp, a candle, a phone screen) is called motivated, and motivated light is the single biggest difference between footage that looks filmed and footage that looks rendered. Our film has two sources, one candle and then a string of bulbs, and both are in the frame.
Quality. Quality is how hard the shadows are. Hard light from a small source (the sun, a bare bulb) gives sharp-edged shadows and texture; soft light from a large or diffused source (a window with a curtain, an overcast sky, a wall of string lights) gives gentle shadows and flattering faces. A single candle is small but close, so its light is soft on the hands and falls off fast into darkness, which is exactly the shadow we want to hide a dog in.
Color. Color is temperature, measured in kelvin, and it is one of the most reliable words a model understands. Figure 2.2 puts the three questions and the kelvin scale on one page, and you will use that scale in every prompt for the rest of the series.
Figure 2.2: The three questions of light, with the kelvin scale every prompt should use.
Remember, adjectives drift and numbers do not. “Warm” can mean a sunset or a fireplace, but “2700K” means one thing, and the current models honor it. As a general rule, name the source, name the quality, and put a kelvin number on the color, and you will get the light you asked for far more often than not.
The lighting vocabulary
Film built a vocabulary of named lighting styles on top of those three questions, and most of the names are a hundred years old. The story goes that Cecil B. DeMille had to invent the phrase “Rembrandt lighting” around 1915 to explain to a nervous business partner why half an actor’s face was in shadow on purpose. The name stuck, and it still works in a prompt today.
Figure 2.3 lays out the dozen styles you will actually use, with the mood each one carries. Read it as a menu, not a syllabus; a film usually commits to two or three of these and stays there.
Figure 2.3: The lighting vocabulary, twelve named styles and the mood each one carries.
Our dinner uses two styles in sequence. The first half is low key and practical, one candle doing all the work, and the second half is high key and practical, string lights doing all the work. Both halves are motivated, which is why the change reads as a story beat rather than a rendering glitch.
The camera package and the film stock
Light describes what is in front of the lens; the camera package describes the lens and what is behind it. A crew names the body, the focal length, the aperture, the frame rate and the shutter, and each of those words changes the image in a way the model has learned. You do not have to own any of it to use the words.
Focal length. Focal length sets how much of the room you see and how the space feels. A 24mm lens stretches a room and makes movement feel fast; a 35mm or 40mm feels like a person standing in the room; an 85mm flattens and isolates a face. We use 40mm today because we want to be a guest at the table, not a fly on the wall.
Depth of field. Depth of field is how much of the scene is sharp, and it is set by the aperture. “Shallow depth of field” gives you the soft background that reads as cinema; “deep focus” keeps everything sharp and reads as documentary. Shallow focus also hides a dog at the far end of a table, which is convenient.
Frame rate, shutter and stock. Twenty-four frames per second with a normal shutter gives motion blur that looks like film; sixty frames per second looks like sport or a phone. Then there is the stock: a little grain, a soft glow around bright sources (halation), and slightly imperfect color are what a viewer’s eye reads as photographed rather than generated. Figure 2.4 shows the whole package and the one line it becomes in a prompt.
Figure 2.4: The camera package, and the single line it turns into at the top of every prompt.
Color grammar
Color is the last layer of the look and the one viewers feel most and notice least. Think of color as three dials: temperature, saturation and contrast, and each dial has a meaning that decades of films have taught the audience.
Temperature is warm against cool, and it maps to comfort against distance. Warm light says home, safety, nostalgia, and sometimes danger by fire; cool light says night, technology, loneliness, calm. Our dinner is warm all the way through, because the joke is affectionate and the film should feel like a hug.
Saturation is a reality marker. Rich saturated color feels heightened, like a musical or a dream; muted color feels real, like a documentary. Contrast is intensity: a wide gap between the darkest and brightest parts of the frame reads as drama, a narrow gap reads as soft and gentle. Figure 2.5 maps the three dials and the one convention to be careful with.
Figure 2.5: Color grammar, three dials the audience already knows how to read.
As a general rule, a film gets one grade and a scene gets its own light. Changing the grade between shots is what makes AI sequences feel like a slideshow of unrelated clips; changing the light inside a shot is what makes them feel directed. Our whole film is one grade, warm and softly contrasted, with one change of light inside it.
How the model reads it
The model reads light the way it read the shot in Article 1: the first sentence locks the subject and the action, and the look is direction that follows. The difference is that light has to be described twice today, once for each state, with the cause of the change in between. Figure 2.6 shows the look header, the block of words that carries the camera package, the grade and the lighting, and where it sits in the prompt.
Figure 2.6: The look header. Camera package, grade and lighting in one block, reused across a film.
Get Micheal Lanham’s stories in your inbox
Join Medium for free to get updates from this writer.
State, cause, state. Write the first lighting state in its own time window, then the cause of the change as an action the character performs (he pulls a cord, she opens the blinds, the sun clears the roof), then the second state. Models follow a lit switch far better than a vague “the room brightens,” because the switch is an action and actions are what the first-sentence contract is made of.
Hide with light, reveal with light. In Article 1 we hid the pool outside the frame; today we hide the dog inside the frame, in shadow. That means the constraints line has to say it plainly: the far end of the table is in darkness, and the dog is not visible before the lights come on. The model will not infer the joke; it has to be told what the audience may not see yet.
The imperfection words. Glossy, symmetrical, flawless output is exactly what viewers now spot as machine-made, so the look header carries deliberate flaws: fine grain, slight halation around the candle and the bulbs, a lived-in room, and the phrase “not a commercial.” Think of these as the fingerprints that make the footage look photographed.
The workflow
The loop is the one from Article 1 with one extra step, a light plan. Before writing the prompt, write the two states and the cause in a table, because if the cause is not an action you can see, the change will not happen on cue.
- State A: One candle, 1850K, low key, far end of the table in darkness, the dog unseen
- Cause: He reaches up and pulls the cord of the string lights
- State B: String lights, 2700K, high key and warm, the whole room visible, the dog revealed at the head of the table
And the shot spec, one row as before.
- Framing: Medium shot from the side of the table, locked off
- Angle: Eye level, 40mm
- Move: None until a gentle push-in in the last six seconds
- Duration: 15 s
- What changes, and when: Lights on at 6 s; dog revealed by 9 s; he sits and raises a glass at 10 s
- Sound: Candle hiss, cutlery, the switch, bulb buzz, tail thumping; no music, no dialogue
Here is the prompt. Notice the look header sits right after the opening sentence, and the lighting is described twice with the cause between.
The generation prompt
Settings
Model: Seedance 2.5, text-to-video, no reference images
Duration: 15 seconds
Aspect ratio: 16:9
Resolution: 720p
Audio: on
Seed: any (write it down if you like the result)
Prompt
A man in a rolled-sleeve shirt sets a candlelit dinner table
with painstaking care in a dim dining room, lit only by one
candle, and when he reaches up and switches on a string of
warm bulbs above the table, the light spreads across the
room and reveals a golden retriever in a black bow tie
sitting upright at the head of the table.
Look: 40mm, shallow depth of field, 24 fps, fine 35mm film
grain, slight halation around the candle and the bulbs, warm
and softly contrasted, natural skin tones, a lived-in family
dining room, not a commercial.
[0 to 6 seconds] Locked-off medium shot from the side of the
table, eye level. Low-key lighting: a single candle flame is
the only source, a warm 1850K glow on his hands and face,
soft shadows falling off fast into darkness. He straightens
a fork, pours water into a glass, and adjusts a folded
napkin. The far end of the table is in complete darkness.
[6 to 9 seconds] He reaches up and pulls the cord of the
string lights hanging above the table. The bulbs flicker
once and come on at a warm 2700K, and the light spreads
across the whole room: cream walls, framed photos, and an
unlit bone-shaped cake on a plate at the far end of the
table.
[9 to 15 seconds] High-key warm light on everything. At the
head of the table a golden retriever in a black bow tie sits
upright on a dining chair, ears up, tail wagging against the
chair back. The man sits down opposite, raises his water
glass to the dog, and the dog tilts its head. A slow, gentle
push-in toward the two of them.
Audio: the faint hiss of a candle flame, cutlery and glass
being set down, the click of the light switch, the soft buzz
of the bulbs coming on, a tail thumping against wood, one
soft huff from the dog. No music, no dialogue.
Constraints: one continuous shot with no cuts, one man and
one dog only, no other people, no readable text, the dog
must not be visible before the lights come on, the far end
of the table stays dark for the first six seconds, no
flicker after the lights are on, no identity change, no
extra limbs.What we generated this with. Seedance 2.5, because it changes lighting state on an in-shot action and renders its own sound. The only vendor-specific element is the bracketed time windows, so the prompt reads on any current model that supports timestamps. Fifteen seconds at 720p costs roughly three and a half to seven dollars depending on the host.
The finished film: 15 seconds from the single generation above, with no edits. This preview is silent. Watch it with sound.
Judging the result
Watch it once for the story and once for the light. The joke depends on two things happening in order, so check those first.
- The dog is visible in the dark. Fix: strengthen the darkness (“the far end of the table is pitch black, nothing visible beyond the candle’s reach”) and move the dog’s first mention entirely into the third window.
- The lights are on from the start, or never come on. Fix: make the cause bigger and earlier in the sentence (“he stands, reaches up, and pulls the cord; the bulbs snap on”) and repeat “at six seconds” in the constraints.
- The image is glossy and airbrushed. Fix: add a second imperfection (“a scuffed table, a chipped plate, a slightly crooked photo frame”) and keep “not a commercial” at the end of the look line.
- The bulbs flicker for the whole second half. Fix: “flicker once, then steady” in the second window and “no flicker after the lights are on” in constraints, both of which are already there; if it persists, replace “flicker” with “come on.”
If the light changes on cue and the dog appears with it, you have directed with light. Everything in the next five articles assumes you can.
Where this goes next
You now control what the viewer sees (Article 1) and how it feels (this article) inside a single locked-off frame. Next time the camera finally moves through a room, and we will follow a waiter carrying a sparkler through a packed restaurant to find out who is turning a hundred.
This is Day 2 of 7 in the Learn AI Filmmaking series:
This series is built on my book, Learn AI Filmmaking. Each of these lessons, from the logline to the final cut, is covered there in depth.
Check out my best-selling book AI Agents In Action, 2nd Edition by downloading your copy here. Sponsored by Anaconda Inc.
