Most AI video tools have solved the easy part: generating movement. The hard part is deciding what the frame should feel like before a single prompt is written. That decision is cinematography, and it separates a clip that looks technically impressive from one people actually remember.
The film tradition that includes painters of light such as Peter Pau — known for wuxia epics and romantic dramas built on silhouette, mist, and warm interior light — offers a practical vocabulary. You do not need to copy specific films. You need to borrow the underlying decisions: where light comes from, how much depth the frame has, what color is doing emotionally, and how the camera behaves. This guide turns those decisions into a repeatable workflow for any modern text-to-video or image-to-video tool.
Start With Intention, Not With a Prompt
Beginning with a prompt is the most common reason AI video looks generic. A prompt describes content; it does not describe intent. Before you type anything, write one sentence that states the emotional job of the shot. "A woman waits" is content. "A woman waits, and the room feels like it is holding its breath" is intent, and intent dictates every technical choice that follows.
Next, define the shot's dramatic function in the scene. Is it establishing scale, revealing a relationship, building unease, or releasing tension? A single clip can only do one of these well. When you ask a model to deliver scale, intimacy, and mystery at once, it tends to produce a wide, empty composition with soft light and no point of view.
Finally, collect two or three reference frames — stills, paintings, or photographs. Do not ask the model to imitate an artist's style. Instead, name the properties you see: "top light through a window slot, hard shadow diagonal across the floor, cool blue exterior visible through a warm doorway." Property-based references translate into prompts far more reliably than style names, and they keep output consistent across a series of shots.
Reading Light the Way a Cinematographer Does
Key, fill, and the emotional temperature of a scene
Every lighting setup answers three questions: where is the primary source, how much do the shadows open up, and what separates the subject from the background? The ratio between key and fill is the emotional dial. A high ratio — hard key, almost no fill — reads as drama, danger, or loneliness. A low ratio — soft key with generous fill — reads as safety, comedy, or commercial polish.
In prompts, describe direction and quality rather than equipment. "Single hard key from camera left, deep falloff into shadow, small warm rim light" produces a far more specific result than "cinematic lighting." Add falloff language: "shadow detail drops off quickly behind the subject." Falloff is what makes generated light feel physical instead of painted on.
Motivated light sources and practical fixtures
Great cinematography rarely lights a room; it lights the sources inside the room. A table lamp, a doorway, a window, a phone screen, or a car headlight justifies the direction and color of every beam. When you give a model a visible in-frame source, the light stops looking arbitrary.
Practical sources also solve continuity. If the same lamp is visible in three shots, the audience accepts that the light behaves consistently between them. For AI work this is also a generation trick: a visible source gives the model an anchor, which reduces the drifting shadows that appear when illumination has no stated origin. Describe the fixture, its color temperature, and its position relative to the subject in every shot where it appears.
Composition That Survives Generation
Depth layering: foreground, midground, background
AI video flattens scenes. It fills the frame evenly, applies shallow depth of field everywhere, and forgets that a room has corners. Counter this by explicitly asking for layers: "out-of-focus foreground element at frame right, subject in the midground, distant window with cool haze in the background." Three planes give the eye somewhere to travel and make the image feel photographed rather than rendered.
Silhouette is the strongest version of this. Placing a dark subject against a bright, hazy background collapses detail and forces the audience to read posture and shape. It is also one of the easiest looks to generate, because the model does not need to resolve faces or fabric texture. Use it when you want scale, mystery, or the feeling of a figure moving through a hostile environment.
Negative space, thirds, and deliberate imbalance
Centered framing communicates control. Off-center framing with empty space communicates tension or loneliness — the empty area is where the audience expects something to enter. Decide which reading you want, then place the subject accordingly and state it in the prompt: "subject small in the lower third, large empty sky above."
Beware of automatic symmetry. Many models default to balanced, centered, postcard compositions because they are statistically safe. If you need imbalance, state it twice: once in the composition instruction and once by describing what occupies the opposite side of the frame.
Color as Narrative, Not Decoration
Color should tell the audience where they are emotionally before dialogue or action does. A warm amber interior against a cold blue exterior is a classic shorthand for safety versus exposure. Desaturated midtones with one saturated accent pushes attention to a single object. Monochrome schemes with a single hue shift suggest memory or a different time.
Practical guidance: pick two dominant colors and one accent, then hold that palette across every shot in the sequence. Consistency matters more than sophistication. Three shots that share a palette read as a scene; three shots with different palettes read as a mood board.
In AI workflows, color drift is the most visible continuity failure. Describe temperature in Kelvin-like terms, such as "warm 3200K practicals, cool 5600K daylight through the window," and repeat that color sentence verbatim in each prompt. Where the tool allows reference images, supply a graded still as a color anchor. Then finish in post: a light grade with matched shadows and slightly cooled highlights can unify clips that were generated separately.
Camera Language: Lens, Movement, and Shot Size
Shot sizes and what they communicate
Wide shots establish geography and make people small. Medium shots carry dialogue and body language. Close-ups carry internal state. Most AI clips fail because the prompt asks for a wide, cinematic, epic shot when the story beat is actually a close-up of a hand.
Plan shot size before motion. A slow push-in on a medium shot reads as realization. A locked wide with a tiny moving figure reads as inevitability. The same subject, lighting, and palette can produce opposite emotions purely through framing and movement.
Movement: slow push, lateral track, locked frame
Models favor drifting, floating camera moves because they look dynamic in short samples. Restraint reads as confidence. Specify one movement per shot and describe its speed: "slow push-in over four seconds, almost imperceptible," or "locked frame, subject exits left, camera does not move."
Locked frames are underused and cheap to generate, because the model does not have to invent parallax. They also cut together well; a series of static shots with strong internal motion, such as fabric in wind or rain in a doorway, feels more directed than a sequence of unrelated swoops.
Focal length, distortion, and perspective
Lens language changes spatial feeling. Long-lens compression makes backgrounds loom behind subjects and flatters faces; wide lenses exaggerate proximity and make spaces feel unstable. Describe the effect rather than the millimeter: "compressed background, distant mountains appearing close behind the subject" versus "wide perspective, foreground hand large and distorted."
A quick translation table
| Cinematic intent | Prompt language | What to check |
|---|---|---|
| Isolation and drama | hard key from camera left, minimal fill, deep shadow falloff | shadow direction stays consistent, features readable |
| Epic scale | subject small in lower third, layered haze, wide perspective | subject does not merge with background texture |
| Intimacy | soft window key, low contrast, close framing | skin tones stay stable, no waxy smoothing |
| Unease | off-center framing, cool practical source, slight tilt | horizon does not drift mid-shot |
| Memory | warm haze, reduced saturation, shallow depth | palette holds across the sequence |
| Kinetic energy | lateral tracking, foreground blur, motion streaks | subject edges stay clean, no ghost limbs |
Treat the table as a translation layer, not a formula. Take one row, write a prompt, generate several variants, and change only one variable per run. If you change lighting and camera movement together, you learn nothing from the comparison.
A Repeatable Workflow From Mood Board to Final Clip
Build a small reference library
Collect twelve to twenty frames you genuinely like, organized by lighting direction, palette, and shot size. Tag them in plain language. When you need a look, you are no longer searching for inspiration; you are selecting from a shelf.
Write a shot list with one variable per shot
For each shot, record five decisions: shot size, light direction and ratio, palette, camera movement, and duration. If two consecutive shots share all five, one of them is redundant. If they share none, the sequence will feel fragmented.
Iterate in passes
Pass one is composition: lock framing and subject placement with low motion. Pass two is light: refine direction, ratio, and falloff. Pass three is motion: add the camera move. Pass four is finish: upscale, stabilize, and grade. Trying to solve everything in a single prompt produces a clip that is mediocre at all four.
Finish in the edit
No generated clip arrives finished. Trim to the emotional beat, not to the clip's natural length. Add sound design before color: footsteps, cloth movement, room tone, and a low drone will do more for perceived quality than an extra grade pass. If a shot only works for two seconds, use two seconds.
Evaluating Output: What to Keep, What to Regenerate
Judge every generation against five criteria. First, subject consistency: does the face, hair, and wardrobe survive the whole clip? Second, lighting continuity: do shadows stay on the same side, and does the light source remain motivated? Third, motion physics: do limbs, fabric, and water behave plausibly? Fourth, artifact load: warped hands, melting edges, flickering texture, and text-like smears. Fifth, emotional read: does a viewer who sees the clip without context describe it the way you intended?
Rank the five in order of importance for your project, then accept any clip that satisfies the top three. Chasing perfection on the last two criteria usually costs more time than the improvement is worth. A slightly imperfect clip that reads correctly will outperform a technically clean clip that reads as nothing.
Common Mistakes That Break the Cinematic Illusion
- Describing style instead of properties. "Shot like a famous film" tells the model almost nothing. Name light direction, contrast ratio, palette, and lens effect.
- Stacking movements. Push-in plus orbit plus handheld shake produces mush. One movement per shot.
- Ignoring the background. A beautifully lit subject in front of a flat, evenly bright wall still looks artificial. Give the background its own light and depth.
- Varying the palette between shots. Color drift destroys scene continuity faster than any generation artifact.
- Overfilling the frame. Empty space is a composition tool, not wasted pixels.
- Grading before editing. Lock the cut first; per-clip grades applied to an unfinished sequence rarely match.
- Trusting a single generation. Produce several variants and select. Generation is cheap; judgment is the scarce resource.
- Neglecting sound. Silent clips feel synthetic. Room tone and physical sound effects restore credibility instantly.
FAQ
How much cinematography knowledge do I need to start? Very little. Learn three things: where the light comes from, how much contrast the shot needs, and how big the subject is in frame. Those three decisions cover most of the visible difference between amateur and professional-looking output.
Should I use reference images or text prompts? Use both. Reference images anchor palette and lighting quality; text narrows subject, framing, and motion. Text-only pipelines drift fastest on color, while image-only pipelines often lock in composition you cannot adjust.
How do I keep a character consistent across shots? Fix wardrobe, hair silhouette, palette, and light direction, and change only shot size and angle between generations. Consistency comes from constraint, not from more description.
Why do my generated shadows look wrong? Usually because no light source was stated. Add a visible practical source, describe its position relative to the subject, then repeat that sentence in every prompt for the scene.
Is a locked camera boring? No. Static framing with strong internal motion is one of the most reliable looks in AI video, and static shots cut together more cleanly than drifting ones.
How long should each clip be? As short as the story beat allows. Most cinematic sequences work best with clips of two to five seconds, cut on movement or on a light change rather than on a fixed rhythm.
When should I stop iterating? When the top three evaluation criteria are met and the clip reads as intended. Beyond that, additional passes usually trade time for negligible visible gain.
Where to Go From Here
Cinematography, whether practiced with a camera or with a generative model, is a discipline of subtraction. You choose one light, one palette, one movement, and one idea per shot, then defend those choices through the whole sequence. Study the work of master image-makers to build vocabulary, but build your own shelf of references, your own translation table, and your own evaluation criteria.
The practical next step is small: pick a scene of three shots, write the five decisions for each, generate several variants per shot, and finish the sequence in the edit with sound. Repeat that loop a few times and the techniques stop feeling like rules and start feeling like instincts.
