# Subject Motion — Video Generation > How subjects (people, animals, objects) should move in video prompts. AI models handle motion very differently from static images. ## Motion types ### Subtle motion (RECOMMENDED for stability) - Breathing - Blinking - Slight head turns - Hair movement in wind - Fabric rustle - Eyes shifting These work reliably in ALL AI video models. ### Moderate motion - Walking - Standing up / sitting down - Reaching for object - Turning whole body - Talking (lip movement) - Pouring liquid - Opening door Most models handle these but watch for artifacts. ### Complex motion - Running - Dancing - Fighting - Jumping - Throwing - Multi-person interaction - Sports Risky — models often produce distortions, weird limb behavior. ### Extreme motion - High-speed action - Acrobatics - Vehicle stunts - Crowds in motion Avoid unless using Sora or Veo for short clips with careful prompting. ## Motion description language ### USE ACTION VERBS WITH MODIFIERS GOOD: ``` She slowly turns her head from right to left, eyes scanning the room. ``` GOOD: ``` He gradually raises his hand to his face, fingers brushing his temple. ``` BAD: ``` She moves her head. // Too vague ``` BAD: ``` He does a thing. // No specificity ``` ### USE "BEGINS TO" FOR NATURAL TIMING GOOD: ``` She begins to walk slowly toward the door, taking three measured steps. ``` This gives the model timing cues — start state, action, end state. ### SPECIFY START AND END STATE GOOD: ``` Starting position: standing still, hands at sides, looking forward. Action: slowly turns head 45° to the left, eyes following. End position: head turned, looking off-screen left, body still facing forward. ``` This gives clear motion vector. ### AVOID INSTANT TRANSITIONS BAD: ``` He walks across the room. // No timing — model guesses ``` GOOD: ``` He takes 3 slow steps across the room over the duration of the clip. ``` ## Lip sync notes If subject is speaking: - Kling 2.x has best lip sync - Veo 3.1 is acceptable - Most others: avoid showing speaking from front (use OTS or back) For dialogue: ``` Camera: medium shot from front Subject speaks slowly, calmly: "I waited for you." Lip movements should match this dialogue. ``` ## Multi-subject motion When two subjects move together: - Specify each independently - Specify spatial relationship - Specify timing relationship ``` Subject A (left of frame) raises her hand in greeting. Subject B (right of frame) responds by stepping forward and nodding. Action timing: A's gesture begins first, B responds 1 second later. ``` ## Physics constraints AI models struggle with: - Defying gravity - Object permanence (object disappearing) - Mass/weight realism (light vs heavy objects move similarly) - Liquid dynamics (water, fire are tricky) - Rapid acceleration Use motion within natural human range. ## Default for AI video When unsure: **minimal motion**. A subtle breath + slight head turn is more reliable than complex action. --- ## Director thinking — actions are AUTHORED, not extracted The video-prompt-engineer is not a passive interpreter of the script. The script gives the **frame** (location, mood, characters, camera intent). The **specific actions inside the frame are the director's decision** — and in the AI pipeline, the prompt writer IS the director. This means: if the script says "an old Baku street, a parked carriage, a man on it, two distant pedestrians" — it is **not enough** to write "the driver does something." You must decide, as a director: - **What is this character doing in this moment?** (waiting for market, returning from work, dropping off a passenger, just arrived) - **What is the inner story of the frame?** (a city waking up, an evening winding down, a moment before something changes) - **What concrete physical action shows that inner story?** (adjusting his cap against the chill, glancing down the street, lighting a pipe, checking a pocket watch) - **What does the animal do in this moment?** (resting head down, alert with ears forward, drinking from a trough, standing patiently) - **Why are the pedestrians walking down this street?** (residents going to work, mother and child returning home, two merchants in quiet conversation) Every prompt must carry a **specific, authored moment** — not a vague checklist of "natural motion." Vague prompts produce vague clips that feel like stock footage. Authored prompts produce clips that feel like cinema. ### How to author the moment For every frame, write a one-line **inner story** before writing the action list: ``` Inner story: a quiet street wakes up at dawn — the man with the cart is waiting for market, the city is yawning into the day, light is climbing. ``` Then derive every action from that inner story: - The man's action **expresses** waiting + dawn (cap adjust against chill, long look toward the light) - The horse's action **expresses** patience + early morning (head low, weight shift, ear swivel) - The pedestrians' action **expresses** unhurried morning (calm walk, quiet conversation) - The environment's action **expresses** awakening (curtain stirs in opened window, lantern flame thinning in daylight, dust in golden light) Every action serves the inner story. Nothing is generic. ### What to refuse in the brief If you receive a brief that only says "subject moves naturally" or "make it feel alive" — **push back**. Ask the script writer / director persona what the inner moment is, or, if you are the director in the pipeline, decide it yourself and state your decision in the AZ description. Do not generate a prompt with unauthored, generic motion. --- ## Forbidden actions need a real reason — never write groundless prohibitions When you write `the horse does not step, the driver does not turn his head, the pedestrians do not stop walking` — every one of those negatives must answer the question **"why?"** with a real, story-grounded reason. If you cannot answer it, you are not directing — you are restricting blindly. ### The two prohibitions are universal — everything else needs justification In observational documentary grammar, only two prohibitions apply by default to every subject: 1. **No direct camera awareness** — characters do not look at the camera, wave at it, perform for it, or break the fourth wall. The camera is observing; subjects do not know it is there. 2. **No anachronistic / modern behavior** — period subjects do not check phones, make modern gestures, wear modern clothing, ride in cars not yet invented. Beyond these two, every additional "subject must not do X" needs a story reason: | Restriction | Valid reason (allow) | No reason (drop the restriction) | |---|---|---| | Horse does not bolt | Yes — bolting would tear the carriage out of frame, breaks composition | — | | Horse does not step at all | Where's the reason? Horses shift weight constantly. **DROP** | | Driver does not dismount | Yes — dismounting changes the staging, leaves an empty bench | — | | Driver does not turn his head | Where's the reason? Real humans look around. **DROP** | | Pedestrians do not stop | Yes — they're walking past, stopping would make them a focal point | — | | Pedestrians do not breathe | Where's the reason? They're alive. **DROP** | ### The reflex to "lock everything down" is wrong A common failure mode: the prompt writer reads a poetic script line ("frozen time", "still photograph", "the world holds its breath") and reflexively writes a wall of negatives that turn every living subject into a statue. The result on screen is dead: a CGI freeze-frame, not a documentary. The right reflex is the opposite — **start from "everything is alive"** and only forbid what would genuinely damage the story (camera awareness, dramatic action that doesn't fit the inner moment, anachronism). Everything else stays. ### Audit your own prompt Before you ship a prompt, read every negative line and ask: **"If I delete this restriction, what bad thing happens on screen?"** - If you can name a concrete bad outcome → keep the restriction - If your answer is "I don't know, I just felt like it shouldn't move" → delete the restriction This audit will cut 50–70% of the negatives in a typical first draft, and the resulting prompt will produce a clip that feels alive instead of taxidermied.