Phases of motion and the motion prompt
Lesson 10 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Three phases of any action
Every action has three phases: preparation, the action itself, and recovery, when the body returns to balance. A jump is a crouch, flight and a landing that absorbs the impact. Kicking a ball is the wind-up, the kick and the leg following through after the ball. Models often spend the whole shot on the main action, and the video cuts off mid-motion or loops in a weird repeat. So plan a shot as a tiny 5–8 second story and leave the last seconds for recovery: the person stops, breathes out, looks into the camera. That kind of ending is also easy to cut to the next shot.
A specific verb
“Dances”, “works out”, “moves around”: for a model that’s an invitation to improvise, and it usually tangles arms and legs. Pros write motion like a choreographer: which steps, in what order, with which foot. Not “dances” but “spins once on her left heel”. Not “works out” but “steps forward and drops into a squat”. A specific verb narrows the model’s choices, which leaves less room for glitches. A good test: could your friend repeat the movement from your description without asking a single question? If yes, the model will get it too.
The order of actions
A model reads a prompt roughly like a script: what’s written earlier usually happens earlier. So lay the motion out in steps and connect them with first, then, finally. Write the camera as a separate line so it doesn’t mix with the body: “runs toward the camera” and “the camera moves toward her” are very different things for the model. Timecodes like “(0–2s)” work in some models and get ignored by others, so test on one draft in your model. Look at Mira’s prompt: three phases of a run in three lines, the camera separate. Text like that is easy to edit one line at a time.
Hands and objects
Hands are the hardest thing for video models: they’re small in the frame, there are many fingers, and they overlap each other in every frame. That’s where six fingers come from, a cup growing into a palm, a drawstring held “through” the hand. Rules lower the risk. A simple grip: holds a cup, pulls a drawstring. An object of plausible size with one hand on it. No fast finger work like guitar chords or typing. The hand is either fully in frame or out of it, because fingers cut off at the frame edge often melt too. And if the contact is complex, hide it with a cut.
The first frame decides
In image-to-video, the model continues motion from your picture, so the pose in the first frame has to lead into the action. If Mira sits on a kerb in the start frame and the prompt says she “runs off”, the model either teleports her to standing or breaks her legs on the way up. Choose a start frame from which the action begins naturally: mid-step for a run, a slight crouch for a jump. Or write the transition: “she stands up first, then runs”. Many models also let you set the last frame, and then the model fills in the motion between two poses. That’s the easiest way to get an exact ending.
Two people in frame
With two people, the difficulty grows more than double: the model has twice as many arms and legs, and they overlap on contact. Hugs, handshakes and high-fives in close-up are classic spots where fingers grow into each other. What works: a wide shot where hands are small and details less visible; brief contact; a cut exactly at the moment of contact (the viewer fills in the slap of palms); each person with their own clear action instead of one tangle. If you really need a close-up touch, shoot it as a separate shot with one hand in frame.
The refinement loop
A perfect shot on the first generation is rare even for pros, so they have a process. First, 2–4 drafts in a cheap mode to test the idea. Then pick the best and check it frame by frame against the weight checklist from lesson 8. Then the main rule: change one thing at a time. If you rewrite the whole prompt, you won’t know what helped, and you may lose what already worked. Generate the final in high quality only once the draft is good. That’s how Mira saves credits and learns which words in her prompts actually work.
Check yourself
- The prompt “she dances energetically” gives chaos: arms and legs get tangled. What should she do?
Correct answer: Describe 1–2 specific moves: “steps left, then spins once on her right heel”. Specific steps narrow the model’s choices, so it no longer has to improvise. - A 5-second shot cuts off in mid-air, in the middle of the jump. How to fix it?
Correct answer: Describe the landing and recovery, and shorten the preparation or use 8 seconds. All three phases must fit into the shot. The model spends time on what’s described, so describe the ending too. - Start frame: Mira sits on a kerb. Prompt: “sprints away”. The result is a teleport. What should change?
Correct answer: Use a frame where she’s already mid-step, or write “stands up first, then runs”. The model continues motion from the first frame. The pose must lead into the action, or the transition has to be described. - The draft has three problems: feet slide, the hood freezes, the light is flat. What’s the best way to fix it?
Correct answer: One fix per iteration, starting with the most visible. One change at a time shows exactly what worked and doesn’t break what was already good.
← 12 principles: bringing motion to life The world reacts to motion →