Learn AI Filmmaking Day 6 — Effects That Belong in the World
Day 6 of 7: the ten pillars of a believable effect
Series note. This is the sixth of seven articles that teach AI filmmaking from the ground up. So far our films have obeyed the laws of the world. This article breaks them on purpose, with the framework that decides whether an audience believes an effect or spots it as a sticker on top of the footage.
The film this article builds: 30 seconds from a single generation. Watch it with sound.
The brief
Georges Méliès told the story for the rest of his life. In 1896 he was filming a Paris street when his camera jammed for a few seconds; when he projected the reel, a bus had turned into a hearse and the pedestrians had jumped from one place to another. He had discovered the substitution splice by accident, and within six years he had used it, and a hundred tricks like it, to send a rocket into the eye of the moon.
Méliès understood something that still decides whether an effect works. The trick itself is easy; the hard part is making the world around the trick behave as if it really happened. A video model can generate an explosion in one line of prompt, and the explosion will look like a sticker unless the room lights up, the debris obeys gravity, the camera flinches and the floor stays wet afterwards.
So here is the brief. We are going to make a thirty-second film with one effect that has a full arc, a slow-motion peak and a reaction shot, from a single generation. Its logline: at the school science fair, a nervous kid’s papier-mache volcano erupts far bigger than planned, and the drenched judge gives it two thumbs up.
The framework we will use is the ten pillars, shown in Figure 6.1. Think of the pillars as a checklist of everything a real event does to the world around it, which is exactly what a rendered event forgets to do.
Figure 6.1: The ten pillars of a believable effect. Everything a real event does to the world around it.
Cause, peak, aftermath
Every effect is a three-act story. Something visible triggers it, it reaches a peak of energy, and it leaves the world changed, and an effect that skips any of the three reads as false even to a viewer who cannot say why. A fireball that appears from nowhere is a glitch; a fireball that follows a spark and leaves scorched pavement is an event.
Our volcano has the cleanest cause in cinema, a kid pouring vinegar into a crater, and the prompt lingers on it for six seconds so the audience has time to dread what comes next. The peak is the eruption, and the aftermath is pink foam on everything, dripping from a basketball hoop, for the last thirteen seconds of the film. Figure 6.2 puts the arc on a timeline with the seconds we will give each act.
Figure 6.2: Cause, peak, aftermath. The three-act structure of an effect, timed for our volcano.
As a general rule, give the cause more time than the peak. The eruption is over in seven seconds, and the audience only enjoys it because the first ten seconds taught them what to fear.
Light and residue
Two pillars do more than the rest to sell an effect, and they are the two most often forgotten: the effect must light the world, and the world must stay marked afterwards. A real explosion throws orange light on every wall in the frame; a real flood leaves everything wet. When a rendered effect does neither, the eye reads it as pasted on top.
Our foam is pink, wet and lit by flat gym fluorescents, so the prompt says the foam catches the light with a wet glisten and throws a pink bounce on nearby faces. The residue is the whole final act: foam sliding off the judge’s shoulders, hanging from the hoop, pattering on the floor, and a single finger wiping one lens of her glasses. Figure 6.3 shows the color temperatures effects produce and the residue each one leaves.
Figure 6.3: Effects light the world and leave marks. Color temperature and residue by effect type.
Scale, occlusion, physics, particles
The middle pillars are about space. An effect has a size, and the audience learns that size only from things they already know the size of, so a person standing near it, a folding table, a basketball hoop above it, all tell the viewer how big the eruption is. Our foam rolls over the table edge and hangs from the hoop, and those two objects are the ruler.
Occlusion is the effect passing in front of and behind things: foam that flies past the camera in the foreground and blocks the judge for a moment is foam that exists in three dimensions, while foam that never overlaps anything is a layer. Physics and particles are the rest of the argument: droplets arc and fall under gravity, big blobs fall faster than fine mist, and each particle has a life from spawn to splat. Figure 6.4 collects the four pillars with the prompt phrases that carry them.
Figure 6.4: Scale, occlusion, physics and particles. The four pillars that put an effect in three dimensions.
The camera is a witness
Here is the pillar that separates an effect that was filmed from one that was rendered. A real camera reacts to a real event: it lurches at the blast and settles over a second, it loses focus and finds it again, it smears fast debris into motion blur and blooms bright sources into flare. Write those reactions into the prompt and the audience believes a person was holding the camera when it happened.
Slow motion belongs here too, and it is one of the effects only a camera can do. We ramp into slow motion at the peak and back to normal speed for the aftermath, which stretches the seven seconds the audience paid for and lets every droplet have its arc.
Sound is the tenth pillar and the realism multiplier: a blast is a sharp transient, a deep body and a long tail that rings in whatever room it happened in, and a gym has a very particular ring. Figure 6.5 shows the camera and sound vocabulary we will use.
Figure 6.5: The camera is a witness. How a real camera and a real room react to an event.
When the prompt gets too big
Ten pillars is a lot of words, and a prompt can get so detailed that the model loses the plot. The book’s advice still holds: build the effect progressively in your head (start with the baseline, add the pillars in pairs), but write the full prompt once, and if the result degrades, cut rather than add. Contradictions are the usual culprit, and “slow motion” beside “quick frantic cutting” or “pitch dark” beside “bright and colorful” will confuse any model.
Two decisions keep our prompt inside the limit. We prioritize the pillars a foam eruption needs most (physics, particles, residue, camera, sound) and mention the others once; and we split the effect from the reaction by giving the reaction its own time windows, which is also where the story lives. Figure 6.6 is the decision framework, and the last card on it is the one that matters most.
Figure 6.6: When the prompt gets too big. Prioritize, split, remove contradictions, and let the effect serve the story.
Remember, the reaction is the story. The eruption is spectacular for seven seconds, but the film is about a judge deciding, with foam in her hair, that the kid deserves it. The book calls this using effects as character revelation, and it is why the last thirteen seconds of our prompt describe a face more than a foam.
Effects only a model can do
Most of this article applies to any effect ever filmed, but a video model can also do things no camera can. It can shift style inside a shot, so a hand-drawn character steps through a door and becomes photographic; it can move a camera through a keyhole or orbit a room in one unbroken take; it can run time forward and backward within a shot, and it can morph one object into another across a match cut. Every one of those still needs the pillars, because a style shift with no lighting logic is a glitch and a morph with no residue is a screensaver.
Use them the way Méliès used his splice: as a story beat with a threshold the audience can see. As a general rule, one impossible thing per film, and make it the turn. Our impossible thing is modest, a science-fair volcano that behaves like a real one, and the next article uses a subway platform’s lights to do a little more.
How the model reads it
The structure is the one you know, with the pillars folded into the time windows where each one occurs. The cause window carries scale references (the table, the hoop, the judge leaning in); the peak window carries physics, particles, light, occlusion and camera response, and is flagged as slow motion in its first words; the aftermath windows carry residue and sound; the reaction windows carry the story. The pillars are not a list at the end; they are woven into the moments where a witness would notice them.
Say what must not happen early. The foam cannot appear before the cause is complete, so the constraints say so, exactly as we hid the pool in Article 1 and the dog in Article 2. Anticipation is the pillar the audience feels most and the one the model skips first.
Plausible, not perfect. The words “physically plausible” in the constraints do real work, and so does the phrase “far bigger than anyone expected” in the opening sentence, because it licenses the scale without licensing nonsense. We want a volcano that could happen in a gym, just once, on the worst day of a judge’s career.
The workflow
The loop gains a pillar checklist, filled in for this effect before the prompt is written. The right-hand column is the phrase that goes in the prompt, so the checklist is the prompt in pieces.
Get Micheal Lanham’s stories in your inbox
Join Medium for free to get updates from this writer.
Pillar: 1 Cause, peak, aftermath
- For the volcano: Pour, gurgle, blast, foam everywhere
- Prompt phrase: “pours vinegar”, “a low gurgle”, “the crater blows”, “foam keeps dripping”
Pillar: 2 Light
- For the volcano: Flat fluorescents, wet pink bounce
- Prompt phrase: “catching the fluorescent light with a wet glisten, a pink bounce on nearby faces”
Pillar: 3 Occlusion
- For the volcano: Foam past the lens, blocking the judge
- Prompt phrase: “flies past the camera in the foreground and briefly blocks the judge”
Pillar: 4 Scale
- For the volcano: Table, hoop, kids
- Prompt phrase: “over the table edge”, “hanging from the basketball hoop”, “the kids step back”
Pillar: 5 Physics
- For the volcano: Arcs, gravity, heaviest first
- Prompt phrase: “arcing outward and falling under gravity, the heaviest blobs first”
Pillar: 6 Particles
- For the volcano: Blobs, droplets, mist
- Prompt phrase: “breaking into blobs and droplets of every size, a fine mist hanging last”
Pillar: 7 Atmosphere
- For the volcano: A haze of mist
- Prompt phrase: “a faint pink mist hangs in the air”
Pillar: 8 Camera
- For the volcano: Lurch, settle, slow motion
- Prompt phrase: “the camera lurches with the blast and steadies over a second”, “slow motion”
Pillar: 9 Residue
- For the volcano: Foam on judge, hoop, floor
- Prompt phrase: “foam sliding off her shoulders, hanging from the hoop, pattering onto the floor”
Pillar: 10 Sound
- For the volcano: Transient, body, gym tail, silence, cheer
- Prompt phrase: “a sharp wet blast with a deep body and a long gym-reverb tail”
Here is the prompt.
The generation prompt
Settings
Model: Seedance 2.5, text-to-video, no reference images
Duration: 30 seconds
Aspect ratio: 16:9
Resolution: 720p
Audio: on
Seed: any (write it down if you like the result)
Prompt
At a school science fair in a gymnasium, a nervous
twelve-year-old boy in oversized safety goggles pours
vinegar into a lumpy papier-mache volcano and it erupts far
bigger than anyone expected, blasting a column of pink foam
over the table and drenching the judge, who wipes one lens
of her glasses and slowly raises both thumbs.
Look: 35mm, 24 fps, fine 35mm film grain, flat school-gym
fluorescent light with a wet pink bounce off the foam on
nearby faces, handheld documentary feel, saturated but
realistic colors, a real school gym, not a commercial.
[0 to 6 seconds] Medium shot, eye level. A gym lined with
folding tables and tri-fold poster boards, a basketball hoop
high in the background. The boy tips a plastic jug and pours
vinegar into the crater of the brown papier-mache volcano on
his table. A judge in a blazer with a clipboard leans in
close, unimpressed. Other kids crowd behind the table.
[6 to 10 seconds] A low gurgle. The volcano trembles, a
first pink bubble swells at the crater, and the kids around
the table take one step back. The judge does not move.
[10 to 17 seconds] Slow motion. The crater blows: a thick
column of pink foam blasts upward, catching the fluorescent
light with a wet glisten, breaking into blobs and droplets
of every size, arcing outward and falling under gravity, the
heaviest blobs first, a rolling wave of foam pouring over
the table edge and across the floor. Foam flies past the
camera in the foreground and briefly blocks the judge. The
clipboard flies out of her hand. The camera lurches with the
blast and steadies over a second.
[17 to 24 seconds] Normal speed. The judge stands drenched
head to toe in pink foam, her glasses coated, foam sliding
off her shoulders in slow drips. A faint pink mist hangs in
the air. Foam hangs from the basketball hoop above and
patters onto the floor. Everyone is silent. She lifts one
finger and wipes a single lens of her glasses clean.
[24 to 30 seconds] She looks at the boy for a long beat,
then slowly raises both thumbs. The gym erupts in cheering,
the boy's face splits into a grin, and pink foam keeps
dripping from the hoop.
Audio: gym echo and kids murmuring, the glug of vinegar
pouring, a low rising gurgle, then a sharp wet blast with a
deep body and a long gym-reverb tail, foam splattering on
the table and floor, droplets pattering from the hoop, dead
silence, one squeak of a finger on a wet lens, then a wall
of cheering. No music, no dialogue.
Constraints: cuts only where the time windows change, one
boy and one judge in focus with other kids in the background
only, no readable text on the poster boards or the
clipboard, no foam before ten seconds, physically plausible
arcs and drips, the foam must light nearby faces and stay on
every surface it lands on, no identity drift, no extra
limbs.What we generated this with. Seedance 2.5, because it handles slow-motion time windows, lights the scene from the effect, and renders the transient and tail of a sound in the right room. The time windows are the only vendor syntax; any current model that supports timestamps reads the rest. Thirty seconds at 720p costs roughly seven to fourteen dollars depending on the host.
The finished film: 30 seconds from the single generation above, with no edits. This preview is silent. Watch it with sound.
Judging the result
Freeze a frame at the peak and look at the judge’s face and the wall behind her: if neither has any pink light on it, the effect is a layer. Then watch the last ten seconds and count the surfaces that stayed wet.
- Foam appears before the pour. Fix: “no foam before ten seconds” is in the constraints; add “the crater is dry and empty” to the first window.
- The eruption looks like paint or cartoon goo. Fix: add “wet, glossy, translucent at the edges, like real soap foam” to the peak window and cut “pink” to a single mention.
- Nothing stays wet. Fix: repeat the residue in the reaction window (“foam still on her shoulders and in her hair as she raises her thumbs”).
- Slow motion is ignored. Fix: open the peak window with “In slow motion, at one-quarter speed,” and add “returns to normal speed” to the start of the aftermath window.
- The judge is dry or a different person. Fix: describe her once in the opening sentence (“a judge in a navy blazer with short grey hair and glasses”) and keep “one judge” in the constraints.
If the room lights up pink, the foam obeys gravity, the camera flinches, and the judge is still dripping when she raises her thumbs, the effect belongs in the world. That is the difference between spectacle and story, and it is the last skill before we put everything together.
Where this goes next
Six articles, six layers: the shot, the look, the camera, the character, the cut, the effect. Next time we use all of them in one thirty-second prompt, with references, dialogue, coverage, sound design and one small impossible thing, and a tuba player who empties a subway platform of its phones.
This is Day 6 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.
