VFX and Effects Prompting for AI Video

TL;DR: "VFX" usually means an effect composited onto footage, with a mask and adjustable timing. AI video models don't do that: they generate a whole shot, effect baked in. Good for a full effect-driven shot, previs, and elements to grade over. Organised by effect, with 28 parameterised prompts and an honest sort of what a rewrite fixes, what a setting fixes, and what nothing fixes.

Is "VFX" Even the Right Word for This?

Mostly, no, and saying so up front saves you a wasted afternoon. In production, VFX means an effect composited onto filmed footage, with a mask that isolates it, timing you can nudge frame by frame, and the freedom to redo the smoke pass without reshooting the actor. That workflow assumes layers.

A text-to-video model doesn't produce layers. It produces one finished clip, generated in a single pass, where the fire, the smoke or the magic effect is baked into every pixel from the start. There is no mask to soften, no separate pass to retime, no way to pull the explosion two frames later without regenerating the entire shot and hoping the rest holds. What looks like a VFX shot on the timeline is, structurally, closer to a single unadjustable take.

That's not a reason to avoid these tools for effects work. It's a reason to be precise about what you're asking for, because the honest description of the output changes what a good prompt looks like.

So What Are These Tools Actually Good For?

Three things, and none of them is "add this effect to my existing footage."

Generating the whole effect-driven shot. You describe the explosion, the storm, the transformation, and the camera, and the model generates a complete clip built around that effect from frame one. This is the main use case below.

Previs. Before you commit a VFX budget, or before you brief a real compositor, a generated version of the shot shows direction, timing and mood cheaply. Nobody expects the previs pass to be the deliverable.

Elements to grade or comp over. A clean burst of sparks, a plate of drifting smoke, a hologram flicker shot against a simple background: generated as its own short clip, then pulled into a real editor or compositor as one layer among several. This is the closest the workflow gets to traditional VFX, and it works because you're treating the output as raw material, not a finished shot.

What none of them do: mask an effect onto footage you already shot, key it against existing content, or let you adjust the timing after generation without a new take.

What Makes an Effect Read as Real, Not Like a Sticker?

Interaction. An effect that exists in its own bubble, with nothing around it reacting, reads as pasted on no matter how detailed the effect itself is.

Name what the effect touches. Fire and magic effects should spill light onto the nearest wall, face or surface. Debris should obey something resembling gravity, scattering low and settling rather than hanging in the frame. Smoke or dust should be displaced by whatever is moving through it, not sit static while a car drives past it.

Scale cues do real work here, because these models have no ground truth for how big or heavy anything is unless you tell them. "A boulder-sized chunk of rubble" behaves differently in the model's output than "debris," even though both are debris.

The other lever is the camera's own reaction: a flinch, a whip-pan away from the blast, a half-second of blown-out exposure that recovers, a handheld shake on impact. A camera that behaves as if the effect is really there sells the effect more than another adjective on the effect itself does. Our camera movement vocabulary reference has the terms that actually move the render, checked against what Kling documents.

Fire and Explosions: Why Do They Look Fake?

Usually because nothing around the fire acknowledges it exists. Real flame throws hard, flickering light on the nearest surfaces; a generated fireball often sits in the frame like a sticker. Name the light spill and the camera's reaction explicitly, and describe the beat you want the ignition to land on, since these models have no internal sense of "now" unless you give them one.

Wide shot of an abandoned warehouse at night, a fuel drum ignites near the
centre of frame, orange-white fireball rolling upward and outward, harsh
flickering light throwing long shadows across the nearby brick wall and
crates, camera pushes back half a step and the exposure blows out for a
beat before recovering, low rumble on impact

Close interior shot, a small gas leak ignites at the stove, a controlled
burst of blue-orange flame licking upward, light catching the metal
cabinets and glassware on the counter, steam and smoke curling off the
scorched surface, handheld camera flinches back on ignition

Night exterior, a vehicle explodes at [DISTANCE] from camera, fireball
expanding with a visible pressure wave rippling the air around it, debris
of [SCALE, e.g. car-door-sized panels] arcing outward and falling, camera
shakes on the shockwave, brief lens flare toward the blast

Seedance 2.5 documents integer-second timestamps in its own tutorial, so if you're on that model you can place the ignition at an exact second inside a longer generation rather than describing it only in relative terms like "a few seconds in." Seedance 2.0 only responds to shot numbers, not timestamps, so the same instruction there needs a shot-grammar rewrite instead.

Smoke and Atmosphere: What Makes It Read as Volume?

Density falloff and displacement. Flat, uniform haze reads as a filter over the image; smoke that thins toward its edges and reacts to whatever moves through it reads as a physical substance in the room.

A figure walks through a fog-filled corridor lit by a single overhead bulb,
fog thick and low near the floor, thinning toward the ceiling, swirling and
parting visibly around the figure's legs as they walk, backlit so the
particles catch the light

Interior shot, smoke drifts from an extinguished candle on a table, thin
grey wisps curling and dispersing slowly in still air, warm side light
catching the smoke's edge, shallow depth of field, everything else static

Wide shot of a smoke-filled arena, colored stage smoke pooling low across
the floor, a performer's movement visibly pushes a channel through the
smoke as they cross, haze thickening again slowly behind them

No vendor publishes a smoke-density or displacement parameter. This is prompt-only craft, and it's also one of the more reliable prompt-fixable effects on this list, because the failure is usually that nothing in the prompt described the interaction, not that the model can't render it.

Water and Liquids: Why Does Volume Never Conserve?

Because there's no fluid simulation underneath any of this, only a model predicting plausible-looking frames. A poured glass may not contain the volume it started with by the last frame; a wave may lose or gain mass mid-motion. No published parameter changes this on any consumer video model as of this writing, so treat liquid work as the effect category with the lowest ceiling on this whole page.

Close-up, red wine pours into a wide glass on a dark table, liquid
swirling and settling, single droplet running down the outside of the
glass afterward, soft studio side light catching the surface, static camera

A wave breaks against rocks on a coastline, white spray thrown upward and
outward on impact, water receding and pulling debris and foam back with it,
camera locked off, low afternoon side light

Overhead shot, a droplet falls into a still pool of water, circular ripples
expanding outward from the point of impact, ripples slowing and fading
before reaching the frame edge, camera static, soft top light

For anything where the water's exact volume, splash shape or surface behaviour has to be correct on delivery, this is squarely the third bucket below: a different tool, not a better sentence.

Rain, Snow and Fog: How Do You Prompt Weather?

By separating foreground and background particle behaviour and by describing what the weather does to surfaces, not just what falls from the sky. Rain that only exists as streaks in the air, with no wet reflection on the ground, reads as an overlay rather than a place.

Night street exterior, heavy rain falling, streetlight reflecting in wet
asphalt, a figure walks under an umbrella, near-camera raindrops streak
fast and blurred, background rain slower and softer, distant neon signs
smeared in the wet reflections

A quiet forest clearing under falling snow, flakes drifting slowly and
unevenly, thin layer of snow visibly accumulating on branches and the
ground over the shot, breath fogging faintly in the cold air, flat
overcast light, camera static

Coastal road at dawn, dense fog rolling low across the tarmac, headlights
of an approaching car cutting two soft cones through the haze, visibility
dropping to a few metres, camera slow push forward, muted desaturated
colour

Where a model publishes a real negative-prompt field, it's genuinely useful here for suppressing a specific unwanted artefact (blur, extra vehicles) rather than for controlling the weather itself, which stays a positive, descriptive job.

Magic and Energy Effects: What's Actually Promptable?

Almost all of it, since no vendor ships a "magic" preset, and that's an advantage: you're not fighting a closed enum, you're describing light, colour and interaction the same way you would for fire.

A figure raises one hand and a swirling ball of blue energy forms above
their palm, light pulsing and casting a cool blue glow on their face and
clothing, faint crackling arcs of light jumping off the edge of the sphere,
dust motes in the air catching the glow, camera slow push in

A rune circle ignites on the stone floor around a standing figure, thin
lines of gold light tracing the pattern outward from the centre, light
reflecting upward onto the figure's face and the surrounding walls, faint
haze rising from where the lines burn brightest

A staff strikes the ground and a shockwave of pale green light ripples
outward across the floor, nearby loose debris and dust kicked up and
scattered by the wave, the light fading to embers as it reaches the frame
edge, camera braces as if from the impact

Grok Imagine is worth naming here specifically because xAI's own documentation lists no negative_prompt, no seed and no style presets anywhere in its API corpus, checked 27 August 2026. Every constraint on an energy effect's colour, shape or intensity has to be stated positively, in the prompt itself; there's nothing to lean on besides the sentence.

Transformation and Morphing: How Do You Prompt a Change?

By describing the in-between state explicitly, because the model has no physical model of "becoming" and will otherwise cut straight from one state to the other rather than showing a transition.

A woman's reflection ripples in a mirror, the reflection slowly reshaping
into a wolf, fur pushing through where skin was, the transformation moving
from the face outward, dim moonlit interior, camera locked off on the
mirror

A block of ice sitting in direct sun slowly melts across the shot, edges
softening and rounding first, meltwater pooling and spreading outward
beneath it, a faint object embedded inside becoming visible as the ice
thins

If the transformation needs to run longer than one generation holds cleanly, several models let you chain clips by carrying the last frame forward: Seedance 2.5's return_last_frame produces a last_frame_url you feed into the next call, and Luma Ray 3.2 continues a shot via generation_id. Neither is a transformation feature specifically; both are a practical way to stage a slow change across two linked generations instead of asking one clip to do all the work.

Destruction and Debris: Why Does Debris Float?

Because gravity isn't simulated, it's guessed at from training data, and guesses drift furthest on objects the model has seen fewer of at your specific scale. Name the material and roughly how heavy it is; "chunks of drywall" and "steel girders" should not fall the same way, and telling the model which one you mean is the only lever you have.

A brick wall collapses inward, individual bricks and chunks of mortar
breaking free and falling, dust punching outward at the base of the
collapse, camera shakes and steps back, dust cloud rolling toward camera
and catching backlight

A wooden crate is struck and bursts apart, splinters and packing debris
scattering low and outward across the floor, largest fragments landing
first, dust settling over several seconds, static camera, hard side light

A glass window shatters from an impact off-screen, shards bursting inward
in a radial pattern, smallest fragments trailing last, light glinting off
falling glass, camera behind the glass, slow motion

Kling's own documentation notes a case worth knowing here: motion that is physically impossible for the described subject sometimes gets reinterpreted as a camera move instead of failing outright. If a destruction prompt keeps producing a camera pan when you asked for object motion, that reinterpretation is one plausible explanation, not a broken prompt.

Particles and Sparks: How Do You Direct Them?

By direction and falloff, since sparks with no stated trajectory tend to spray symmetrically in every direction, which reads as a canned effect rather than something that just happened.

An angle grinder cuts into metal, a dense stream of orange sparks arcing
downward and outward from the contact point, sparks dimming and dying
before reaching the floor, close-up, hard practical light from the sparks
themselves illuminating the operator's gloved hands

A campfire crackles at night, embers lifting off the flame in a loose
column, caught briefly by a gust and scattered sideways before dying out
mid-air, warm firelight on nearby faces, static wide shot

Vidu documents a movement_amplitude field for exactly this kind of directional control, but check before you trust it: Vidu's own pages disagree with each other about whether the parameter governs the objects in frame or the camera, and on the Q3 model the field is echoed back in the response without changing the output at all. Test it on your account before writing a prompt that depends on it.

Sci-Fi Holograms and Displays: What Sells the Illusion?

Treating the light as emissive and flat rather than physically lit, plus parallax as the camera moves. A hologram that behaves like a solid object with normal shading reads as a physical prop; one described as flickering, semi-transparent light with scanlines reads as a projection.

A translucent blue holographic map floats above a table, faint scanlines
flickering across its surface, edges slightly transparent, casting a cool
blue rim light on the hands reaching toward it, camera arcs slightly to
one side, revealing the display's depth through parallax

A holographic assistant flickers into existence beside a desk, thin
vertical scanlines and slight static at the moment it appears, its light
casting a faint blue glow on the wall behind it, the projection stabilising
after a beat of flicker

A wall-mounted sci-fi display flickers with scrolling data, occasional
horizontal glitch tearing across the screen, the display's glow lighting
the operator's face from below, camera static, dim otherwise-dark room

No vendor documents a hologram or display preset. This is entirely craft, which makes it one of the more reliably prompt-fixable categories here: the failure mode is almost always a missing interaction cue, not a model limit.

Slow-Motion Impact: How Do You Prompt the Moment?

By describing what keeps moving after the main subject stops, since real slow motion reads through secondary motion: fabric, liquid, dust and debris that lag behind and continue after the impact itself has landed.

A boxer's punch connects, impact captured in extreme slow motion, sweat
flying off on contact in a wide arc, the opponent's face and cheek
visibly deforming under the strike, camera locked off, hard directional
light catching the flying droplets

A dropped vase shatters on a stone floor in extreme slow motion, first
fragments separating before the sound would reach you, dust puffing
outward from the point of contact, glass shards continuing to bounce and
settle for a beat after the initial break

A diver enters water in slow motion, the surface breaking upward in a
crown-shaped splash, air bubbles trailing behind as the body submerges,
water surface rippling outward and slowly settling after entry

Genuine slow motion is usually a retiming choice made after generation, not a native parameter most vendors expose. Describe the physical slowness in the prompt itself, in-camera, rather than expecting a slow-mo toggle to exist.

Where Does Generated Video Actually Break?

Sort the failure before you touch the wording, the same way our image prompt troubleshooting guide sorts image failures into three buckets. Video effects split the same way, and the split matters because rewriting a parameter problem wastes exactly as much time here as it does with images.

Prompt-fixable. Missing light spill, no smoke displacement, no camera reaction. These are craft omissions: you didn't tell the model what should react, so nothing did. A rewrite genuinely fixes most of these, often on the first try.

Parameter or setting. Whether the clip has audio at all, whether you can place an effect at an exact second, whether an effect is a closed preset rather than a prompt target. Kling's settings.audio defaulting to off, Seedance 2.5's timestamp support versus 2.0's shot-only grammar, Pika's closed enum: none of these move with a better sentence, because the sentence was never the control.

Capability limit. Volume conservation in liquids and fire, debris obeying real gravity, smoke that fully respects a moving subject's mass. These are physics approximations breaking down, and no vendor publishes a simulation parameter that fixes them. Scale cues and camera language narrow the gap. They do not close it. If the shot has to be physically correct on delivery, the honest move is a different tool, or an artist, not another paragraph.

Which Plan Do You Actually Need for This?

We generate the prompt. We do not generate the video, and we are not a compositor: nothing here masks, keys or retimes anything, and none of it makes fire behave more like fire. What it does is stop you rewriting the same corrected prompt from memory every time you switch models.

Video Prompt Generation sits on the Advanced and Team plans, not Pro, per the comparison table on our pricing page. The Free plan includes 5 prompt enhancements per day, forever, per our FAQ page. If you're building longer effect-driven sequences rather than single shots, structuring a 20-second generation on LTX covers pacing a multi-beat take, and prompting Grok Imagine's Video 1.5 covers a model with real audio and reference support but zero negative-prompt or seed control. Pair either with our sound design guide once the visual effect is landing the way you want.