
Smoke reads as a soft grey blur, which is why it looks like the simplest effect in the world. In practice it is one of the hardest, because the human eye is extraordinarily good at reading fluid behavior. Real smoke is a turbulent volume: density changes through its depth, it self-shadows, it catches a rim of light from whatever sits behind it, and its edges break into filaments that never repeat in the same way twice.
Procedural noise alone gets you a texture that is statistically smoke-like and behaviorally nothing like smoke. It sits still, breathes uniformly, and reveals its repeating pattern within a few seconds of screen time. Hand-animated 2D smoke drawn frame by frame can be beautiful, but it costs hours per second of footage and gets expensive the moment a director asks for a different color or a slower rise.
The workflow in this article is a hybrid, and the hybrid is what makes smoke affordable again. You author a small set of high-quality static smoke plates in Paint.NET, a fast layer-based raster editor that most creators already have installed. You then hand those plates to an AI video tool that understands volumetric motion and let it animate them. Finally you composite, retime, and grade the result like any other visual effects element.
Splitting the job this way puts each tool on the part it does best. Paint.NET handles precise, art-directed density, silhouette, and color. AI video handles turbulence, dissipation, and the slow rolling motion that is so tedious to key by hand. Your compositor handles integration. It also keeps you in control of your own look instead of rolling the dice on a single text prompt and hoping.
The payoff is a library. Once you have built and organized a set of plates, you can generate twenty variations of the same shot in an afternoon instead of rebuilding the effect from scratch each time.
Smoke fails more often at the planning stage than at the painting stage. Before you open the editor, write a one-paragraph shot brief that answers six questions.
Answer those and you have a spec you can actually paint against. Skip them and you will produce generic fog that looks acceptable in isolation and wrong in the cut.
The format decisions you make now determine every technical choice later.
| Decision | Why it matters | Practical default |
|---|---|---|
| Plate background | Determines the blend mode you can use later | Pure black for additive, transparent for coverage |
| Resolution | AI models soften detail, so modest plates often survive | 1920x1080 plates for 4K composites |
| Frame rate | Many AI tools output a fixed rate | Match your timeline to avoid retiming artifacts |
| Plate count | More variants means more editorial choice | 6 to 12 per look |
| Generation length | Longer clips drift further from the source plate | 3 to 5 seconds per generation |
| Color | Warm and cool smoke behave differently against skin | Author neutral, grade later |
Collect three to five real smoke clips or stills that match your intent: a cigarette curl, a flare stack plume, dry ice spilling over a table, dust kicked up by a tire. Note direction, speed, turbulence scale, and how the light behaves. This board is what you will compare against when you review generations, and it is far more useful than your memory of what smoke looks like.
The goal here is not a finished effect. It is a set of static images with clean density, a readable silhouette, and true black where there should be nothing. Everything downstream depends on that.
Start a new image at your target resolution with a black or transparent background. Build the plate across four layers, each on its own so you can adjust them independently: base density, mid structure, edge detail, and rim highlight. Name them. You will thank yourself when you revisit the file in a week.
Use the built-in cloud generators under the Render menu to lay down large-scale structure. Keep the scale large so you get a few big volumes rather than uniform static. Follow with a heavy Gaussian blur to soften the whole thing into a mass rather than a texture.
Then do the single most important step in the entire workflow: crush the blacks. Use Levels or Curves to push everything below roughly ten to fifteen percent brightness to absolute black. AI video models interpret black as "nothing here," and compositors do the same. A plate with a muddy grey background will produce a muddy grey halo around your smoke in the final shot, and no amount of grading will fully remove it.
Smoke needs a definite shape to read as smoke rather than as haze. Select regions with a feathered freeform selection, then push them with distortion tools such as Bulge or Twist. A directional motion blur is the fastest way to imply a flow direction, and it also helps the AI model understand which way the smoke should travel.
If you have a community plugin pack installed, a smudge tool is excellent for pushing density around by hand, and an alpha mask utility lets you reuse one good silhouette across several plates with different internal detail.
Large smooth masses read as fog. Smoke reads as smoke because of the small structures inside it. Duplicate the mid layer, apply a fine-grained cloud or texture distortion to it, then mask that detail to the edges of the main mass so the interior stays smooth and the perimeter gets torn up. A light motion blur along the flow direction ties the detail to the movement.
Adding a second texture pass at a different scale and blending it at low opacity gives you variety without obvious repetition, which matters when the same plate is reused across multiple generations.
Smoke that looks beautiful at 100 percent zoom often vanishes at 25 percent, and 25 percent is roughly how it will appear on a phone screen. Toggle the zoom a few times while you work. If the shape disappears, increase the contrast between the dense core and the background rather than adding more detail.
Export high-quality PNGs with a clear naming convention that encodes direction, density, and version, for example plumes_rising_dense_v03_black.png. Keep a contact sheet of every plate in the set so you can scan them at a glance when generating. Consistency here saves more time than any single editing trick.
With a plate library in hand, you have four practical ways to animate it. They behave differently, and choosing the wrong one is the most common reason a smoke shot looks fake.
This is the reliable default. You supply a single plate and a prompt that describes only motion, and the model keeps the design largely intact while adding flow, turbulence, and dissipation. Keep the generated clip short, three to five seconds, and keep the motion instruction modest. The tighter the prompt, the less the model invents.
Supplying both a starting and an ending plate gives you control over the arc of the shot: dense to wispy, low to high, left to right. This is the right choice for a reveal, a title transition, or any moment where the smoke has to clear by a specific time.
When smoke needs to interact with a subject, a hand, a car, or a falling object, generating the entire shot can be more convincing than compositing a plate over live action. You lose control of the plate design and gain interaction, which is often the better trade for a hero shot.
Feeding the model a reference image of a specific material, such as dry ice, steam from a kettle, or fine dust, pushes the look toward that material's behavior. This is the fastest way to match a new shot to an existing library of plates you already like.
Smoke prompts fail in two directions. Too short and the model invents a whole scene. Too long and it tries to satisfy every clause and produces mush. The sweet spot for image-to-video is a sentence describing motion, direction, speed, and light.
| Intent | Terms that move the needle |
|---|---|
| Motion | billow, drift, curl, roll, rise, seep, dissipate, wisp |
| Direction | upward, left to right, toward camera, across frame |
| Density | thin, wispy, dense, opaque, layered haze |
| Lighting | backlit, rim lit, volumetric shaft, ambient, low key |
| Camera | locked off, static tripod, slow push in |
Say "the smoke curls upward and to the left, thin at the edges and dense at the base, backlit with a warm rim" and you will get something noticeably different from "smoke moving." Be specific about one thing at a time rather than stacking adjectives you cannot evaluate.
Generate eight variations, not one. Keep the seed when a result is close and iterate on the prompt from there. Save the prompt text inside your project file so you can reproduce a look months later, because prompt behavior shifts between model versions and you will want the reference.
Text and watermarks, obviously, but also sudden brightness changes, hard geometric edges, morphing into objects, and heavy camera shake. Smoke should move within the frame; the frame itself should stay still unless you deliberately want movement.
A raw generation rarely drops into a timeline unchanged. Compositing is what turns a decent AI clip into a believable element.
For plates exported on black, Screen or a linear additive blend is the standard choice because black becomes transparent and bright values add light. Lighten gives a softer, less blown-out result. Overlay is useful when the smoke should pick up and react to the color underneath it rather than glow independently. For transparent plates, use normal blending with the alpha channel and check for fringing.
AI-generated smoke frequently moves slightly faster than real smoke, because the model is optimizing for visible change. Slow the clip to seventy or eighty percent and it immediately reads as heavier and more physical. For a richer look, layer two generations moving in slightly different directions and speeds; the parallax between them creates depth that a single pass cannot.
Track or match-move the smoke to the camera if the shot is moving. Rotoscope the smoke behind foreground objects so it appears to occupy the same space as the scene. Add a subtle light wrap where the bright smoke meets a dark silhouette edge. Match grain and noise to the surrounding footage, and slightly defocus the layers you intend to read as further away. These small moves do more for believability than another generation pass.
Match your black level to the shot, then grade the smoke toward the scene's color temperature. Warm practical lights should tint the smoke warm; a cold moonlit exterior should tint it cool. Keep saturation low, since highly saturated smoke reads as colored gas rather than smoke unless that is the intent.
The model reinterprets the texture slightly every frame, so the smoke shimmers. Fix it with shorter clips, a lower motion strength, or by cross-dissolving between two takes with a twenty-frame overlap. Adding fine grain over the composite also masks residual flicker.
Usually caused by an over-described prompt or an ambiguous plate. Simplify the prompt to motion only, clean up the plate so the silhouette is unmistakable, or mask the problem region and composite a clean pass over it.
The plate's background was not truly black. Re-crush the blacks in Paint.NET before exporting and re-generate. If the halo persists, it is alpha fringing from a transparent export; try a black-background plate with additive blending instead.
One pass produces one motion, and one motion looks synthetic. Stack two or three passes with different directions and speeds, and vary the retiming on each.
Models often bake a warm or cool cast into the output. Author neutral plates, then grade after compositing rather than fighting the model. If the drift is severe, desaturate the clip almost entirely and rebuild the color in your compositor.
Here is how a ten-second smoke shot breaks down in practice.
| Stage | Time | Output |
|---|---|---|
| Brief and reference board | 20 min | Shot spec and three references |
| Plate authoring in Paint.NET | 45 to 60 min | 8 to 12 plates |
| Generation batches | 15 min queue, 20 min curation | 6 to 10 usable clips |
| Compositing and grading | 60 to 90 min | Finished shot |
| Review and versioning | 20 min | Delivery file and project archive |
Run generations in the background while you paint the next plate set. The queue time is free, and batching your prompts keeps you from context-switching between art direction and technical review.
Before you call the shot finished, walk through this list once.
Yes. The built-in cloud renders, Gaussian blur, Levels, Curves, motion blur, and the distortion tools are enough to build a solid plate set. Plugins mainly add speed and convenience, not capability you cannot reach manually.
Eight to twelve per look is a comfortable working number. That gives you enough variety to avoid obvious repetition across a sequence while keeping the set small enough to manage.
Start with Screen for black-background plates, switch to Lighten if the smoke looks blown out, and use Overlay when you want the smoke to take on the color of the scene behind it rather than add brightness.
Three to five seconds is the practical range. Longer clips tend to drift away from the source plate and accumulate flicker, which costs more time to fix than generating two shorter takes and joining them.
For plate authoring, no. For generation, a machine with a modern GPU makes iteration dramatically faster, but you can also work with hosted tools and batch your prompts to make queue time productive.
Yes, and it is one of the best uses of the technique. Generate a clip, then use a ping-pong or cross-dissolve technique to hide the seam. Keep motion slow and mostly upward so the loop point is hard to spot.
Almost always because the plate lacks a defined silhouette. Fog is uniform; smoke has a source, a direction, and a clear density gradient. Rebuild the plate with a stronger core, a directional flow, and torn edges, and the result will read correctly.

Alexander
Vidu Q4 alternatives compared: Kling 4.0, Seedance 2.5, Wan 3.0, and Runway Gen-4 — 4K support, input modes, durations, audio, and what each costs in credits on Domer.
The top 4K AI image generators on Domer compared: Flux 3 Klein, Nano Banana 2.1, GPT Image 2.5, GPT Image 2, and Nano Banana Pro — 4K credit costs, reference limits, and speed.
Nano Banana 2.1 vs GPT Image 2.5 Flare compared: credit costs per resolution, aspect ratios, reference images, text rendering, and which tasks fit each model.