AI crowds. In, out or multiplied, plate untouched.
A crowd used to mean extras, catering and a second unit, or weeks of tiling and roto. Our AI populates a held wide with background extras from a prompt, multiplies the twenty people you could afford on the day, or lifts a crowd off a plate and gets the background back, in one pass, with the camera move, lighting and environment exactly where they were. Artist-checked before anything leaves.
Louvre courtyard · 10 s · 1080p · the same plate both ways: crowd lifted off, or repopulated from a prompt
ProRes
What a crowd unit and a paint team do in weeks, the model does in a pass
We are not a VFX lab with faster artists. We are an AI lab: the tedious, per-frame work that fills a vendor’s schedule is done by the model in a single pass, and artists are spent only where judgment matters. The result is the same deliverable, in a day or two instead of weeks.
| The job | Traditional VFX lab | Pixazo AI |
|---|---|---|
| Background extras | Cast, dressed, fed and directed, or tiled by hand | Generated from shot type, environment and density words |
| Crowd removal | Roto and paint, frame by frame | Removed with their shadows and the background rebuilt in one pass |
| Turnaround | Weeks | 24–48h on a standard shot |
| Cost driver | Head count and artist-hours | Shot class; one free test shot gives the number |
Where the artist still matters: extras that touch the cast, exact head counts and hero faces. Every frame is checked before it leaves.
Extras that hold their place through the shot
The slider at the top of the page shows the removal direction. This is the other one: a held wide populated by AI from shot type, environment and density words.
Generated wide shot · 8 s · 1080p · three hikers hold spacing and scale; a fourth walks in at about 7 s
Why this is the honest test
The same plate exists both ways. Run it one direction and the pass clears the crowd and rebuilds what was behind every person. Prompt the empty plate the other direction and it repopulates from shot type, environment and density words, with extras that hold spacing and scale as the camera moves.
What actually comes back
Three different jobs hide under the word crowd. Tell us which one you have and the deliverable follows. If you are not sure, send the shot and we will tell you.
| Deliverable | Format | Notes |
|---|---|---|
| Populated plate | EXR / DPX / ProRes, your frame rate and colour space | Your empty or sparse plate with background extras generated from the prompt. Grain matched to source. The hero action is untouched. |
| Duplicated crowd | EXR / DPX / ProRes | The extras you shot, tiled and varied across the frame so the same twenty faces do not repeat in a row. Works best from two or three passes of the same setup. |
| Clean plate | EXR / DPX / ProRes, plus the reusable clean plate | Every person removed and the background reconstructed for the full move. The clean plate is also delivered on its own so your comp team can reuse it. |
| Crowd element + alpha | Per-frame EXR with embedded alpha, or PNG sequence | Generated or duplicated extras as a separate element with a grey-valued matte, for comps you want to own. Delivered through our matte pipeline. |
| Density variants | 2–3 versions of the shot | sparse, steady foot traffic, packed. Cheap to produce once the first version is approved, so ask for them in the same brief. |
The question that decides the price: is the crowd behind the action, or does it have to touch it? Background extras behind a held wide are a standard shot. Extras that walk through the cast, hand them things or react to them are a different scope, and we will say so before quoting.
AI crowds cannot do every shot
The model is very good at a place and a density. It is bad at a number. Knowing that before the brief saves a round of notes.
- Held wides and simple moves. Courtyards, streets, beaches, stadium bowls, concourses. The Louvre plate above is a slow push; the generated hikers shot is a held wide with a fourth extra walking in at seven seconds.
- Density by description. A precisely named place implies how many people belong in it.
Sparse,steady foot traffic,packedandshoulder-to-shoulderall land differently and reliably. - Clean plates for the whole move. Removal reconstructs the background across the move, not frame by frame, so the camera can keep going.
- Duplication from thin coverage. Two or three passes of the same extras are enough to fill a stand or a square without the repeats reading.
- Period and genre. Wardrobe and era come from the prompt and the plate; a 1940s platform or a fantasy market are not harder than a modern street.
- Exact head counts. The model cannot count.
Exactly 43 people
becomes a density, not a number. If the count is contractual, it is a duplication job from real extras. - Extras that interact with the cast. Hand-offs, collisions, eye-lines and reactions need roto, tracking and an artist, and are quoted as compositing.
- Hero extras. A specific face, a specific wardrobe, a continuity character in the background: shoot them. Generated crowds are for the people nobody is looking at.
- Heavy foreground occlusion. Removing a crowd that the cast walks through means rebuilding what was behind moving actors; it is possible, but it is paint, not a pass.
- Whip pans and handheld chaos. Fast, erratic moves break the spacing and scale that make generated extras read as present. We will tell you on the test shot.
What works in the prompt: lead with shot type and environment, not a head count. Wide, held, rainy railway concourse at rush hour, steady foot traffic
gives the model a place and a behaviour; add 60 people
gives it nothing to anchor to.
How a crowd shot moves through
What moves the number
We do not publish a rate card, because the same word covers a 24-hour pass and a week of paint. These are the drivers; the test shot gives you the actual figure.
| Driver | Cheaper | Dearer |
|---|---|---|
| Job type | Populate an empty plate; clear a crowd behind the action | Clear a crowd the cast walks through; extras that touch the cast |
| Camera | Held, locked-off, slow push or pan | Handheld, whip pans, rack focus through the crowd |
| Length | Under 10 s per shot | Long oners; re-usable clean plates bring the per-second cost down |
| Deliverables | Finished plate only | Separate crowd element with alpha, density variants, clean plate on its own |
| Volume | A sequence or a show, priced per shot class | One-off shots |
Faster than a rate card: send one shot, ideally the one your current vendor is slowest or dearest on. We return a finished plate and a firm per-shot number for that class of work. If you are already being quoted elsewhere, tell us the number you need to land under.
Frequently asked questions
Can AI add a crowd to an empty shot?
Yes. To add crowd to video — or add extras to video — the AI crowd generator builds digital extras from the shot type, the environment and density words such as sparse, steady foot traffic or packed, and holds their spacing and scale as the camera moves.
Can AI remove a crowd from a shot?
Yes. Whether you need to remove people from video, remove crowd from video, or remove bystanders from video, a clean-plate pass removes the people and their shadows and rebuilds the background across the camera move. See AI clean plates.
Can I ask for an exact number of people?
No. The model works from density, not head counts. If a count is contractual, it is a crowd replication VFX job that lets you duplicate people in video from extras you shot.
Can generated extras interact with the cast?
Not in a single pass — full crowd simulation with hand-offs, collisions and eye-lines needs roto and an artist, and is quoted as compositing.
How long does it take and what does it cost?
24–48 hours on a standard shot. We do not publish a rate card. Send one shot free and you get a finished result plus a firm per-shot number for that class of work.
Crowds rarely arrive alone
A clean plate is usually the first half of a cleanup and paint brief, and a crowd element is only useful with a good alpha matte behind it. Archive crowds come through restoration and upscaling first. All of it sits inside the AI VFX Studio: 15 production capabilities, 200+ models, human-finished on every shot.
Send us the empty square. Or the full one.
One shot, no charge, no commitment. You get a finished plate, the clean plate or crowd element alongside it, and a firm number, and you compare all three against what you have already been quoted.
Send a test shotCrowd generation · Clean-plate removal · 24–48h standard shot · Artist-finished
