How to Generate Crowd and Background Extras in AI Vertical Drama Without Faces Repeating
A wedding reception. Forty guests behind the lead as she reads the message that ends her engagement. The reviewer pauses on frame twelve. The same woman in the green dress is standing at two tables at once. Two frames later, a man near the bar has the face of the groom. The scene has to go back.
Repeating faces are one of the most common reasons a crowd shot fails review in AI vertical drama. Generators tend to collapse toward a small set of familiar faces when asked for many people at once. They also tend to borrow features from the strongest character reference in the prompt, which is usually the lead. The result is a background that looks populated at a glance and wrong on a second viewing. Viewers on a phone screen may not name the problem, but they feel it. The scene reads as artificial.
This guide sets out a working method for crowds and background extras. It covers how to plan the crowd before generation, how to build a pool of extras that stays distinct, how to protect lead faces from bleed, and how to check a crowd shot before it reaches review.
1. Why Crowds Fail in the First Place
Understanding the failure modes makes them easier to prevent. There are three that account for most rejected crowd shots.
The first is convergence. When a prompt asks for a large group of people with little description, the generator fills the space with averaged faces. Those averaged faces resemble each other because they are drawn from the same statistical centre. Ten people become three faces repeated in different clothing.
The second is reference bleed. If the shot includes a character reference for the lead, the generator may apply some of those features to nearby figures. This is most visible with distinctive traits such as a particular hairstyle, a strong jaw or a recognisable eye shape. A background extra suddenly looks like a sibling of the lead.
The third is temporal drift. In a moving shot, background faces can change between frames. A guest who is clearly middle aged at the start of a pan is younger by the end. This happens because background figures receive less attention in the generation process than the subject, so they are less stable over time.
Each failure has a different fix. Convergence is solved by planning and description. Reference bleed is solved by separating lead and crowd in the workflow. Temporal drift is solved by framing, focus and shot length. A good crowd method handles all three.
2. Plan the Crowd Before Anyone Writes a Prompt
The most effective step happens before generation begins. Decide what the crowd needs to do in the story, how much of it will be visible, and how many distinct people the viewer will actually see.
Decide the Crowd Role
A crowd in vertical drama usually plays one of a few roles. It can establish a place, such as a busy restaurant or an office floor. It can react, such as guests gasping at a public slap. It can isolate, such as a lead standing alone in a sea of strangers. Each role needs a different amount of facial detail. An establishing crowd can be soft and distant. A reacting crowd needs a few readable faces. An isolating crowd can often be shot almost entirely out of focus.
Count the Readable Faces
In a 9:16 frame, the number of readable background faces is far smaller than in a horizontal frame. A typical medium shot might show three to six faces clearly enough to matter. Count those for each scene. That number, not the total head count, is what the generation plan must protect. Everyone else can be suggested through shape, colour and movement.
Write a Crowd Sheet
For scenes with recurring crowds, write a short crowd sheet. List each readable extra by a simple label, such as Guest A or Waiter B. Give each a brief description covering approximate age, build, hair, skin tone, and one distinctive wardrobe element. Keep these descriptions deliberately varied across the list. The crowd sheet becomes the reference document for every operator who works on the scene, and it stops the same averaged face from being generated by default.
3. Build an Extras Pool That Stays Distinct
For series that return to the same locations, a reusable pool of background extras saves time and protects consistency. The same cafe regulars or office colleagues can appear across many episodes without the generator inventing new ones each time.
Generate Extras as Their Own References
Create each recurring extra as a standalone reference image, generated separately from any lead character. Use neutral lighting and a clear, front facing view. Give each one a distinct combination of age, build and features that will not be confused with others in the pool. Name and store these references in the same system used for lead character references, with a clear label that marks them as background.
Build in Deliberate Variety
When building the pool, push variety on purpose. Vary age across a wide range. Vary body shape, hair length and texture, facial hair, glasses and accessories. Vary skin tone in a way that reflects the setting of the story. A pool where every extra falls within a narrow band of appearance will converge in generation no matter how carefully it is prompted.
Keep the Pool Separate From the Leads
Store background references in a clearly separate folder or category from lead references. This prevents an operator from attaching the wrong reference by accident, and it makes it simple to confirm that no lead reference is being used in a crowd only shot. Separation in storage leads to separation in generation. The two should never share a reference slot unless the shot calls for both.
4. Protect the Lead From Reference Bleed
Reference bleed is the most damaging crowd failure because it touches the most important face in the frame. A background figure that resembles the lead confuses the viewer and weakens the drama of the scene.
Generate in Layers Where Possible
One reliable approach is to generate the background crowd and the lead separately, then combine them in compositing. The crowd plate is generated with no lead reference attached at all. The lead is generated against a clean or matched background with full reference support. The two are then combined, with lighting and grade matched in post. This removes the possibility of bleed entirely, at the cost of a compositing step.
When Generating Together, Push the Crowd Back
When a scene must be generated in a single pass, give the lead clear spatial priority. Place the lead in the foreground and the crowd in the midground or background. Use shallow depth of field so the crowd softens with distance. Describe the crowd in terms of wardrobe, movement and placement rather than faces. The less the prompt asks the generator to render background faces in detail, the less likely it is to borrow from the lead.
Describe Contrast Explicitly
When a few background faces must be readable in the same shot as the lead, describe them in deliberate contrast. If the lead is a young woman with long dark hair, describe the nearest extras as an older man with grey hair, a woman with short blonde hair, a teenager in a cap. Contrast gives the generator a clear reason to make them different. Without it, the nearest default is often the lead.
5. Control Temporal Drift in Moving Shots
Crowd faces that change during a shot are harder to catch than repeats, because each single frame may look fine. The problem only appears in motion.
Keep Crowd Shots Short
Background stability decreases as shot length increases. Short crowd shots give the generator less time to drift. Vertical drama pacing already favours short shots, so this usually fits the edit naturally. When a longer crowd shot is needed, consider building it from two or three shorter generations joined at a natural cut point, such as a reaction or a turn.
Limit Camera Movement Across Faces
A slow pan across a crowd asks the generator to maintain every face it passes. A static shot or a gentle push toward the lead asks much less. Where a scene needs movement, move toward the lead rather than across the crowd. This keeps attention on the stable subject and reduces the number of background faces that must hold over time.
Use Focus as a Tool
Soft focus on the background is not only an aesthetic choice. It is a consistency tool. A background that is slightly soft hides small changes in face shape and expression. In a 9:16 close up or medium shot, a softened background is natural and expected. Use it whenever the story does not require the crowd to be sharp.
6. Use Simulation Thinking for Large Crowds
For very large crowds, such as a stadium, a street protest or a packed train station, generating every visible person as a detailed individual is neither practical nor necessary. A better approach borrows from the logic of crowd simulation used in film and games, where large numbers of figures are treated as a system rather than as individuals.
Separate Hero Extras From the Mass
Divide the crowd into two groups. Hero extras are the few faces near the camera that the viewer will read. The mass is everyone else. Hero extras come from the extras pool with full reference support. The mass is described in terms of density, colour, movement and direction, with no individual faces requested.
Describe Behaviour, Not People
For the mass, describe what the crowd is doing rather than who is in it. People moving toward an exit. A crowd parting around a central figure. Guests turning toward the stage. Behaviour gives the generator a pattern to follow. It also creates movement that reads as natural on a phone screen, where individual faces in the far background are barely visible.
Build Depth in Layers
Large crowd scenes benefit from being built in depth layers. A foreground layer with hero extras. A midground with softer figures. A background that is little more than shapes and light. Each layer can be generated or adjusted separately, and the combination produces a sense of scale without asking any single generation to render hundreds of distinct faces.
7. Check Every Crowd Shot Before Review
A short, consistent check before a crowd shot is submitted catches most problems while they are still cheap to fix. Operators and quality reviewers can use the same checklist.
Scan for duplicates. Look across the frame for any two figures with the same face, and repeat the scan at the start, middle and end of the shot.
Compare against the lead. Check that no background face shares a distinctive feature with any principal character in the scene.
Watch the shot in motion at full speed and at reduced speed, looking for faces that change age, shape or expression without reason.
Confirm against the crowd sheet. Check that readable extras match their descriptions and that recurring extras look as they did in earlier episodes.
Check continuity with adjacent shots. A guest seated at a table in one shot should not be standing at the bar in the next unless the scene allows for it.
Log any failure with the frame number, the type of failure and the likely cause. Over a series, these logs show which scene types and prompt patterns produce the most crowd errors, which allows the team to adjust its method before the same failure repeats.
8. Keep Recurring Crowds Consistent Across Episodes
Vertical drama often returns to the same places. The family dining room. The office reception. The restaurant where the couple first met. Viewers notice when the background changes without reason, even if they cannot say what changed.
Treat recurring background figures the same way as recurring locations. Record which extras appear in each location, where they usually stand or sit, and what they usually wear. Reuse the same references when the location returns. If an extra must change, such as a new receptionist after a story event, introduce the change deliberately and log it so later episodes follow the new version.
This record does not need to be complex. A short entry in the continuity notes for each location is enough. What matters is that the information exists and that every operator working on the location can find it. Consistent backgrounds make a world feel lived in. Inconsistent ones make it feel generated.
Axis AI Studios Perspective
Crowd scenes are where many AI vertical drama productions show their seams. The lead holds. The background gives the game away. Axis AI Studios treats background figures as part of the production plan rather than an afterthought, because a scene is only as convincing as its least stable face.
The approach at the studio rests on planning the crowd before generation, keeping a distinct pool of background references separate from leads, controlling shot length and focus to limit drift, and checking every crowd shot against a consistent list before it reaches review. None of these steps is complicated. The value comes from applying them every time, across every episode, so that a series returning to the same cafe for the tenth time still feels like the same cafe.
For platforms and commissioning teams looking for a production partner that holds the background to the same standard as the lead, the conversation starts at business@axisaistudios.com. Write to Axis AI Studios.
FAQ
Why do AI generators repeat faces in crowd scenes?
When asked for many people with little description, generators tend to produce averaged faces drawn from the same statistical centre, so several figures end up looking alike. The effect is stronger when a lead character reference is present, because nearby figures can inherit its features. Planning readable extras and describing them in deliberate contrast reduces both problems.
Should crowds and leads be generated in the same pass?
They can be, but separating them is the most reliable way to prevent reference bleed. Generating the crowd plate with no lead reference attached and compositing the lead afterwards removes the risk entirely. When a single pass is necessary, place the lead in the foreground, soften the background and describe the crowd by wardrobe and behaviour rather than faces.
How many distinct background faces does a vertical drama scene need?
Fewer than most teams expect. In a 9:16 medium shot, usually only three to six background faces are readable enough to matter. Those faces need distinct references and descriptions. Everyone else can be suggested through shape, colour, movement and soft focus.
Further Reading
For scenes where two principal faces share the frame and both must hold, the guide to generating two character scenes without losing either face covers framing and reference methods that keep both leads stable.
For teams building a pool of recurring extras alongside their lead references, the walkthrough on version control for character references covers how to name, store and update references across a long series.
For productions that return to the same cafe, office or dining room across many episodes, the method for building a reusable location library covers how to keep interiors consistent from one episode to the next.

Let's set
the new standard together.
If you're working on something, we'd like to hear about it.
