Prompting Crowds and Multiple Subjects

TL;DR: Multiple subjects AI image prompts fail two different ways: two named people can swap attributes (attribute binding), while six or more turns into a generic blur (density overload). Structure named subjects as labeled segments, cap named figures around five or six, and check your model's reference-image limits before promising a client a specific headcount.

What's Actually Different About Prompting Multiple Subjects?

A prompt with multiple subjects AI image models have to render is really two separate engineering problems wearing one name. The first is an identity problem: two or three named people, each with their own clothing, hair, and features, and the risk that the model attaches the wrong attribute to the wrong person. Researchers call this attribute binding, and our dedicated piece on the mechanism covers the research behind it in full, including why adding more descriptive detail makes a two-person prompt worse rather than better.

The second problem shows up once you cross roughly six named figures, and it isn't a binding error at all. It's density. Ask for ten individually described people and most current models stop trying to keep every one of them distinct; they render a plausible-looking mass instead, with faces and clothing blurring together the way a real photograph does at that range. That's not necessarily a failure. A wedding photo genuinely wants a crowd in the background and two or three sharp faces in front. The trick is deciding, before you write the prompt, which of your subjects need individual treatment and which ones are texture.

How Do You Prompt Two People Without Swapping Their Attributes?

Two people is the single most common multi-subject request, a couple, two colleagues in a before-and-after shot, an interview setup, and it's also exactly where attribute binding bites hardest, because two named subjects with distinct attributes is the smallest case researchers can reliably reproduce the failure in. The fix doesn't need the full labeled-segment treatment a six-person group needs; it just needs each person's description kept as a self-contained clause rather than woven together in one sentence.

Compare a prompt that invites the model to mix the two people up against one that separates them cleanly:

Risky: A chef in a white jacket and a server in a black apron stand
together in a busy restaurant kitchen.

Safer: In a busy restaurant kitchen: a chef wearing a white jacket
on the left, a server wearing a black apron on the right.

Two small changes carry the weight here: a spatial anchor for each person (left, right) and moving the shared setting to the front so it isn't competing with either subject's clause for the model's attention. Neither change requires more words, just a different arrangement of the same ones.

How Do You Structure a Prompt for Three to Six Named Subjects?

Past two people, spatial anchors alone stop being enough, and the single highest-leverage fix becomes structure rather than vocabulary. Instead of folding every person into one flowing sentence, give each subject their own labeled block, the same setting details repeated in each, so the model has an unambiguous unit to attach attributes within rather than a paragraph of pronouns and adjectives to sort out on its own.

Setting (shared): a rooftop bar at golden hour, string lights, city skyline behind them

Subject 1: a woman with short black hair, wearing a cream blazer, holding a glass of white wine
Subject 2: a man with a beard and glasses, wearing a navy suit jacket, leaning on the railing
Subject 3: a woman with long red hair, wearing a green dress, laughing mid-conversation

Composition: all three standing in a loose triangle, medium shot, natural evening light

This is a text-only version of the same idea vendors ship as structured or JSON prompting elsewhere on this site. If your tool accepts a JSON prompt, the same information holds together even more reliably as an explicit schema, because there's no prose ambiguity left for the model to resolve at all:

{
  "setting": "a rooftop bar at golden hour, string lights, city skyline",
  "subjects": [
    { "id": 1, "description": "woman, short black hair, cream blazer, holding white wine" },
    { "id": 2, "description": "man, beard and glasses, navy suit jacket, leaning on railing" },
    { "id": 3, "description": "woman, long red hair, green dress, laughing" }
  ],
  "composition": "loose triangle, medium shot, natural evening light"
}

Keep each subject's block short. The failure mode this avoids is stacking four or five adjectives onto each of three people, which hands the model fifteen attribute-subject pairs to keep straight instead of three. Two or three concrete, non-overlapping details per person is usually enough to tell them apart without overloading the pairing.

How Many People Can a Reference Image Actually Hold Together?

Text alone only gets you so far once a scene needs specific, recognisable people rather than described ones. Reference images anchor identity instead of asking the model to invent it, which sidesteps a chunk of the binding problem, but every model publishes a hard ceiling on how many you can combine, and that ceiling is easy to miss until a client asks for a six-person team photo. Our full walkthrough of prompting Gemini's image models covers the rest of what these models do; here's the part that matters specifically for headcount.

Google's Gemini API documentation currently names three distinct Nano Banana models, each with its own reference-image budget, and all three add up to the same total of 14 images per request:

Read the character-reference row literally: it's the closest thing to a documented answer for "how many distinct people can I safely composite into one shot." Nano Banana 2 Lite has no character-reference slot at all, which matters if a fast, cheap tier is tempting for a group-photo job specifically. Google's own model description for that tier says it plainly: it's "Not optimized for multiple reference inputs or multi-turn sequential editing." That line is about combining several reference images and iterating across turns, not a claim about accuracy inside a single crowded frame, so don't stretch it further than it says.

A six-person team headshot is the case where this ceiling actually bites. Nano Banana Pro's five-character slot covers a small team fine, but a sixth person has nowhere to go. The workaround isn't a bigger prompt, it's splitting the job: run the five highest-priority faces through the character-reference slots for accuracy, then either generate the sixth person separately and composite them in, or fall back to a detailed text description for whichever face matters least to the client. Trying to force a sixth reference image into a five-slot budget doesn't fail loudly. It just silently drops or blends one of the people you asked for, which is a worse outcome than planning around the limit up front.

How Do You Prompt an Actual Crowd, Not Just a Few Named People?

Once you've decided a scene genuinely needs a crowd rather than a small named group, describing every person individually stops being useful, or even possible within a prompt's practical length. Crowd prompting works better as environment description than as a subject list: name two or three anchor figures in the foreground with real detail, and describe everyone behind them as density, activity, and mood instead of as individuals.

Foreground anchors:
- a street vendor in a red apron, mid-gesture, calling out to customers
- a woman in a yellow raincoat checking her phone, standing just behind him

Background crowd (no individual description):
a busy evening market street, dozens of shoppers moving between stalls,
umbrellas up against light rain, string lights overhead, shallow depth of
field so the background reads as motion and color rather than distinct faces

Shallow depth of field and motion language do real work here: they give the model permission to render the background as texture, which is what it was going to do anyway past a handful of figures, instead of fighting it toward individuating people you never described.

Negative prompting for crowd scenes is genuinely inconsistent across vendors, which is worth checking before you assume a syntax carries over. Some models publish a real negative_prompt parameter with its own character limit. Google's Nano Banana models have none at all; their own prompting guidance recommends what they call "semantic negative prompts", contrasting "no cars" against "an empty, deserted street" as the positive alternative. Applied to a crowd, the same move means describing a quiet, mostly empty corner of the market rather than insisting there be no people in the back, which tends to hold up better across models than a boilerplate exclusion list. Our free negative-prompt generator handles the per-model syntax differences if you'd rather not track which vendor accepts which parameter.

What Goes Wrong Specifically in Dense Crowd Scenes?

Crowds concentrate the same anatomy-rendering problems that show up anywhere in AI image generation, just at higher frequency because there are more limbs, hands, and faces in the frame for something to go wrong with. Extra fingers or an extra limb is a documented, well-studied failure that gets worse as figure count rises, particularly at the edges of a crowd where the model has the least context for how a partially-occluded body should resolve. If a client-facing crowd shot needs to hold up to close inspection, plan for a masking or inpainting pass on the two or three figures nearest the camera rather than expecting one generation to get everyone right.

It's also worth being honest about what a crowd prompt is and isn't for. If the brief actually needs the same five or six people to stay recognisable across a whole campaign, a product shoot with a consistent cast, or a video sequence, that's a different job from composing them correctly in one frame. Keeping a character consistent across multiple generations or video clips is its own discipline with its own fixes, and reaching for crowd or group-photo techniques there will produce a different person in every shot.

A Pre-Flight Checklist for Multi-Subject Prompts

Run through this before you submit a prompt with more than one named subject:

1. Count your named subjects. One or two: write normally. Three to six: use labeled segments. Seven or more: describe a crowd, not a subject list.
2. Give each named subject two or three non-overlapping details, not five. Overlapping detail across subjects is what invites attribute binding.
3. Repeat the shared setting identically in every subject's block if you're generating separately to composite later.
4. Check your model's reference-image and character-reference caps before promising a specific headcount of recognisable people.
5. Decide foreground versus background explicitly. Anchor figures get individual description; background figures get density and motion language instead.
6. Reach for positive scene description before a negative-prompt list, especially on models like Nano Banana that don't support one at all.
7. Plan a masking pass on foreground faces and hands in any crowd shot that needs to survive close inspection.
8. Confirm whether the job needs one correct frame or a consistent cast across many. Those are different problems with different tools.

Palette control is the piece this checklist doesn't cover on its own: once a multi-subject composition is structurally correct, keeping specific colours accurate and distinct across every subject is a related but separate discipline, and our guide to exact colour and palette prompting covers it directly.