Multiple Consistent Characters in One AI Scene

To keep multiple consistent characters in one AI scene, give each character their own reference image and their own fixed description block, tell the model which image is which person and where each one stands, and repair one character at a time instead of regenerating the whole scene. Stay inside what your model documents: character consistency for up to four characters on Nano Banana 2.1 and five on Nano Banana Pro, up to ten references on FLUX 3 Image, and four on Midjourney's Edit Model.

Limits checked in October 2026 against Google's Gemini image generation docs, Black Forest Labs' FLUX.2 overview, character consistency guide and FLUX 3 Image editing and bounding box docs, Midjourney's Edit Model docs, OpenAI's image generation guide and ComfyUI's multiple LoRAs and inpainting tutorials. The method below is assembled from those vendors' prompting guidance. It is not a scored test of ours, and the example characters are invented.

This page assumes you can already hold one face. If not, start with the consistent character AI guide, which covers the single-character methods this one builds on.

What you will have at the end

A prompt structure that keeps two to four characters separate, an order for assembling a shared scene, and a repair routine for the frame where one face slipped. It works in any tool that reads reference images, and it adapts to LoRAs.

Why two characters drift when one does not

With one character, every trait in the prompt belongs to the same person. With two, the model has to decide who owns the copper hair and who owns the beard, and it gets no help from a sentence that lists both. The usual failures are swapped features, two faces averaged into siblings, and a wrong head count. Midjourney's docs acknowledge the first one directly, with a tip for when features get mixed up between characters.

The fix is ownership. Each trait, outfit and reference image is assigned to a named character, and each character is assigned a place in the frame. That is the whole method; the rest is doing it in the right order.

AI consistent multiple characters: what each tool documents

Check the limit before you plan a cast. A scene with five recurring characters is outside what most models document.

Two notes on that table. Google's limit is about character images, not total references, so on Nano Banana 2.1 four characters with one image each uses the whole documented allowance. And FLUX 3 Image is the only one here with documented placement control: a prompt can end with rows that give each element an id, a box and a description, with box coordinates from 0 to 1000. Black Forest Labs says boxes guide placement and are not clipping masks.

The character-blocking method

Blocking is the theatre word for deciding where each actor stands. Here it means writing the prompt in owned blocks instead of one flowing sentence.

A full prompt has five parts, in this order: the image roles, one block per character, the blocking, the scene, and a keep line. This is an example with two invented characters:

Image 1 is MARA. Image 2 is JONAH.

MARA: woman in her late 20s, shoulder-length copper hair with a centre parting,
freckles across the nose, slim build. Wearing a grey wool coat.

JONAH: man in his early 30s, short black curly hair, trimmed beard,
broad shoulders. Wearing a navy bomber jacket.

Blocking: MARA sits on the left of a cafe table, facing the camera.
JONAH sits on the right, turned toward MARA, mid-laugh. Two people only.

Scene: a small cafe by a window, soft morning daylight, eye level, 4:5 vertical.

Keep MARA's face and hair exactly as in image 1, and JONAH's exactly as in image 2.
Do not blend their features.

Every part is there for a documented reason. Naming images by number is how Black Forest Labs' docs address references. The separate blocks follow Google's advice to be hyper-specific and to break complex scenes into steps. The keep line borrows the wording of Google's own edit example, which starts with "keep everything the same". For ready-made variations, our Gemini face consistency prompt pack has a group-scene set.

Consistent characters in the same scene: the assembly order

Order matters more with two characters than with one, because every regeneration puts both faces at risk again.

Step six is where most of the saving is. Google calls multi-turn conversation the recommended way to iterate on images, and its consistency example says to include previously generated images in later prompts. An edit that names one person leaves the other face alone; a fresh generation rolls the dice on both.

Two characters, AI image consistency: copy the features word for word

The second half of the method is boring on purpose. Once a block is written, it is pasted, never retyped. A model has no way to know that "the redhead" in one prompt and "the woman with copper hair" in the next are the same person.

This is a writing convention, not a measured result. It follows from how these models read prompts: the more consistently a trait is worded and the more clearly it is attached to one person, the less the model has to guess. The same idea sits behind the description blocks in our ChatGPT consistent character guide and the identity block in Flux consistent character.

Two character LoRAs in one scene

If your characters live in LoRAs, the blocking idea still holds, but the assembly changes. Loading two face LoRAs together does not give you two faces.

1. Generate the scene with neither LoRA loaded, or with only one. Get the composition, poses and light right first.
2. Mask the first character's face in the mask editor and inpaint it with only that character's LoRA loaded. Result: one correct face, and the other person untouched.
3. Mask the second character and repeat with the second LoRA.
4. Check both against their heroes, then save the frame as a reference for the next shot.

ComfyUI's inpainting tutorial covers the masking steps, and its grow-mask setting softens the edge of the repainted area. If a LoRA puts its face on every person in the frame even when used alone, it has absorbed more than the identity; our guide to face LoRA overfitting covers that diagnosis.

Group photo, AI, same characters: going past two

Three and four characters work the same way with more blocks, as long as the tool documents that many. Past the limit, change the approach instead of adding references and hoping.

• Stay inside the documented number. Four character images on Nano Banana 2.1, five on Pro. Google's own example asks for an office group photo of the people in five attached portraits.
• Combine characters into one reference. Midjourney's docs suggest putting the characters into a single reference image when their features mix. A clean side-by-side sheet also frees up reference slots.
• Build in passes. Compose two or three people, lock the frame, then add the next person with an edit that keeps everyone already in the frame unchanged.
• Keep faces large enough to matter. In a wide group shot each face is a small part of the frame. Black Forest Labs notes that elements given very small boxes often fail to appear at all, so frame tighter, or fix faces afterwards with edits.

For moving scenes, the same cast needs a different binding step; AI video character consistency covers that. And if references stop holding once the cast is on screen for weeks, that is the point to train a LoRA per character; the AI Influencers program takes you through reference tools first, then LoRA training for consistent faces and ControlNet for poses.

Troubleshooting a multi-character scene

Whose faces you can put in a scene

Two characters means two sets of rights. Google's docs carry a plain reminder to make sure you have the necessary rights to any images you upload, and every Nano Banana image includes a SynthID watermark. Build scenes from original personas. If a real person appears, including you, that person needs to have agreed to the use, and a second real person needs their own consent, not yours. The likeness section of consistent characters without a LoRA lists the rules by tool.

Multiple consistent characters: FAQ