Multi-Character Scenes: Keeping Two Locked Identities Consistent

Getting one AI character to stay consistent across a set of images is hard enough. Putting two of them in the same frame, both recognizable, both holding their own face, is where most workflows fall apart. The two identities blend into a single average-looking person, or one character quietly borrows the other's hair, jaw, and eyes halfway through the batch.

I hit this wall constantly when building scenes for virtual influencers who have a recurring co-star. After a few hundred failed generations I settled on a workflow that actually holds. This guide walks through why the bleed happens and how to lock two identities at once.

Quick Answer: Two characters stay distinct when you keep their identities in separate spatial lanes and separate conditioning. In practice that means regional prompting or a masked two-pass approach so each face is driven by its own reference, never a shared one. For most creators the reliable path is generating the scene composition first, then locking each face in a second masked pass. Tools like Apatero.com handle the identity lock for you, and in ComfyUI you get the same result with regional conditioning plus per-region IPAdapter.

:::tip[Key Takeaways]

• Face-bleed happens because a single prompt and a single identity signal get applied to the whole canvas
• Separate each character into its own spatial region so their conditioning never overlaps
• Lock faces in a second masked pass rather than hoping the first pass gets both right
• Keep one reference per character and never average two references into one
• Compose the scene first, identities second, for the most stable results :::

Why Two Characters Bleed Into One

Before fixing it, it helps to understand what the model is actually doing. A diffusion model does not know that your image contains two separate people. It sees one canvas and one set of instructions. When your prompt says "a woman with red hair and a man with a beard," the model tries to satisfy both descriptions everywhere at once.

The result is predictable. Attributes leak across the frame. The woman picks up a hint of the beard structure in her jaw. The man's hair warms toward red. If you add a face reference on top, that reference gets applied globally, so both people drift toward the same face.

This is not a bug you can prompt your way out of. Writing "two distinct people" harder does not help. The fix is structural. You have to give each identity its own space and its own signal.

Understanding this early saves hours. Most people try to solve a spatial problem with better adjectives, and better adjectives were never going to work.

The Core Principle: Separate Lanes

Every reliable multi-character method comes down to one idea. Each character gets a lane, and the model is only allowed to think about that character inside that lane.

There are three ways to enforce lanes, in rough order of ease:

• Regional prompting, where you draw boxes and assign a prompt to each box
• Masked identity passes, where you generate the scene first and then repaint each face inside a mask
• Latent compositing, where you generate each character separately and stitch them into one latent before a final harmonizing pass

You do not need all three. For two characters, regional prompting or a masked second pass will carry you most of the way. I lean on the masked pass because it is the most forgiving when poses overlap.

If you want a deeper primer on the region tooling itself, our regional prompter ComfyUI guide covers the node setup in detail. This post is about the identity strategy on top of it.

Method One: Regional Prompting for Clean Compositions

Regional prompting is the fastest way to get two characters that do not blend. You divide the canvas into regions, usually a left half and a right half, and give each region its own prompt and its own identity conditioning.

The workflow looks like this in practice:

1. Set your base prompt to describe only the shared scene, such as the setting, lighting, and mood
2. Define two regions, one per character, with a small overlap in the middle
3. Assign each character's description and face reference to its own region
4. Lower the base prompt weight so the regional prompts dominate their zones

The reason this works is that the identity signal for character A never touches region B. Their conditioning lives in different parts of the canvas, so they cannot average together.

Regional prompting has one weakness. It assumes the characters occupy roughly separate areas. If your two characters are hugging, overlapping, or one is behind the other, the hard region boundary starts to show. For those poses, the masked pass is better.

Method Two: The Masked Two-Pass Lock

This is my default for anything with contact or overlap. The idea is simple. Do not try to get both faces right in one shot. Get the composition right first, then lock each face separately.

Here is the full sequence:

• Pass one generates the whole scene with both characters roughly described. Do not worry that the faces are generic or slightly wrong yet. You are only locking pose, framing, and clothing here.
• Draw a mask over the first character's face and run an identity pass using only that character's reference. The rest of the image stays frozen.
• Draw a mask over the second character's face and repeat with the second reference.

Because each identity pass sees only one face and one reference, there is zero opportunity for the two identities to mix. Character A is repainted while character B is literally masked out of the computation.

The trick that makes this reliable is keeping your denoise strength moderate on the face passes. Push it too high and the new face ignores the head angle from pass one. Keep it in a middle range and the locked face snaps onto the existing pose cleanly. If you already build single-character sets, this is the same discipline described in our face consistency techniques, just applied twice with masks.

One Reference Per Character, Always

The single most common mistake I see is people feeding a stack of mixed references into one IPAdapter and hoping the model sorts it out. It will not. Averaged references produce an averaged face, and that average is exactly the blended look you are trying to avoid.

Keep it clean:

• Character A has its own reference or its own trained identity
• Character B has its own reference or its own trained identity
• These two signals never enter the same node at the same time

If you are working from trained LoRAs instead of reference images, the same rule holds. Two character LoRAs at full strength on the whole canvas will fight. Scope each one to its region or its masked pass. Our two-character LoRA tutorial goes deep on balancing LoRA weights when both are active.

Compose First, Identity Second

The mental model that fixed this for me was to stop treating the scene and the identities as one problem. They are two problems, and solving them in the wrong order is why results feel random.

Composition is a layout problem. Where are the bodies, what is the pose, what is the framing, what are they wearing. Identity is a face problem. Whose face goes on which head.

When you solve composition first and freeze it, the identity pass has a stable target to attach to. When you try to solve both at once, every reroll changes the composition and the identity together, so you never learn what actually moved the needle.

This is also why platforms that specialize in consistent characters feel so much easier. Apatero locks the identity as a first-class object, so you compose scenes and the same face follows into each one without a manual mask pass. If you would rather not wire the node graph yourself, that is the shortcut.

Handling the Hard Cases

A few situations still need extra care even with a solid workflow.

Characters of similar appearance. Two people with the same hair color and build are the hardest case because there is less for the model to hold onto. Exaggerate one distinguishing feature per character in your references. A different hairstyle or a strong wardrobe contrast gives the identity pass an anchor.

Interacting hands and faces. When one character touches the other's face, masks overlap and the passes can fight. Shrink your masks to the core of each face and let the harmonizing pass blend the edges. Do not mask the whole head when only the face needs locking.

Group scenes beyond two. The same principles scale, but the cost grows. Every added character is another region or another masked pass. Three is manageable. Past four, consider generating pairs separately and compositing, which our multi-image editing guide covers.

A Repeatable Checklist

When a scene is not holding, I run through this list in order:

1. Is each character in its own region or mask, with no overlap in conditioning
2. Does each character have exactly one reference, never a blend
3. Did I lock the composition before touching the faces
4. Is my face-pass denoise moderate, not maxed out
5. Are the two characters distinct enough in the reference to begin with

Nine times out of ten the problem is on this list, usually a shared reference or a face pass that ran too hot.

Frequently Asked Questions

Why do my two AI characters always end up looking like the same person? Because a single prompt and a single identity signal are being applied to the whole canvas. Split each character into its own region or masked pass so their conditioning never mixes.

Can I keep two characters consistent across a whole series, not just one image? Yes. Lock each identity as a reusable reference or trained character, then apply the same locked identities to every new scene. The composition changes per image while the two faces stay fixed.

Do I need ComfyUI for this, or is there an easier way? ComfyUI gives you full control with regional conditioning and masked passes. If you would rather skip the node graph, a platform like Apatero locks identities for you so two characters follow into new scenes automatically.

What causes face-bleed when two characters get close together? Overlapping masks or regions. When the spatial lanes touch, their identity signals can mix. Tighten each mask to the core of the face and let a light final pass blend the seams.

Is regional prompting or the masked pass better? Regional prompting is faster for characters in separate areas. The masked two-pass approach is more reliable when characters overlap or make contact. Many creators use both, regions for layout and masks for the final identity lock.

Wrapping Up

Two consistent characters in one scene is not a prompt problem, it is a structure problem. Give each identity its own lane, feed each one a single clean reference, lock the composition before the faces, and repaint each face in its own masked pass. Do that and the bleed disappears.

Start with one pair you care about, get the workflow holding, then reuse those locked identities across an entire series. Once the two faces stay put, building multi-character content stops feeling like gambling and starts feeling like a pipeline.