Multi-Character Scenes: Two Locked Identities, Zero Bleed

The first time I tried to put two locked AI characters in the same frame, I ended up with one character whose face was about 60 percent her, 40 percent the other character. Identity soup. The hair color landed on the wrong head. The eye shape blended. The pose was right but the casting was wrong. I spent the rest of that night reading every multi-character workflow thread on Reddit and most of them had the same problem.

Two characters in one frame is the moment most AI workflows quietly collapse. Single-character work is solved in 2026. Multi character AI scene consistency is still where the field has open ground. The fix is a three-layer stack of regional prompts, dual IPAdapter, and dual LoRA tied to canvas regions. Once you have it dialed in, you can ship two-character scenes that hold both identities at production quality.

Why a Single Prompt Always Bleeds Two Characters

The first thing to understand is why single-prompt approaches fail. When you write "a young woman with red hair and a young man with dark hair sitting at a cafe table," the diffusion model does not parse the sentence semantically the way you do. It builds a latent representation that includes both characters as features in the same image space.

In that shared latent, the red hair has a probability of landing on either character. The dark hair has a probability of landing on either character. The model picks based on training-data distributions, which favor certain compositions over others. You end up with one character whose hair color is a blend of both intended descriptions. Their eye shape is a blend. Their proportions are a blend. The result is the identity soup I started this article with.

This is not a prompting problem. It is a latent-space architecture problem. No amount of better prompt phrasing fully fixes it because the underlying model treats the canvas as a single subject space rather than as multiple subject spaces.

The 2025 versions of Flux and SDXL got slightly better at separating subjects in the same frame because of training-data improvements. But "slightly better" is not the same as "production reliable." Real multi-character work still needs structural separation in the workflow, not just better text.

This is why most AI comic creators and most AI brand campaigns with multiple characters either use a single-character-per-frame approach (cut to character A, then cut to character B, never both at once) or accept the bleed. The third path, the one this article covers, is to build the structural separation yourself.

Regional Prompting Explained for Practitioners

Regional prompting is the foundational technique. You divide the canvas into named regions and assign a separate prompt to each region. The diffusion model treats each region as its own subject space during the denoising process.

In ComfyUI the regional prompt nodes have been around for a few years. The Attention Couple node is the workhorse. You define rectangles or masks on the canvas, attach prompt text to each rectangle, and the model conditions on the right prompt per region during generation.

For two-character scenes the canvas typically splits into three regions. Left region for character A. Right region for character B. Shared region for the background, lighting, and shared elements. You write the prompt for each region separately. The model then renders character A in the left, character B in the right, and the shared elements stitch them together.

The prompt structure looks like this. Left region: "a young woman with red hair, freckles, blue sweater." Right region: "a young man with dark hair, glasses, brown jacket." Shared region: "cafe interior, warm lighting, wooden table, afternoon sun."

This alone reduces bleed substantially. From maybe 60 percent identity match to around 80 percent. The remaining 20 percent gap is what dual IPAdapter and dual LoRA fix.

Splitting the Canvas: Left Region, Right Region, Shared Region

The exact way you split the canvas matters more than people realize. Most tutorials show a clean 50/50 vertical split. That works for symmetrical two-character compositions where both characters take up equal space. It does not work for asymmetric compositions where one character is foreground and the other is background, or where the characters are at different scales.

The split should match your composition. For a cafe-table scene with both characters seated and roughly equal size, a 50/50 vertical split works. For a scene where character A is foreground and character B is in the background, the foreground character gets the bigger region. For a scene with one character standing and one character seated, the standing character's region is taller.

The shared region is where you place the background, the lighting, the table or object both characters interact with, and any environmental detail. It is the smallest of the three regions usually, sitting between the two character regions.

There is a small but important technique here. You can let the character regions and the shared region overlap slightly at the edges. The overlap zone gets blended naturally and avoids hard seams where the two characters meet. For the cafe scene, the table itself sits in both character regions and the shared region. The overlap lets it render coherently.

In Apatero AI, the multi-persona workflow handles the regional split with a visual interface. You drag rectangles on the canvas preview. Drop a persona into each rectangle. The shared region inherits the background prompt. Much faster than building the node graph in ComfyUI from scratch.

Dual IPAdapter: One Per Region

Regional prompts get you 80 percent of the way. Dual IPAdapter gets you the rest of the way. The idea is to run one IPAdapter FaceID per region with a separate reference image per character.

The setup. Character A reference image plugs into IPAdapter FaceID instance 1, attached to the left region. Character B reference image plugs into IPAdapter FaceID instance 2, attached to the right region. Both run during the same generation pass. Each IPAdapter strengthens identity in its own region without affecting the other.

This works because the IPAdapter conditioning is region-aware. The FaceID embedding for character A only applies in the left region. The FaceID embedding for character B only applies in the right region. The model cannot cross-contaminate because the conditioning channels are separate.

Weight tuning is the gotcha. Each IPAdapter weight should sit between 0.7 and 0.85. Too high (above 0.9) and the IPAdapter overpowers the regional prompt structure. Too low (below 0.6) and you get drift back toward the bleed problem. I tested 25 different weight combinations on the same two-character reference set and 0.78 to 0.82 was the sweet spot for both adapters.

Hot take. Most multi-character tutorials online tell you to use a single IPAdapter with a composite reference image (a side-by-side composite of both characters). That approach fails for the same reason single prompts fail. The latent representation blends. Two separate IPAdapters on two separate regions is the structurally correct solution.

Two LoRAs With Activation Words Tied to Regions

The third layer is dual character LoRAs. If you have trained a LoRA for each character (most reliable approach for ongoing work), you load both LoRAs and tie their activation words to the region prompts.

The setup. LoRA for character A loaded at weight 0.8. Activation word for that LoRA included in the left region prompt. LoRA for character B loaded at weight 0.8. Activation word included in the right region prompt. The shared region prompt has no activation words.

This is where the structural separation gets really clean. LoRA activation words act like keys. The activation word only "unlocks" the LoRA when it appears in the prompt being processed at that point. Because the region prompts are separated, the LoRA activations are region-scoped. Character A's LoRA only fires in the left region. Character B's LoRA only fires in the right region.

When you stack all three layers (regional prompt, dual IPAdapter, dual LoRA), you get identity match rates that approach single-character workflows. In my testing the two-character match rate across 30 generations was around 88 to 92 percent for both characters simultaneously. That is production quality.

The catch is workflow complexity. Building this stack in ComfyUI takes about 90 minutes the first time. Once saved as a workflow, generation is straightforward. But the initial setup is heavy.

Pose ControlNet to Stabilize Both Bodies

The fourth layer is pose ControlNet for body stability. Without pose control, the regional prompt approach can produce two characters whose bodies overlap awkwardly or stand in poses that contradict the scene description.

The fix is OpenPose ControlNet using a reference pose image that has both bodies pre-posed. You either generate the pose reference manually (using a 3D pose tool like Magic Poser or even drawing stick figures) or you take a real photo of two people in the desired pose and convert it.

The pose reference feeds into a ControlNet OpenPose node. The OpenPose conditioning applies to the whole canvas (it does not need regional gating because it is structural rather than identity). Both characters now hit the right poses without random arm placement or floating limbs.

Weight tuning for pose ControlNet sits around 0.6 to 0.8. Lower than character IPAdapters because pose is a softer constraint that you want to allow flexibility around. Hard locking pose at 1.0 weight tends to produce stiff-looking compositions.

For comic-style work the pose ControlNet is essential because the audience expects characters in deliberate poses across panels. For lifestyle or photo-style work it can be optional if you trust the model to produce natural poses on its own.

Shared Lighting Without Breaking Either Identity

Lighting is the one thing you actually want to bleed across regions. Both characters in the same scene should be lit by the same source. If character A is lit by golden hour sun and character B is lit by overcast soft light, the composition reads as a clumsy collage.

The fix is to keep lighting language in the shared region prompt only. The character regions describe identity and outfit but not lighting. The shared region carries the lighting language for both. Because the regional prompting allows partial blending at boundaries, the lighting language affects both characters without overwriting their identity.

The shared lighting prompt structure I use. "Warm afternoon light, soft directional from window left, gentle fill on faces, subtle shadow on table." That language applies to the whole scene. Both faces receive consistent lighting because the prompt anchored the lighting at the scene level.

This is the one place where I let region boundaries get blurry deliberately. Identity is hard-walled at the region level. Lighting is soft-walled across regions. The result is two consistently identified characters who clearly inhabit the same physical space.

Common Failure Patterns and Fixes

A few specific failure modes I have seen and how to fix them.

Failure mode one. Character A's outfit ends up on character B. Caused by outfit description in the prompt being parsed as a global feature rather than regional. Fix is to make sure outfit text is in the regional prompt for the right character, never in the shared region.

Failure mode two. Both characters render at the same height when one should be sitting and the other standing. Caused by no pose conditioning. Fix is to add pose ControlNet with a reference pose that shows the height difference.

Failure mode three. The faces are correct but the eye gaze is wrong (both characters looking the same direction when they should be looking at each other). Caused by no explicit gaze direction in either region prompt. Fix is to add gaze direction to each region prompt ("looking right toward the other character" in the left region, "looking left toward the other character" in the right region).

Failure mode four. The whole image looks like a side-by-side collage rather than a unified scene. Caused by too-strong region weights or no shared region. Fix is to ensure the shared region prompt has substantial scene content (background, lighting, environment) so the model has something to bind the two character regions together.

Failure mode five. The shared region renders empty or muddled. Caused by shared region prompt being too sparse. Fix is to write a full background prompt for the shared region with the same level of detail as the character region prompts.

Encoding Two-Person Scenes as an Apatero Workflow

The whole stack above runs cleanly in Apatero AI with much less manual setup than ComfyUI. The multi-persona feature handles the regional plumbing, the dual IPAdapter logic, and the shared region setup as a visual workflow rather than a node graph.

The Apatero workflow. Open the multi-persona tab. Drop persona A into the left region. Drop persona B into the right region. Write the shared region prompt for background and lighting. Click generate. The platform handles the dual IPAdapter calls, the regional prompt splitting, and the optional pose ControlNet as a single backend operation.

Setup time in Apatero is around 5 minutes for a two-character scene versus 90 minutes to build the equivalent ComfyUI graph the first time. After that you save the workflow and reuse it.

For my own work I run the multi-character setup in Apatero for everything except niche edge cases where I need more control than the UI exposes. For those edge cases I drop into ComfyUI. The 90/10 split between hosted and custom is the same pattern I describe in the comparison guide on this site.

Working Beyond Two Characters

The same approach extends to three or four characters in a single frame but with diminishing reliability. Three-character scenes hit around 80 to 85 percent identity match in my testing. Four-character scenes drop to 65 to 75 percent. Past four characters the regional canvas gets too crowded and the IPAdapter conditioning starts overlapping in ways the model cannot cleanly separate.

For larger scenes the better approach is multi-pass compositing. Generate each character separately on a transparent background using single-character workflows. Composite them into the scene in a final pass using a tool like Krita or Photoshop. This is more manual but the identity quality is much higher.

Most production use cases for AI influencers and AI brand work involve two characters or fewer (a couple, a duo, a host plus guest). The two-character workflow described in this article covers the bulk of what most people actually need.

FAQ

Can I do this with only IPAdapter, no LoRAs?

Yes but identity match drops from 88 to around 75 percent. For one-off use that is fine. For ongoing series work I recommend training character LoRAs for stability.

Do I need ComfyUI or can I use Automatic1111?

A1111 has regional prompting and dual IPAdapter support via extensions but the workflow is less reliable than ComfyUI. For multi-character work specifically, ComfyUI or Apatero are the better options.

What about multi-character video?

Multi-character video is still hard in 2026. The same identity-bleed problem exists in video generation, and the regional-prompting solutions are not as mature in video models. Workarounds involve generating each character separately and compositing.

Will Flux Kontext do this natively?

Flux Kontext has some multi-reference support that can hold two characters in a scene but the consistency is closer to 70 to 75 percent. The full structural stack described in this article still wins on quality.

Can I use one trained LoRA and one IPAdapter?

Yes, asymmetric setups work. Stronger character (the one you have a LoRA for) uses LoRA plus IPAdapter. Other character uses IPAdapter only. Identity match is slightly different between the two but it can be a useful workflow when you only have a LoRA for one of the characters.

How do I handle the same character interacting with itself?

Two-of-the-same-character scenes work the same way. Both regions use the same persona reference and LoRA. The result is two instances of the same character, useful for "before and after" or "two moods of the same person" content.

What about pose conflicts when both characters are touching?

When characters physically touch (a hug, a handshake), the pose ControlNet needs to be especially clean. Generate the pose reference carefully. Often easier to use a real photo of two people in that exact pose as the OpenPose source.

Is this overkill for casual content?

Yes, possibly. For casual one-off content with two characters, single-prompt with --cref or basic regional prompting is fine. The full stack matters when you are shipping production volume or need very high identity quality.

The Takeaway

Multi character AI scene consistency is solvable. It requires structural separation, not just better prompting. Regional prompts split the canvas. Dual IPAdapter locks identity per region. Dual LoRAs tie activation words to regions. Pose ControlNet stabilizes bodies. Shared lighting language stitches everything together.

The first time you build this stack it takes a long evening. After that it is a saved workflow you reuse. The output is two locked identities in one frame, which opens up an entire category of content (comics, couple shots, host-guest scenes, multi-character brand campaigns) that single-character workflows cannot serve.

For the hosted approach that hides most of the plumbing, Apatero AI's multi-persona feature makes the workflow accessible without the node graph. Related guides worth reading: the LoRA plus IPAdapter consistency recipe that underpins the single-character side, the IPAdapter weight tuning guide for dialing in the per-region adapters, and the comic page workflow which uses this technique heavily. External references worth bookmarking: the ComfyUI Attention Couple documentation and the cubiq IPAdapter implementation guide.