Multi-Character LoRA Generation: Create AI Art with 2+ Characters
Generating multiple characters in a single AI artwork is one of the most challenging LoRA techniques to master. This guide walks you through a systematic four-step method for multi-character LoRA generation, plus tips for managing your LoRA collection.
⚡ TL;DR
If you only need one takeaway: use Tsubaki.2 for multi-character generation whenever you can. The DiT architecture handles 2+ character composition with dramatically less feature bleeding than SDXL — no regional prompting tricks required. When you must stay in SDXL (because the character LoRA you need only exists there), use the right: / left: regional prompting formula plus LoRA weights around 0.65.
This is Part 4 of the PixAI LoRA series. If you’re new to LoRAs entirely, start with Part 1 — What is LoRA? first.
Want to follow along with your own multi-character generation?
The examples in this guide use Tsubaki.2 — PixAI’s flagship model for multi-character work.
Try Tsubaki.2 Free →📚 The complete LoRA series
Why Multi-Character LoRA Generation Is Uniquely Hard
Single-character LoRA work is well-understood at this point. Pick a base model, find a LoRA on the marketplace, add the trigger words to your prompt, generate. Done.
Two characters in one image is a different problem entirely. You’re not just asking the model to render two people in a scene — you’re asking it to keep two completely separate identity sets straight, while two character LoRAs both pull the latent space toward their respective trained features. The result, more often than not, is one of three failures:
Feature bleeding
Hair colors swap. Eye colors mix. One character ends up wearing the other’s outfit. Signature accessories disappear entirely.
One LoRA dominates
Both characters end up looking like the same character. The “stronger” LoRA wins, the weaker one disappears as if it wasn’t loaded at all.
Wrong number of subjects
You asked for 2 characters, the model produces 3. Or just 1. The character count tag gets quietly overridden.
The good news: each of these has a fix, and the fix depends almost entirely on which base model you start with. Let’s start there.
Start With the Right Model: Why DiT Beats SDXL Here
Before any prompt technique matters, the base model you pick sets the ceiling for what’s achievable. For multi-character work, this matters more than anything else you’ll do downstream.
To make this concrete, we ran the same scene through three different approaches: two characters from Honkai: Star Rail — Evernight and March 7th — discussing a weekend trip on a sunny city street. School uniforms, casual setting, both characters’ canonical visual features should be intact. Here’s what came back.
What the comparison actually shows
Approach A (SDXL naive) produces two characters that are visually identical. Evernight’s red eyes, black gloves, question mark earring, white coattail — all gone. The two character LoRAs fought, March 7th’s training dominated, and the result is essentially “two March 7ths in a school uniform.” This is what happens by default when you load two character LoRAs and write a normal prompt.
Approach B (SDXL + regional prompting + reduced weights) is meaningfully better. With the right: / left: formula and LoRA weights dropped to 0.65, you can see one character is more distinctive than the other now. Outfits are more layered. But the hair colors still aren’t quite right, and there’s still residual bleeding. This is the realistic ceiling for SDXL.
Approach C (Tsubaki.2 DiT) looks like a different category of output entirely. Evernight’s red eyes are present. Her dark hair accessory is there. March 7th has her blue eyes. The two characters are clearly different people, interacting naturally. And we didn’t use right: / left: syntax — just natural-language description.
“DiT architecture solves multi-character composition at the model level. Most of the SDXL workarounds become unnecessary.”
When you still need to use SDXL
There’s a real reason you might stay on SDXL even after seeing the comparison above: character LoRA availability. The SDXL ecosystem has been around longer, and many character LoRAs — especially for niche game and anime characters — were trained on Illustrious, NoobAI, or other SDXL-derived bases. If the specific character you want is only available as an SDXL LoRA, that’s your stack.
The next two sections cover both routes — SDXL with regional prompting in Section 3, and the simpler DiT workflow in Section 5. Choose based on where your character LoRA actually lives.
📖 For more on PixAI’s model lineup and when to pick which one: Tsubaki.2 — Everything You Need to Know
The Regional Prompting + LoRA Formula (SDXL Route)
If you’re on SDXL, regional prompting is the most reliable technique we have. The core idea: divide the canvas into named regions, then describe each character only within its region. The model learns to apply each set of descriptions only to the corresponding canvas area, which dramatically reduces feature bleeding.
The base regional prompting syntax was originally documented for non-LoRA BL/yaoi work by PixAI community creator ATone — see the BL/Yaoi prompting guide for the full original formula. What follows is how to adapt it specifically for multi-character LoRA generation.
The structure
— Multi-Character LoRA Prompt Template —
# Global block — applies to whole image
2girls, [scene description], [character A] at right, [character B] at left,
<lora:character_A_lora:0.65>, <lora:character_B_lora:0.65>
# Right region — character A only
right: [character A trigger words], [A’s visual details], [A’s action]
# Left region — character B only
left: [character B trigger words], [B’s visual details], [B’s action]
Three changes from single-LoRA practice
Drop LoRA weights from 1.0 → 0.6–0.7
When two character LoRAs compete, both at full strength saturate the latent space and cause one to override the other. Reducing both gives each room to coexist. Start at 0.65 and adjust by ±0.05 if needed. For the deep dive on what weight does, see Part 2 — LoRA Weight Settings.
Place each LoRA’s trigger words inside its region block
Don’t put both characters’ triggers in the global prompt — that gives both LoRAs permission to apply everywhere. Anchor each trigger to one side. If you’re unsure about trigger word placement, see Part 3 — LoRA Trigger Words.
Declare positions in the global prompt first
Before the region blocks, tell the model who goes where: Acheron at right, Bronya at left. This primes the spatial composition before the regions get applied. Skipping it makes the regions less reliable.
Pro Tip — Negative Prompts Matter More Now
Add feature mixing, features bleeding between characters, mixed hair colors, swapped outfits to your negative prompt. These are explicit instructions to the model to avoid the most common bleed patterns.
Worked Example: Two Character LoRAs in One Scene
Let’s walk through the actual prompt that produced Approach B from Section 2 — Evernight and March 7th sharing a phone screen on a sunny street. Here’s the full prompt, broken down.
Why this works — line by line
Global block sets the scene, position, and LoRA loading
Number of subjects (2girls), shared activity, background, lighting — and crucially, the position declarations (Evernight at right, March 7th at left) that prime the spatial composition before the region blocks apply.
Both LoRAs at 0.65 — not 1.0
At 1.0 each, they compete for latent dominance. At 0.65, both have room to apply their training without overpowering. If after a few iterations one character still looks “weaker” than the other, nudge that one up to 0.7 while leaving the dominant one at 0.6.
Right region anchors Evernight’s identity
The trigger words (Evernight (Honkai: Star Rail)) appear only in this block. Specific visual anchors (red eyes, black gloves, question mark earring) reinforce the character’s distinctive features so the LoRA actually delivers them.
Left region keeps March 7th confined
Same logic, opposite side. Her trigger appears only in the left block. Her visual descriptors (blue eyes, slim body, energetic expression) reinforce her training without crossing into Evernight’s region.
The DiT Route: When Natural Language Replaces Regional Syntax
If you’re on Tsubaki.2, the workflow looks completely different — and significantly simpler. No right: / left: syntax. No weight balancing. Often no character LoRA needed at all (Tsubaki.2 recognizes many popular characters natively).
Here’s the prompt that produced Approach C from Section 2 — same scene, same characters, completely different writing approach.
Three writing differences from SDXL
Write descriptive sentences, not tag lists
Tsubaki.2 was trained on natural language. “A girl with red eyes wearing a black skirt” outperforms 1girl, red eyes, black skirt.
Use “On the right / On the left” prose
Spatial composition becomes part of the natural-language description. No right: / left: syntax — DiT understands prose positioning.
Often skip character LoRA entirely
Tsubaki.2 recognizes many popular game and anime characters from natural language. Try without a LoRA first; add one only if the result isn’t accurate enough.
Pro Tip — Turn on Prompt Helper
PixAI’s Prompt Helper restructures your prompt for Tsubaki.2 automatically. After generation, check the Generation Task panel to see how it transformed your input — it’s the fastest way to learn what natural-language structure DiT actually rewards.
Common Pitfalls When Stacking Character LoRAs
Even with the right model and structure, multi-character work surfaces problems you don’t see in single-character generation. Here are the four most common, with fixes.
PITFALL 01
“My second character disappeared”
Symptom: The model produces two of the dominant character, ignoring the second LoRA entirely.
Fix: Drop both LoRA weights to 0.6, then bump the weaker one to 0.7. Add stronger visual anchors (signature outfit, distinctive accessory) to that character’s region block.
PITFALL 02
“Three characters appeared, I asked for two”
Symptom: Character count tag gets overridden. Sometimes called “extra body” syndrome.
Fix: Reinforce 2girls in the prompt and add 3girls, multiple girls, crowd to negatives. If still happening on SDXL, the LoRA itself may be over-trained on multi-character data — try a different LoRA for that character.
PITFALL 03
“They both end up female / both look young”
Symptom: SDXL’s training bias toward female/young characters overrides one of your prompts.
Fix: Use strong reinforcement (2boys or male focus, tall male) and aggressive negatives (1girl, feminine, female). For mature characters, add mature, adult and negate child, young.
PITFALL 04
“Interactions look unnatural / wooden”
Symptom: Characters separated cleanly, but they’re not actually interacting — just standing near each other.
Fix: Add explicit interaction tags to the global block: looking at each other, eye contact, holding hands, sharing a phone. For DiT, describe interaction in natural prose: “they walk arm in arm” or “she shows the phone to her friend.”
📖 For stacking LoRAs at a more advanced level — including non-character LoRAs alongside character ones — see LoRA Stacking Guide (PixAI Mastery, Part 3).
Going Beyond Two Characters: 3 or More
Two characters is the upper limit of what current models reliably handle. At three or more, even Tsubaki.2 starts producing noticeable bleeding, and SDXL regional prompting breaks down quickly.
If you genuinely need a three-character scene, here are the honest options, ranked from best to most laborious:
Try Tsubaki.2 first, accept some bleeding
For 3-character group shots, Tsubaki.2 still handles the composition better than anything else available. You may get one character with slightly merged features, but the overall result is usually publishable with light hand cleanup.
Generate characters separately, composite manually
Generate each character on a transparent or simple background, then composite into a shared scene using PixAI Edit Pro or external software. More work, dramatically better consistency.
Use Reference Pro for character-consistent additions
Generate the first two characters together, then use Reference Pro to add the third character to the existing scene with natural-language instructions while preserving the first two.
“There’s no shame in compositing. Some of the best multi-character work in the community uses AI for the heavy lifting and manual touch-ups for what AI can’t quite get right.”
In summary
The multi-character cheat sheet
✓ Tsubaki.2 (DiT) handles multi-character with the least friction
✓ On SDXL: use right:/left: regional blocks + 0.65 weights
✓ Place each character’s trigger words inside their region block only
✓ Add bleeding terms to negative prompts explicitly
✓ 3+ characters: accept some bleeding or composite manually
Frequently Asked Questions
Can I use more than two character LoRAs at once?
Technically yes — PixAI allows multiple LoRA slots. Practically, results degrade fast past two. Three character LoRAs in one image will almost always produce visible bleeding even on Tsubaki.2. For three or more, see Section 7 — manual composition is usually a better path than stacking three LoRAs at once.
Why do I need to lower LoRA weights for two characters but not for one?
A character LoRA at weight 1.0 saturates the latent space to enforce that character’s features everywhere. With two competing LoRAs both at 1.0, the saturation conflicts and the stronger one wins. Dropping both to ~0.65 gives each enough “influence” to apply locally without overpowering the other globally. For the full theory, see Part 2 — LoRA Weight Settings.
Does Tsubaki.2 work with SDXL-trained character LoRAs?
No. LoRAs are base-model-specific. An Illustrious-trained LoRA won’t work properly on Tsubaki.2, and vice versa. If your target character only has an SDXL LoRA available, your options are: (a) stay on SDXL with regional prompting, (b) check if Tsubaki.2 can generate the character from natural-language description alone (it often can for popular characters), or (c) train a DiT version of the LoRA — see DiT LoRA Training Guide.
What if my two characters are supposed to look similar (like sisters or clones)?
Counterintuitively, this is harder. The less visual contrast between characters, the easier it is for the model to merge them. If characters share a hair color or outfit, exaggerate other differences — different expressions, accessories, poses, or hair styles. Add explicit difference cues to your prompt: “older sister with longer hair, younger sister with twin tails.”
How many generations should I expect before getting a good multi-character result?
More than single-character work. Plan on 4–8 generations per scene with seed and minor prompt adjustments. The first try almost never works perfectly. Treat the early outputs as a debugging tool: which character is dominating? Are the regions actually being respected? Adjust weights or anchors based on what you see, not what you hoped.
Continue your LoRA journey
Ready to put multi-character generation to work?
Open Tsubaki.2 on PixAI, drop in your two-character prompt, and see the difference for yourself. New users get free daily credits — enough to test multi-character generations across multiple seeds before deciding what works.
Start Creating Free on PixAI →No credit card required · Daily free credits · Commercial use allowed
