SDXL Dual Character LoRA Workflow | Masked Conditioning & Automatic Prompt Splitting

LoRA Workflow | Masked Conditioning & Automatic Prompt Splitting

Generate two completely different LoRA characters in the same image with clean character separation and minimal feature bleeding.

This SDXL workflow uses masked conditioning to assign each character to its own side of the image, allowing two independent LoRAs to coexist without competing for attention or contaminating each other's features.

Perfect for character interactions, couples, rival characters, storytelling scenes, and any composition that requires reliable multi-character generation.

Download the workflow HERE!

Installation Requirements

This workflow requires the following model types:

• Base Model Recomended: Wai Illustrious V12:ComfyUI/models/checkpoints

• VAE [OPTIONAL] sdxl-vae-fp16-fix: ComfyUI/models/vae

• Upscale Model [OPTIONAL] remacri: ComfyUI/models/upscale_models

• Upscale model used in this workflow: 4x-UltraSharp HERE or HERE!

• Mandatory LoRA Model for HiRES: DMD2 Speed LoRA [SDXL, Pony, Illustrious] HERE!

This workflow requires custom node packages to work, they can be easily installed via the built-in Custom Node Manager. However, the Prompt Splitter custom node responsible for dividing the total prompt into parts dedicated to each character's mask areas cannot be downloaded to the manager, it must be cloned in the custom nodes folder of your Comfyui.

Prompt Splitter: HERE!

✨ Key Features

• Two independent LoRA characters in a single image

• Automatic left/right character separation

• Mask-based conditioning for improved character consistency

• Built specifically for SDXL

• Single prompt input for a streamlined workflow

• Automatic prompt splitting with custom Prompt Splitter node

• Dynamic masks that adapt automatically to resolution and aspect ratio changes

• No manual mask editing required

⚙️ How the Workflow Works

The workflow is designed around a simple 3-line prompt structure.

Line 1 — Global Scene Prompt

This line contains information shared by the entire image, such as:

• Environment

• Lighting

• Camera settings

• Art style

• Composition details

• Shared scene elements

Example:

cinematic lighting, city street at night, detailed background, masterpiece

Line 2 — Character A

This line contains all information related to the character positioned on the left side of the image.

Example:

<lora:characterA>, red hair, school uniform, smiling

Line 3 — Character B

This line contains all information related to the character positioned on the right side of the image.

Example:

<lora:characterB>, black hair, military outfit, serious expression

🔀 Prompt Splitter System

A custom Prompt Splitter node automatically separates the three input lines into independent prompts:

• Global Prompt

• Character A Prompt

• Character B Prompt

The workflow then combines the global scene information with each character-specific prompt before sending them to their respective conditioning branches.

This allows both characters to share the same environment while maintaining completely independent identities.

🎭 Masked Character Separation

The image is automatically divided into two regions:

• Left side → Character A

• Right side → Character B

Each region receives its own conditioning and LoRA influence through dedicated masks.

Unlike traditional multi-LoRA workflows, where all LoRAs affect the entire image, this setup restricts each character's conditioning to its assigned area, greatly reducing:

• Face mixing

• Outfit contamination

• Style bleeding

• Identity conflicts

📐 Adaptive Resolution Support

The masks are generated automatically and scale with the selected resolution and aspect ratio.

You can freely change image dimensions without needing to recreate or adjust masks manually.

Whether you're generating:

• Portraits

• Landscape scenes

• Wallpapers

• Vertical illustrations

the character regions remain properly aligned.

🎯 Ideal Use Cases

• Character interactions

• Couples

• Anime scenes

• Storytelling illustrations

• Rival characters

• Before/after comparisons

• RPG party members

• Original character showcases

💡 Why Use This Workflow?

Running multiple character LoRAs together often produces inconsistent results because every LoRA influences the entire image.

This workflow solves that problem by combining automatic prompt separation with masked conditioning, allowing each character to occupy its own controlled region while still sharing the same scene.

The result is cleaner generations, stronger character identity preservation, and significantly better multi-character compositions.