PixAI SDXL Prompting Guide – Otome Special: Mastering BL / Yaoi Art in PixAI

This guide specifically addresses gender errors and feature bleeding in SDXL models when generating BL/Danmei art. By using Regional Prompting (right:/left:) to define dual-character details, it provides a "Golden Formula for BL," interaction tag cheat sheets, and techniques for physique differences—serving as a practical manual for mastering precise male-to-male interactions.

Editor’s note (PixAI)
This guide was written by one of our community’s standout creators, 阿童 (ATone), and is republished here with credit to the original author.

⚠️ Important Note: Multi-Character Limitations in SDXL

BL / Yaoi art often involves multiple characters. It’s important to note that SDXL models are not particularly strong at handling multi-character compositions.

When generating two or more characters, SDXL may suffer from “feature blending,” where visual traits (faces, hair, clothing) mix between characters. This becomes even more noticeable when generating three or more people.

If your goal is complex multi-character scenes, DiT-based models are generally a better choice, as they handle multi-character composition more reliably.

👉 DiT-Based Models:

  • Natural-language prompts supported (tags optional)
  • Higher-quality multi-character generation (spacing, interaction, balance)
  • Improved character interactions (eye-lines, gestures, overlaps)
  • Enhanced version of Tsubaki with improved aesthetics and detail
  • More refined rendering of faces, clothing, and backgrounds
  • Stronger overall artistic quality and color harmony
  • Official high-speed version of Tsubaki
  • High LoRA compatibility, especially with LoRAs trained on original Tsubaki
  • Authentic Korean Art Style
  • Visually Stunning Character Design
  • Balanced Male & Female Output
  • Versatile Style Adaptation

For a deeper explanation, please refer to ATone’s detailed guide on multi-character generation.


Common Issues

Problem: You want to draw BL, but the result ends up as BG (Boy & Girl) or GL (Girl & Girl).

Reason:

1. Models are biased toward generating females by default.

2. The weight of the 2boys tag isn’t strong enough.

3. Interaction tags (like kiss) are often paired with heterosexual couples in training datasets.

The Solution

1. Use Regional Prompting for Positive Prompts (utilizing right: / left: blocks):

2boys, yaoi, Character A at right, Character B at left, [interaction], [setting]

right: Character A, detailed features and actions
left: Character B, detailed features and actions

❌ Use Negative Prompts (to completely exclude females):

1girl, female, hetero, yuri, breasts, 
feminine, woman, girl

⭐ The Golden Formula for BL Illustrations

This is the most effective structure verified by testing. It allows the model to clearly distinguish between two characters:

2boys, yaoi, [Name 1] at right, [Name 2] at left, [interaction], [scene description], [art style], [quality tags], [LoRA triggers]

right: [Name 1], detailed features and actions of the character on the right
left: [Name 2], detailed features and actions of the character on the left

Formula Breakdown

SectionFunction
First Line (Global)Overall settings: Number of people, relationship, positioning, interaction, and scene.
right:All details specific to the character on the right.
left:All details specific to the character on the left.

Key Points:

  • Use at right / at left to declare character positions in advance.
  • Use right: / left: blocks to describe each character in detail.
  • This method prevents the model from “bleeding” (mixing) the features of the two characters together.

Interaction Tag Reference

Editor’s Note (PixAI):
The sample images below demonstrate how interaction tags can work, but in real generations SDXL may still struggle with multi-character consistency. Feature blending (trait “bleeding” between characters) is common, so results are not guaranteed in a single try.

For example, when adding Dan Heng’s red eyeliner (“red eyeliner”) tag, you may also see it appear on Caelus. You can click each Sample Output link to view the exact prompts/tags used.

Outfit mixing can also happen across characters. If you need higher fidelity to a specific character design, you can try adding a character LoRA. However, many character LoRAs are trained with a strong “1boy” bias, so they may not behave as expected in two-character compositions.

If your goal is consistently accurate multi-character results, we generally recommend using DiT-based models, which handle multi-character composition more reliably.

InteractionTagExample Image(PixAI)
Kissingkiss, kissing, french kiss🖼️ Sample Output
Hugginghug, hugging🖼️ Sample Output
Holding Handsholding hands🖼️ Sample Output
Eye Contacteye contact, looking at another🖼️ Sample Output
Hug from Behindhug from behind🖼️ Sample Output
Wall Slamkabedon🖼️ Sample Output
Princess Carryprincess carry, carrying🖼️ Sample Output
Forehead-to-Foreheadforehead-to-forehead🖼️ Sample Output
Head on Shoulderhead on another's shoulder🖼️ Sample Output
Lap Pillowlap pillow🖼️ Sample Output

Height & Build Differences

✅ Height Difference:

height difference, tall male

Build Difference:

muscular male, large male

💡 Note: taller maleshorter male、and slim male are not valid Danbooru tags. Use tall male paired with specific descriptions to express the difference.


Example Prompts

Example 1: Soft BL

Positive:

2boys, yaoi, Hikaru at right, Sora at left, kiss, eye contact, blush, classroom, school uniform, soft lighting

right: Hikaru, short black hair, blue eyes, gentle smile, slightly taller, hand on cheek
left: Sora, messy brown hair, green eyes, surprised expression, blushing

Negative:

1girl, female, hetero, yuri, breasts, feminine

Example 2: Older Seme × Younger Uke

Positive:

2boys, yaoi, Ren at right, Yuki at left, hug from behind, height difference, bedroom, night

right: Ren, short hair, black hair, narrowed eyes, tall male, suit, smirk
left: Yuki, messy hair, brown hair, slim, casual clothes, blush

Negative:

1girl, female, hetero, yuri, breasts, feminine

Example 3: Equal Pairing

Positive:

2boys, yaoi, Kuro at right, Shiro at left, holding hands, walking, looking at another, city, sunset

right: Kuro, black hair, red eyes, black jacket, smile
left: Shiro, white hair, blue eyes, white coat, smile

Negative:

1girl, female, hetero, yuri, breasts, feminine

Advanced Techniques

Why does the right: / left: syntax work?

This block syntax leverages the “conditional description” structure the model learned during training. When the model sees a block starting with right:, it applies the subsequent descriptions only to the character on the right side of the canvas.

Other Valid Block Formats

Besides right: / left:, the following also work:

Using Numbers:

1: [Character A Features]
2: [Character B Features]

Using Names:

Hikaru: [Character A Features]
Sora: [Character B Features]

Keys to Avoiding Feature Mixing

  1. Declare positions first using at right / at left.
  2. Use detailed blocks (right: / left:) for descriptions.
  3. Create contrast: Ensure hair colors (e.g., Black vs. Blonde) and eye colors (e.g., Blue vs. Red) are distinct.

R18 Related Tags

(List of commonly used tags; please use in accordance with platform guidelines.)

Nudity: nude, naked
Topless: topless male, bare chest
Suggestive: suggestive
Tongue Kiss: tongue kiss, tongue out

Common Troubleshooting

IssueLikely CauseSolution
Female appearsNegative prompt isn’t strong enough.Strengthen the Negative Prompt.
Characters look identicalDescription isn’t specific enough.Increase the visual difference in hair/eyes.
3+ people appearDidn’t specify 2boys.Explicitly state the quantity: 2boys.
Interaction looks unnaturalComposition is too complex.Start with simpler poses/compositions.
Character becomes feminizedTags like uke sometimes trigger female traits.Use specific physical tags (e.g., short hair, blush) instead of abstract tropes.

Related Topics

Index