Homewreker — Yuuma Character LoRA

• Repository: toonsquare/Breaking_the_Perfect_Couple
• Trigger Word: gb_yuuma
• Base Model: astranime_V6 (Stable Diffusion 1.5)
• License: Apache-2.0

---

Overview

This repository contains a character-specific LoRA model trained for Yuuma, a character appearing in the webtoon “Breaking the Perfect Couple”.

The goal of this model is to preserve Yuuma’s facial features, hairstyle, and overall character identity while enabling stable generation across a wide range of compositions, from close-up portraits to full-body shots.

---

Trigger Word

• Required trigger: gb_yuuma

Including gb_yuuma in your prompt helps strongly anchor the generation to Yuuma’s character traits.

---

Dataset Construction & Training Pipeline

This model was not trained on raw collected images alone. Instead, a character-consistency-first data pipeline was designed and executed.

1. Original Image Acquisition (from IP Team)

High-quality reference images were provided directly by the IP team, covering the following compositions:

• Close-up
• Bust shot
• Knee shot
• Full shot

Each composition includes:

• Front view
• Left-side view
• Right-side view

This structure ensures that Yuuma’s facial features, hair, and silhouette remain consistent across different viewing angles.

2. Data Augmentation Using Flux

Using the original images as input, a Flux-based generation pipeline was used to create a large augmented dataset.

• Total generated images: ~1,000
• Generation time: ~12 hours

3. Manual Curation (Critical Step)

All generated images were manually reviewed and curated. The following were removed:

• Duplicate or near-duplicate images
• Incorrect generations (identity drift, malformed anatomy, etc.)
• Low-quality outputs (noise, broken composition, poor details)

After curation, approximately 500–600 images were selected for training.

4. LoRA Training

The curated dataset was used to train a LoRA model on top of:

• Base model: astranime_V6
• Architecture: Stable Diffusion 1.5

The training focused on maintaining strong character identity while allowing reasonable variation in pose, framing, and context.

---

LoRA Training Configuration (Reference)

⚠️ The following configuration reflects the settings used during training as closely as possible. Some values may not be exact and are provided for reference only.

Click to expand training configuration

{
  "LoRA_type": "LyCORIS/LoKr",
  "LyCORIS_preset": "full",
  "adaptive_noise_scale": 0.005,
  "bucket_no_upscale": true,
  "cache_latents": true,
  "cache_latents_to_disk": true,
  "caption_dropout_rate": 0.05,
  "clip_skip": 1,
  "epoch": 20,
  "learning_rate": 0.0001,
  "unet_lr": 0.0001,
  "text_encoder_lr": 0.0001,
  "optimizer": "AdamW",
  "mixed_precision": "bf16",
  "network_dim": 100000,
  "network_alpha": 1,
  "max_resolution": "768,768",
  "min_snr_gamma": 5,
  "seed": 1234,
  "save_precision": "fp16"
}

---

Usage Guide

Basic Prompt

gb_yuuma, [description]

Tips

• Positioning: Place gb_yuuma early in the prompt for stronger character locking.
• Composition: Explicitly specifying the composition (e.g., close-up, bust shot, full body) improves control.
• Style: Extreme stylistic deviations may weaken character consistency.

---

Intended Use & Limitations

• This model is designed for character-consistent image generation of Yuuma.
• It performs best when used with anime-style SD 1.5 checkpoints similar to the base model.
• Forcing radically different art styles or photorealistic prompts may reduce identity stability.

---

License

This model is released under the Apache License 2.0.

You are free to:

• Use the model commercially
• Modify and redistribute it

Please ensure compliance with the Apache-2.0 license terms and any applicable IP usage policies.

---

Credits

• Character: Yuuma (Breaking the Perfect Couple)
• Training & Curation: Toonsquare
• Base Model: astranime_V6