# Blue Archive Style — MOSS-Image LoRA

A Blue Archive (BA) illustration style LoRA for MOSS-Image.

It reproduces the look of the game's official art: clean cel shading,

soft bloom, pastel gradients and that unmistakable glossy anime finish.

Trained on the SFT checkpoint rather than Base, so the style

strength behaves predictably in the model you actually generate with.

> Fan-made style study. Not affiliated with NEXON Games or Yostar.

## Trigger word

> *mularuru**

## What it does

Pushes MOSS-Image toward the Blue Archive look:

- Clean, thin lineart with soft anti-aliased edges

- Two-tone cel shading with gentle gradient shadows

- Bright pastel palette — cyan, pink, cream, lavender

- Glossy specular highlights on hair and skin

- Soft bloom / light haze around bright areas

- Typical BA framing: cute anime girls, school settings, halos

## Good at

- Single-character portraits and half-body shots

- School uniforms, casual outfits, classroom and campus scenes

- Bright, airy lighting

- Halo characters and pastel color schemes

- Clean or simple backgrounds

## Not great at

- Dark, moody or horror scenes — the style pulls bright

- Heavy realism or painterly prompts — it fights them

- Complex multi-character layouts (MOSS handles layout, but this LoRA

adds no composition control)

- Non-anime subjects such as vehicles or architecture close-ups

## Recommended settings

| Setting | Value |

|---|---|

| Base model | MOSS-Image SFT (bf16) |

| LoRA strength | 0.8 (try 0.6 – 1.0) |

| Sampler | Euler |

| Scheduler | sgm_uniform |

| Steps | 50 |

| CFG | 3.0 |

| Shift | 5.0 |

| Resolution | 1024×1024 or higher |

Negative prompt (only active above CFG 1):

worst quality, bad anatomy, watermark, jpeg artifact

> With the 8-step DMD speed LoRA at 0.7 and CFG 1 it still works, but the

> style comes out noticeably softer. Use CFG 3 when style fidelity matters.

## Example prompt

mularuru, 1girl, halo, blue eyes, white sailor uniform,

classroom, soft lighting, looking at viewer

## Training details

| Item | Value |

|---|---|

| Base | MOSS-Image SFT bf16 |

| Network | LoRA (attn + MLP), DiT only |

| Rank / Alpha | 4 / 4 |

| Learning rate | 3e-4, constant + 20-step warmup |

| Epochs / Steps | 25 / 1950 |

| Batch | 1 |

| Resolution | 768 (0.59 MP) |

| Precision | bf16 |

| Dataset | 78 images |

| Text encoder | frozen |

Notes: the text encoder is not trained, so no CLIP/TE weights are

included. Trained at 0.59 MP — it works above that, but 1024+ gives the

best line quality.

## Not recommended

- Strength above 1.2 — colors burn out and edges bloom

- Expecting a specific character: this captures the game's art style,

not any individual character's face or outfit