# YOMI Style — MOSS-Image LoRA

A style LoRA for MOSS-Image, trained on the SFT checkpoint (the

checkpoint most people actually generate with), so there is no

train/inference distribution gap.

## Trigger word

> *yomi**

## Recommended prompt

yomi, bold painterly brushstrokes, vibrant high-contrast colors,

chunky color blocking, dynamic composition, iridescent light leaks,

glowing rim light, 1girl, white background, looking at viewer

The style follows a modern theatrical-anime look: strong linework

dissolved into colour blocks, high-contrast saturated palettes, and

specular highlights with iridescent light leaks.

## Recommended settings

| Setting | Value |

|---|---|

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

| LoRA strength | 0.8 – 1.0 |

| Sampler | Euler |

| Scheduler | simple |

| Steps | 50 |

| CFG | 3.0 |

| Shift | 5.0 |

| Negative-branch skip | layer 22 |

| Resolution | 1024×1024 or higher |

Negative prompt (only active above CFG 1):

worst quality, bad anatomy, watermark, jpeg artifact

Tip: with the 8-step DMD speed LoRA loaded at 0.7 and CFG 1,

the LoRA still works but the style is noticeably softer. Use CFG 3

when style fidelity matters.

## Training details

| Item | Value |

|---|---|

| Base | MOSS-Image SFT bf16 |

| Trainer | Fizgig v8.5.0 (open source) |

| Network | LoRA, attn + MLP |

| Rank / Alpha | 4 / 4 |

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

| Optimizer | AdamW |

| Epochs / Steps | 20 / 1320 |

| Batch size | 1 |

| Resolution | 768 (0.59 MP) |

| Precision | bf16 |

| EMA | 0.98 |

| Dataset | 66 images, single style |

| Captions | WD14 content tags + fixed style description |

| Text encoder | frozen (DiT-only LoRA) |

| Output | 18.2 MB safetensors, 246 modules |

Notes:

- Trained on SFT, not Base — it patches the model you are actually

loading, so strength behaves predictably.

- DiT-only adapter: no CLIP/text-encoder weights are included.

- Trained at 0.59 MP; it generalises upward but 1024+ gives the best

line quality.

- Works with the regular ComfyUI LoRA loader for MOSS-Image.