eight tohyama style LoRA for Anima (base v1.0) and Anima-2.9B

A Danbooru-tag LoRA trained to replicate the art style of the illustrator @eight tohyama — clean linework, soft cel shading, delicate facial features, and light pastel palettes with detailed hair rendering.

Works on both the original Anima base v1.0 and the new Anima-2.9B (40-layer expanded) model.

### Trigger words

```

@eight tohyama

```

Put it at the very front of your prompt (keep-token 0). The tag also blends well with other artist tags for style mixing.

### Supported characters

Trained on the artist's original character roster — works with the following character names (Danbooru tags, use as-is):

| Character | Tags to add |

|---|---|

| anri (main, 10 imgs) | anri — aqua blue eyes, small breasts, blue hair clip, long hair |

| saito | saito — black hair side ponytail, pink eyes |

| suou | suou — black hair, short twintails, blunt bangs, brown eyes, loli/petite |

| Graf Zeppelin (Kancolle) | graf zeppelin (kancolle) — blonde, iron cross, kancolle uniform |

Character name goes right after the artist tag. For copyright characters, follow the name with the series tag kancolle).

### Versions

| Version | Base model | File | Size |

|---|---|---|---|

| 1.0 (base) | anima-base-v1.0.safetensors | eight_tohyama_base_1.0.safetensors | 132 MB |

| 1.0 (2.9B) | Anima-2.9B-preview-v1.safetensors | eight_tohyama_1.0.safetensors | 172 MB |

> ⚠️ Version-specific: the 2.9B build is trained on the 40-layer expanded architecture and is not compatible with base v1.0 (and vice versa). Use the matching file for your base model.

### Training details

- Dataset: 33 images of eight tohyama artwork (Danbooru tags, 1024² resolution, 4 repeats)

- Trained with kohya sd-scripts anima_train_network.py)

- LoRA dim 16 / alpha 16, LR 3e-5, 8-bit AdamW, bf16, constant schedule, seed 42

- base 1.0 build: 40 epochs → best at epoch 30

- 2.9B build: 30 epochs → best at epoch 22

- Bucket training enabled; no regularisation images

### Recommended settings

Anima base v1.0 (with this LoRA at strength 1.0):

| Parameter | Value |

|---|---|

| Sampler | er_sde (or euler_a) |

| Scheduler | simple (or beta57) |

| Steps | 25–30 |

| CFG | 4.0–4.5 |

| Flow shift | 3.0 |

| Resolution | 888×1184 / 1024² / up to 1152×1536 |

| LoRA strength | 0.8–1.0 (1.0 default; lower for subtler style) |

Anima-2.9B (with this LoRA at strength 1.0):

| Parameter | Value |

|---|---|

| Sampler | euler / res-multistep / er-sde |

| Scheduler | sgm-uniform / beta / linear-quadratic |

| Steps | 28–50 (sweet spot ~30) |

| CFG | 3.5–5 |

| Flow shift | 2.0–3.0 |

| Resolution | 812×1216 / 1152×1536 / 16:9 up to 1496×840 |

| LoRA strength | 0.8–1.0 |

### Tips

- Detailed prompts work best; the style tag blends with character tags, series tags and other artist tags

- For face-focused portraits, 3/4 view reduces top whitespace; avoid frame / out of frame wording (triggers picture-in-picture artifacts on 16:9)

- Negative prompt starters: worst quality, low quality, artist name, blurry, watermark, text

- No score tags needed; the model responds to quality tags masterpiece, best quality)

### Notes

- Derivative of [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima) — non-commercial license applies to generated content and redistribution. Mark this LoRA as Non-Commercial on Civitai.

- Training data captions are Danbooru-style tags; keep-token 0 keeps the artist tag at the prompt head.