---
title: SakunaMemoir-Lora
canonical_url: "https://www.modelscope.cn/models/abbuibuibui/SakunaMemoir-Lora"
md_url: "https://www.modelscope.cn/models/abbuibuibui/SakunaMemoir-Lora.md"
repository: abbuibuibui/SakunaMemoir-Lora
last_updated: 2026-10-29
parameters: 798.1M
tensor_type:
  - BF16
library_name:
  - safetensors
downloads: 14
stars: 0
---

# SakunaMemoir-Lora

> SakunaMemoir-Lora - abbuibuibui 在 ModelScope 开源的模型。作名（サクナ）（Sakuna Memoir）— Illustrious SDXL 角色 LoRA

- **Repository**: abbuibuibui/SakunaMemoir-Lora
- **Parameters**: 798.1M
- **Downloads**: 14
- **Stars**: 0
- **Last updated**: 2026-10-29

Source: https://www.modelscope.cn/models/abbuibuibui/SakunaMemoir-Lora

---

# 作名（サクナ）（Sakuna Memoir）— Illustrious SDXL 角色 LoRA

[English version](./README_EN.md) | 中文（当前）

非官方同人 LoRA，角色为《闭关不出的吸血鬼大小姐的苦闷》中的 **作名（サクナ）（Sakuna Memoir）**，底模 **waiIllustriousSDXL v17**（常见文件名 `waiIllustriousSDXL_v170`）。仓库含完整正式 **E1–E7**。横向测评推荐 **E6**，强度 **0.8**。训练图与 caption：[abbuibuibui/SakunaMemoir-Dataset](https://huggingface.co/datasets/abbuibuibui/SakunaMemoir-Dataset)。

非官方衍生物；角色与原作权利归原权利方。

## 文件

| Epoch | 文件 | 说明 |
| --- | --- | --- |
| E1 | `sakuna_memoir-000001.safetensors` | 备选。 |
| E2 | `sakuna_memoir-000002.safetensors` | 备选。 |
| E3 | `sakuna_memoir-000003.safetensors` | 备选。 |
| E4 | `sakuna_memoir-000004.safetensors` | 备选。 |
| E5 | `sakuna_memoir_rE5S933-000005.safetensors` | 备选。 |
| E6 | `sakuna_memoir_rE6S1055-000006.safetensors` | **推荐默认。** 0.6–0.8 综合最稳。 |
| E7 | `sakuna_memoir_rE6S1055.safetensors` | 备选。 |

日常请下 **E6**；其余六个便于对照。

E6 SHA-256：`4D8FF3E487ED403AA338ED771415C1B54D91A8A1529C3DEB1343CB8FBB9520A0`

## 展示图

**E6 @ 0.8** 直出，未修图。Euler a / 28 / CFG 4.5 / CLIP skip 2 / `832×1216`；关闭 Hires、ADetailer。

| 正脸 | 主服装全身 |
| --- | --- |
| ![Face](./showcase/01_face.png) | ![Main](./showcase/02_main_full.png) |

| 替装全身 | 复杂夜景 |
| --- | --- |
| ![Alt](./showcase/03_alt_full.png) | ![Night](./showcase/04_complex_night.png) |

## 触发词与设置

训练 caption 使用 `keep_tokens = 2`。

| 角色 | Token | 观感 |
| --- | --- | --- |
| 身份 | `sakuna_memoir` | silver hair, short hair, blue eyes |
| 服装 | `sakuna_memoir_crimson_lord_uniform` | Light-blue crimson-lord military uniform, dark bow, white flower hair clip |
| 服装 | `sakuna_memoir_white_t_shirt` | White T-shirt casual look |
| 服装 | `sakuna_memoir_loungewear` | Loungewear |
| 服装 | `sakuna_memoir_sleepwear` | Sleepwear |
| 服装 | `sakuna_memoir_swimsuit` | Swimsuit |

推荐：waiIllustriousSDXL v17 或 Illustrious / NoobAI 系；强度 **0.8**；Euler a，28，CFG 4.5，CLIP skip 2，`832×1216`。

示例：

```text
<lora:sakuna_memoir_rE6S1055-000006:0.8>,
masterpiece, best quality, very aesthetic, absurdres, anime screencap,
1girl, solo, sakuna_memoir, sakuna_memoir_crimson_lord_uniform,
silver hair, short hair, blue eyes,
upper body, looking at viewer, outdoors
```

## 为何选 E6

七个正式 epoch 先在 0.8 做正脸 + 主装全身 + 替装全身筛选；**E5 / E6 / E7** 进入权重矩阵（10 场景 × 4 强度 = 120 张直出）。

| 候选 | 观察 | 结论 |
| --- | --- | --- |
| E5 | 可用，细节略软 | 备选 |
| E6 | 正脸 / 服装 / 夜景在 0.6–0.8 最稳 | **推荐默认 @ 0.8** |
| E7 | 接近 E6，1.0 无稳定收益 | 备选归档 |

中文测评笔记：[`evaluation/eval_notes_zh.md`](./evaluation/eval_notes_zh.md)。

## 公开横向测评

联络表：**列 = epoch，行 = 强度**。均为原图缩放，未修图换图。

### Round 1：E1–E7 @ 0.8

| Front portrait | Crimson uniform full |
| --- | --- |
| ![r1a](./evaluation/round1_face.jpg) | ![r1b](./evaluation/round1_crimson.jpg) |

![r1c](./evaluation/round1_white_t.jpg)

### Round 3：E5 / E6 / E7 × 0.4 / 0.6 / 0.8 / 1.0

| 正脸 | 三分之二侧脸 |
| --- | --- |
| ![Face](./evaluation/01_face.jpg) | ![TQ](./evaluation/02_three_quarter.jpg) |

![侧脸](./evaluation/03_profile.jpg)

| Crimson lord uniform full | White T-shirt full |
| --- | --- |
| ![m1](./evaluation/04_crimson_full.jpg) | ![m2](./evaluation/05_white_t_full.jpg) |

![坐姿](./evaluation/07_sitting.jpg)

![简洁背景](./evaluation/08_simple_bg.jpg)

![复杂夜景](./evaluation/09_complex.jpg)

![不写服装词](./evaluation/10_no_outfit.jpg)

![无触发词](./evaluation/11_no_trigger.jpg)

## 训练

- 数据：100 张精选静帧 + 同名 `.txt`；caption 以 `sakuna_memoir, <outfit>` 开头。服装分布：{'Crimson': 80, 'Lounge': 4, 'Sleep': 6, 'Swim': 3, 'WhiteT': 7}。数据集：[abbuibuibui/SakunaMemoir-Dataset](https://huggingface.co/datasets/abbuibuibui/SakunaMemoir-Dataset)。
- 网络：LoRA rank 32 / alpha 16 (UNet + text encoders), AdamW8bit, bf16, 1024 buckets.
- 100 images × repeat 2 × 7 epochs ≈ 1400 optimizer steps (resumed mid-run; formal E1–E7 kept).
- bf16，batch 1，latent 落盘缓存；关闭 flip / color aug。

## 相关仓库

| | Hugging Face | 魔塔 |
| --- | --- | --- |
| 模型 | [abbuibuibui/SakunaMemoir-Lora](https://huggingface.co/abbuibuibui/SakunaMemoir-Lora) | [abbuibuibui/SakunaMemoir-Lora](https://www.modelscope.cn/models/abbuibuibui/SakunaMemoir-Lora) |
| 数据集 | [abbuibuibui/SakunaMemoir-Dataset](https://huggingface.co/datasets/abbuibuibui/SakunaMemoir-Dataset) | [abbuibuibui/SakunaMemoir-Dataset](https://www.modelscope.cn/datasets/abbuibuibui/SakunaMemoir-Dataset) |

## 兼容与限制

角色 LoRA，非通用画风。无服装 token 时中高强度常回落到主训练装。

## 许可

- © Kotei Kobayashi / SB Creative / the relevant production committees.
- 权重：**CreativeML OpenRAIL-M**。
- 仅供同人非商业用途；生成图还需遵守底模许可。
- 禁止用于违法、欺诈、骚扰或剥削用途。