guide.md

# Nieta Lumina2.0 ComfyUI 使用手册 (Lumina2.0 ComfyUI User Manual) - guide_zh.md

 
## 一、环境准备 (Environment Preparation)

### 1. 必备组件下载 (Required Components Download)

| 组件类型 (Component Type) | 下载地址 (Download Address)                               | 备注 (Remarks)                                  |
| :------------------------ | :-------------------------------------------------------- | :---------------------------------------------- |
| Lumina DIT 模型          | [NebulaeWis/NietaAniLumina_Alpha · Hugging Face](https://huggingface.co/NebulaeWis/NietaAniLumina_Alpha) | 后缀为 `.pth` 的模型文件 (Model file with .pth extension) |
| Text Encoder              | [gemma_2_2b_fp16.safetensors](https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/blob/main/split_files/text_encoders/gemma_2_2b_fp16.safetensors) | `gemma_2_2b_fp16.safetensors` 文件                   |
| VAE                       | [ae.safetensors](https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/blob/main/split_files/vae/ae.safetensors) | `ae.safetensors` 文件                             |

### 2. 软件与工作流 (Software and Workflow)

* **ComfyUI**: 必须更新至最新版本。(Must be updated to the latest version).
* **工作流文件 (Workflow File)**:
    * 推荐使用官方优化参数的JSON工作流文件。 (It is recommended to use the official JSON workflow file with optimized parameters.)
    * 示例工作流文件(The example workflow file ) [lumina_workflow.json](./lumina_workflow.json) 
    * 通过 ComfyUI 的 "Load" 功能导入工作流文件即可快速配置。(Quickly configure by importing the workflow file through ComfyUI's "Load" function.)

## 二、模型加载配置 (Model Loading Configuration)

### 节点说明（关键节点参数）(Node Description - Key Node Parameters)

| 节点名称 (Node Name)       | 参数设置 (Parameter Settings)             | 示例值 / 备注 (Example Value / Remarks)                                  |
| :------------------------- | :---------------------------------------- | :----------------------------------------------------------------------- |
| UNETLoader                 | 模型路径选择 (Model path selection)       | `lu2/results_cosine_2e-4_bs64_infallssssuum/checkpoint-e4_s68300/consolidated.00-of-01.pth` (请根据实际下载路径修改 Please modify according to your actual download path) |
| ModelSamplingAuraFlow      | 移位值 (Shift) (Shift value)             | `6.00`（官方推荐值 Official recommended value）                         |
| VAELoader                  | VAE 文件选择 (VAE file selection)         | `ae.safetensors` (请确保已下载并放置到ComfyUI的VAE目录下 Please ensure it's downloaded and placed in ComfyUI's VAEs directory) |
| CLIPLoader                 | Text Encoder 选择 (Text Encoder selection) | `gemma_2_2b_fp16.safetensors`，预设选 `lumina2` (preset select lumina2; 请确保已下载并放置到ComfyUI的CLIP目录下 Please ensure it's downloaded and placed in ComfyUI's CLIPs directory) |
| EmptySD3LatentImage        | 分辨率设置 (Resolution settings)          | 默认 `1024x1552`（支持 `1024 - 1280` 分辨率 Default 1024x1552, supports 1024 - 1280 resolution) |

## 三、提示词规范 (Prompt Specification)

### 1. 系统提示词 (System Prompt)

* 需要前缀以下文本（可修改，对出图有影响）：(The following text needs to be prefixed (can be modified, affects image generation)):
    `You are an assistant designed to generate anime images based on textual prompts. <Prompt Start>`
* 示例 (Example)：
    `You are an assistant... <Prompt Start> 1girl, @wlop,..................................`

### 2. 多语言支持 (Multilingual Support)

支持自然语言中英文混合输入（Gemma2等llm作为te的特性）：(Supports mixed Chinese and English natural language input (a feature of LLMs like Gemma2 as Text Encoder)):
`一只坐在樱花树下的猫，sunlight effect, fantasy background`

### 3. Danbooru Tag 规范 (Danbooru Tag Specification)

* 下划线转空格 (Underscore to space)：`ningen_mamei` → `ningen mamei`
* 括号转义 (Parentheses escaping)：`dressed shadow \(meme\)`
* 画师强化 (Artist emphasis)：使用 `@` 符号（如 `@wlop`）(Use the @ symbol (e.g., @wlop))

### 4. 负面提示词 (Negative Prompts)

需包含相同系统前缀：(Must include the same system prefix):
`You are an assistant... <Prompt Start> blurry, worst quality, low quality`

## 四、采样参数 (Sampling Parameters)

| 参数项 (Parameter Item) | 推荐值 (Recommended Value) | 说明 (Description)                     |
| :---------------------- | :------------------------- | :------------------------------------- |
| 采样器 (Sampler)        | `res_multistep`            | 默认，可自行调整 (Default, can be adjusted) |
| 调度器 (Scheduler)      | `linear_quadratic`         |                                        |
| CFG Scale               | `5.5`                      | 可自行调整 (Can be adjusted)             |
| 步数 (Steps)            | `30`                       | 可自行调整 (Can be adjusted)             |
| 降噪 (Denoise)          | `1.0`                      |                                        |
| 分辨率 (Resolution)     | `1024x1024` 至 `1280x1280` | 其余分辨率效果未验证 (Effects of other resolutions have not been verified) |

## 五、工作流说明 (Workflow Description)

### 1. 流程概览 (Process Overview)

UNET → AuraFlow采样 (AuraFlow Sampling) → KSampler → VAE解码 (VAE Decode) → 输出图像 (Output Image)

### 2. 关键连接 (Key Connections)

* Text Encoder 输出分别连接正 / 负提示词节点。(Text Encoder output connects to positive/negative prompt nodes respectively).
* AuraFlow 节点需插入 UNET 和 KSampler 之间。(The AuraFlow node needs to be inserted between UNET and KSampler).

## 六、注意事项 (Precautions)

1.  括号权重需转义：`\(word\)`。(Parentheses for weights need to be escaped: `\(word\)`).
2.  官方 JSON 工作流 (如 `lumina_workflow.json`) 已优化参数，建议直接导入使用。(The official JSON workflow (like `lumina_workflow.json`) has optimized parameters; it is recommended to import and use it directly).
3.  多语言提示词效果需自行测试。(The effect of multilingual prompts needs to be tested by yourself).
4.  目前Lumina推荐的提示词格式是 "tag + 自然语言"，但也支持纯tag和纯自然语言，大家可以多试试。(Currently, Lumina's recommended prompt format is "tag + natural language", but it also supports pure tags and pure natural language. Everyone is encouraged to try them out).

**注：本手册基于内测版本编写，部分参数可能随版本更新调整。** (Note: This manual is based on the internal test version, and some parameters may be adjusted with version updates.)