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 | 后缀为 .pth 的模型文件 (Model file with .pth extension) |
| Text Encoder | gemma_2_2b_fp16.safetensors | gemma_2_2b_fp16.safetensors 文件 |
| 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
- 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)
- 括号权重需转义：
\(word\)。(Parentheses for weights need to be escaped:\(word\)). - 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 (likelumina_workflow.json) has optimized parameters; it is recommended to import and use it directly). - 多语言提示词效果需自行测试。(The effect of multilingual prompts needs to be tested by yourself).
- 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.)