mirror of
https://github.com/willmiao/ComfyUI-Lora-Manager.git
synced 2026-03-22 13:42:12 -03:00
2b67091986
- Introduced a new WorkflowParser class to streamline workflow parsing and manage node mappers. - Added functionality to load external mappers dynamically from a specified directory. - Refactored LoraLoaderMapper and LoraStackerMapper to handle new data formats for loras and trigger words. - Updated recipe routes to utilize the new WorkflowParser for parsing workflows. - Made adjustments to the flux_prompt.json to reflect changes in active states and class types.
149 lines
3.5 KiB
Markdown
149 lines
3.5 KiB
Markdown
# ComfyUI Workflow Parser
|
||
|
||
本模块提供了一个灵活的解析系统，可以从ComfyUI工作流中提取生成参数和LoRA信息。
|
||
|
||
## 设计理念
|
||
|
||
工作流解析器基于以下设计原则：
|
||
|
||
1. **模块化**: 每种节点类型由独立的mapper处理
|
||
2. **可扩展性**: 通过扩展系统轻松添加新的节点类型支持
|
||
3. **回溯**: 通过工作流图的模型输入路径跟踪LoRA节点
|
||
4. **灵活性**: 适应不同的ComfyUI工作流结构
|
||
|
||
## 主要组件
|
||
|
||
### 1. NodeMapper
|
||
|
||
`NodeMapper`是所有节点映射器的基类，定义了如何从工作流中提取节点信息:
|
||
|
||
```python
|
||
class NodeMapper:
|
||
def __init__(self, node_type: str, inputs_to_track: List[str]):
|
||
self.node_type = node_type
|
||
self.inputs_to_track = inputs_to_track
|
||
|
||
def process(self, node_id: str, node_data: Dict, workflow: Dict, parser) -> Any:
|
||
# 处理节点的通用逻辑
|
||
...
|
||
|
||
def transform(self, inputs: Dict) -> Any:
|
||
# 由子类覆盖以提供特定转换
|
||
return inputs
|
||
```
|
||
|
||
### 2. WorkflowParser
|
||
|
||
主要解析类，通过跟踪工作流图来提取参数:
|
||
|
||
```python
|
||
parser = WorkflowParser()
|
||
result = parser.parse_workflow("workflow.json")
|
||
```
|
||
|
||
### 3. 扩展系统
|
||
|
||
允许通过添加新的自定义mapper来扩展支持的节点类型:
|
||
|
||
```python
|
||
# 在py/workflow/ext/中添加自定义mapper模块
|
||
load_extensions() # 自动加载所有扩展
|
||
```
|
||
|
||
## 使用方法
|
||
|
||
### 基本用法
|
||
|
||
```python
|
||
from workflow.parser import parse_workflow
|
||
|
||
# 解析工作流并保存结果
|
||
result = parse_workflow("workflow.json", "output.json")
|
||
```
|
||
|
||
### 自定义解析
|
||
|
||
```python
|
||
from workflow.parser import WorkflowParser
|
||
from workflow.mappers import register_mapper, load_extensions
|
||
|
||
# 加载扩展
|
||
load_extensions()
|
||
|
||
# 创建解析器
|
||
parser = WorkflowParser(load_extensions_on_init=False) # 不自动加载扩展
|
||
|
||
# 解析工作流
|
||
result = parser.parse_workflow(workflow_data)
|
||
```
|
||
|
||
## 扩展系统
|
||
|
||
### 添加新的节点映射器
|
||
|
||
在`py/workflow/ext/`目录中创建Python文件，定义从`NodeMapper`继承的类:
|
||
|
||
```python
|
||
# example_mapper.py
|
||
from ..mappers import NodeMapper
|
||
|
||
class MyCustomNodeMapper(NodeMapper):
|
||
def __init__(self):
|
||
super().__init__(
|
||
node_type="MyCustomNode", # 节点的class_type
|
||
inputs_to_track=["param1", "param2"] # 要提取的参数
|
||
)
|
||
|
||
def transform(self, inputs: Dict) -> Any:
|
||
# 处理提取的参数
|
||
return {
|
||
"custom_param": inputs.get("param1", "default")
|
||
}
|
||
```
|
||
|
||
扩展系统会自动加载和注册这些映射器。
|
||
|
||
### LoraManager节点说明
|
||
|
||
LoraManager相关节点的处理方式:
|
||
|
||
1. **Lora Loader**: 处理`loras`数组，过滤出`active=true`的条目，和`lora_stack`输入
|
||
2. **Lora Stacker**: 处理`loras`数组和已有的`lora_stack`，构建叠加的LoRA
|
||
3. **TriggerWord Toggle**: 从`toggle_trigger_words`中提取`active=true`的条目
|
||
|
||
## 输出格式
|
||
|
||
解析器生成的输出格式如下:
|
||
|
||
```json
|
||
{
|
||
"gen_params": {
|
||
"prompt": "...",
|
||
"negative_prompt": "",
|
||
"steps": "25",
|
||
"sampler": "dpmpp_2m",
|
||
"scheduler": "beta",
|
||
"cfg": "1",
|
||
"seed": "48",
|
||
"guidance": 3.5,
|
||
"size": "896x1152",
|
||
"clip_skip": "2"
|
||
},
|
||
"loras": "<lora:name1:0.9> <lora:name2:0.8>"
|
||
}
|
||
```
|
||
|
||
## 高级用法
|
||
|
||
### 直接注册映射器
|
||
|
||
```python
|
||
from workflow.mappers import register_mapper
|
||
from workflow.mappers import NodeMapper
|
||
|
||
# 创建自定义映射器
|
||
class CustomMapper(NodeMapper):
|
||
# ...实现映射器
|
||
|
||
# 注册映射器
|
||
register_mapper(CustomMapper()) |