> This document describes, at a high level, the design and implementation of workflows in the InvokeAI frontend. There are a substantial number of implementation details not included, but which are hopefully clear from the code.
Nodes have any number of **input fields** and **output fields**. Edges connect nodes together via their inputs and outputs. Fields have data types which dictate how they may be connected.
During execution, a nodes' outputs may be passed along to any number of other nodes' inputs.
This includes the **Text to Image**, **Image to Image** and **Unified Canvas** tabs.
The user-managed parameters on these tabs are stored as simple objects in the application state. When the user invokes, adding a generation to the queue, we internally build a graph from these parameters.
This logic can be fairly complex due to the range of features available and their interactions. Depending on the parameters selected, the graph may be very different. Building graphs in code can be challenging - you are trying to construct a non-linear structure in a linear context.
The Workflow Editor is a visual graph editor, allowing users to draw edges from node to node to construct a graph. This _far_ more approachable way to create complex graphs.
To support these qualities, workflows are serializable, have a versioned schemas, and represent graphs as minimally as possible. Fortunately, the reactflow state for nodes and edges works perfectly for this.
The reactflow state is, however, structurally incompatible with our backend's graph structure. When a user invokes on a Workflow, we need to convert the reactflow state into an InvokeAI graph. This is far simpler than the graph building logic from the Linear UI:
Graphs are very capable data structures, but not everyone wants to work with them all the time.
To allow less technical users - or anyone who wants a less visually noisy workspace - to benefit from the power of nodes, InvokeAI has a workflow feature called the Linear View.
A workflow input field can be added to this Linear View, and its input component can be presented similarly to the Linear UI tabs. Internally, we add the field to the workflow's list of exposed fields.
OpenAPI is a schema specification that can represent complex data structures and relationships. The backend is capable of generating an OpenAPI schema for all invocations.
When the UI connects, it requests this schema and parses each invocation into an **invocation template**. Invocation templates have a number of properties, like title, description and type, but the most important ones are their input and output **field templates**.
Invocation and field templates are the "source of truth" for graphs, because they indicate what the backend is able to process.
When a user adds a new node to their workflow, these templates are used to instantiate a node with fields instantiated from the input and output field templates.
**Stateful** fields store their value in the frontend graph. Think primitives, model identifiers, images, etc. Fields are only stateful if the frontend allows the user to directly input a value for them.
Many field types, however, are **stateless**. An example is a `UNetField`, which contains some data describing a UNet. Users cannot directly provide this data - it is created and consumed in the backend.
Stateless fields do not store their value in the node, so their field instances do not have values.
"Custom" fields will always be treated as stateless fields.
- If a field is annotated in python as a singular value or class, its field type is parsed as a **SINGLE** type (e.g. `int`, `ImageField`, `str`).
- If a field is annotated in python as a list, its field type is parsed as a **COLLECTION** type (e.g. `list[int]`).
- If it is annotated as a union of a type and list, the type will be parsed as a **SINGLE_OR_COLLECTION** type (e.g. `Union[int, list[int]]`). Fields may not be unions of different types (e.g. `Union[int, list[str]]` and `Union[int, str]` are not allowed).
The majority of data structures in the backend are [pydantic] models. Pydantic provides OpenAPI schemas for all models and we then generate TypeScript types from those.
> In python, invocations are pydantic models with fields. These fields become node inputs. The invocation's `invoke()` function returns a pydantic model - its output. Like the invocation itself, the output model has any number of fields, which become node outputs.
We customize the OpenAPI schema to include additional properties on invocation and field schemas. To facilitate parsing this schema into templates, we modify/wrap the types from [openapi-types] in `openapi.ts`.
When a field is annotated as a primitive values (e.g. `int`, `str`, `float`), the field type parsing is fairly straightforward. The field is represented by a simple OpenAPI **schema object**, which has a `type` property.
When a field is annotated as a pydantic model (e.g. `ImageField`, `MainModelField`, `ControlField`), it is represented as a **reference object**. Reference objects are pointers to another schema or reference object within the schema.
We need to **dereference** the schema to pull these out. Dereferencing may require recursion. We use the reference object's name directly for the field type name.
> Unfortunately, at this time, we've had limited success using external libraries to deference at runtime, so we do this ourselves.
When a field is annotated as a list of a single type, the schema object has an `items` property. They may be a schema object or reference object and must be parsed to determine the item type.
When a field is annotated as a union of a type and list of that type, the schema object has an `anyOf` property, which holds a list of valid types for the union.
After verifying that the union has two members (a type and list of the same type), we use the type for field type name, with cardinality `"SINGLE_OR_COLLECTION"`.
Stateful fields all get a function to build their template, while stateless fields are constructed directly. This is possible because stateless fields have no default value or constraints.
Field outputs are similar to stateless fields - they do not have any value in the frontend. When building their templates, we don't need a special function for each field type.
As described above, the workflow editor state is the essentially the reactflow state, plus some extra metadata.
We provide reactflow with an array of nodes and edges via redux, and a number of [event handlers][reactflow-events]. These handlers dispatch redux actions, managing nodes and edges.
The pieces of redux state relevant to workflows are:
-`state.nodes.nodes`: the reactflow nodes state
-`state.nodes.edges`: the reactflow edges state
-`state.nodes.workflow`: the workflow metadata
#### Building Nodes and Edges
A reactflow node has a few important top-level properties:
-`id`: unique identifier
-`type`: a string that maps to a react component to render the node
-`position`: XY coordinates
-`data`: arbitrary data
When the user adds a node, we build **invocation node data**, storing it in `data`. Invocation properties (e.g. type, version, label, etc.) are copied from the invocation template. Inputs and outputs are built from the invocation template's field templates.
Workflows may be loaded from external sources or the user's local instance. In all cases, the workflow needs to be handled with care, as an untrusted object.
Loading has a few stages which may throw or warn if there are problems:
- Parsing the workflow data structure itself, [migrating](#workflow-migrations) it if necessary (throws)
- Check for a template for each node (warns)
- Check each node's version against its template (warns)
- Validate the source and target of each edge (warns)
When the workflow schema changes, we may need to perform some data migrations. This occurs as workflows are loaded. zod schemas for each workflow schema version is retained to facilitate migrations.
Previous schemas are in folders in `invokeai/frontend/web/src/features/nodes/types/`, eg `v1/`.