18 lines
1.2 KiB
Markdown
18 lines
1.2 KiB
Markdown
## Context
|
|
The existing local image provider can register models and route to OpenAI-compatible `/images/generations`, but it cannot execute ComfyUI workflows or inspect ComfyUI model/node availability.
|
|
|
|
## Goals / Non-Goals
|
|
- Goals: connect to local `127.0.0.1:8188`, inspect ComfyUI status and nodes, run built-in text-to-image workflows, and return generated images to TrueGrowth tasks/assets.
|
|
- Non-Goals: bundling ComfyUI, silently installing models/custom nodes, replacing TrueGrowth business workflows, or exposing remote LAN ComfyUI by default.
|
|
|
|
## Decisions
|
|
- Treat ComfyUI as a local image engine behind TrueGrowth APIs.
|
|
- Keep online text/script providers and existing avatar/voice/video runtimes unchanged.
|
|
- Start with polling `/history/{prompt_id}` after `POST /prompt`; websocket can be added later without changing the public TrueGrowth API.
|
|
- Use API-format workflow JSON templates for SDXL and Flux Schnell.
|
|
|
|
## Risks / Trade-offs
|
|
- Workflow templates depend on node availability and model filenames in ComfyUI.
|
|
- Some Flux installs require custom nodes or newer ComfyUI versions; missing nodes are reported instead of hidden.
|
|
- Model downloads and licenses remain user-confirmed.
|