19 lines
1.8 KiB
Markdown
19 lines
1.8 KiB
Markdown
## Context
|
|
The current image generation path already supports provider profiles, runtime model discovery, model bindings, and image adapters. The safest first local model integration is a provider profile for locally running OpenAI-compatible services.
|
|
|
|
## Goals / Non-Goals
|
|
- Goals: support local `/v1/models` discovery, no-auth local calls, generic `/v1/images/generations` image requests, Model Center image model installation/application, and user-facing defaults for LocalAI-style runtimes.
|
|
- Non-Goals: ComfyUI, A1111/Forge native APIs, remote LAN exposure, direct execution of every downloaded model format, or hidden license acceptance.
|
|
|
|
## Decisions
|
|
- Represent local image services as `openai-compatible` provider profiles with `authType: none`.
|
|
- Keep the request schema as `openai.image.basic-json` so existing image adapter routing remains compatible.
|
|
- Treat optional fields such as `negative_prompt`, `steps`, `cfg_scale`, and `seed` as provider extensions in the JSON body.
|
|
- Store downloaded/imported image models in a lightweight local registry separate from AIGCPanel `config.json` model packages.
|
|
- Model Center can register official model metadata and local paths, but actual generation requires a reachable OpenAI-compatible local service that has loaded the model.
|
|
|
|
## Risks / Trade-offs
|
|
- Some local runtimes use non-OpenAI APIs. They remain out of scope until a dedicated adapter is added.
|
|
- Local runtimes differ in supported parameters. Unsupported extension fields are expected to be ignored or rejected by the local service with a visible API error.
|
|
- Hugging Face model downloads may require large files, authentication, or license acknowledgement. The first implementation exposes official links and supports importing a local path instead of silently pretending all models can be downloaded through the app.
|