ImaGen/docs/architecture.md

ImaGen architecture

ImaGen is intentionally small. The framework owns plumbing; adapters own the upstream API. Each adapter only ever sees its own slice of imagen.yaml.

Layers

        ┌───────────────────────┐
        │   cmd/imagen          │   CLI dispatch
        │   (or HTTP server)    │
        └──────────┬────────────┘
                   │
        ┌──────────▼────────────┐
        │   internal/prompt     │   style preset → prompt suffix
        │   internal/output     │   filename templating, sidecar
        │   internal/config     │   YAML loader, validation
        │   internal/preview    │   tmux-img window spawner
        │   internal/cloud      │   Supabase Storage + imagen.images
        │   internal/usage      │   mai.imagen_usage cost-tracking
        └──────────┬────────────┘
                   │
        ┌──────────▼────────────┐
        │   internal/backend    │   Backend interface + Registry
        └──────────┬────────────┘
                   │
        ┌──────────▼────────────┐
        │   adapters            │   ComfyUI · Replicate · OpenAI · …
        │   (each one register- │   each registers a `type` name on
        │    s in init())       │   `backend.Default` at init time.
        └───────────────────────┘

The Backend contract

type Request struct {
    Prompt         string
    NegativePrompt string
    Width, Height  int
    Steps          int
    Seed           int64
    Style          string
    BackendOpts    map[string]any
}

type Result struct {
    ImageReader io.ReadCloser
    MimeType    string
    Metadata    map[string]any
}

type Backend interface {
    Name() string
    Generate(ctx context.Context, req Request) (*Result, error)
}

Adapters translate Request into whatever the upstream expects. Fields they can't honour (e.g. NegativePrompt on DALL-E) are silently ignored.

Registry

backend.Default holds the process-wide name → constructor map. Each adapter calls backend.Register("<type>", NewX) from its init(). The CLI imports internal/backend (which transitively triggers the mock's init) and any extra adapter packages.

Config flow

imagen.yaml
  backends:
    flux-schnell-local:
      type: comfyui                  ──┐
      base_url: http://mrock:8188      │  framework keeps `type`,
      model: flux1-schnell.safetensors │  hands the rest to the
      default_steps: 4                 │  comfyui adapter as cfg map[string]any
                                     ──┘

The framework never inspects fields below type. That's the adapter's contract with itself, expressed however the adapter wants (typed struct, map lookups, JSON tags — its call).

Output

output:
  directory: ~/Pictures/imagen
  naming: "{date}-{slug}-{seed}.png"
  write_metadata_json: true

Placeholders: {date}, {time}, {slug} (lowercased prompt, alnum-only, truncated to 40 chars), {seed}, {backend}, {ext}. The sidecar JSON contains the prompt, backend instance name, seed, ISO timestamp, and the Result.Metadata map verbatim.

Where adapters fail fast

• Missing required field in their config block — return an error from the constructor; the CLI surfaces it as imagen: backend "X": <err>.
• Unset env-var for credentials — same.
• Network errors during Generate — wrap and return; no retry policy yet (decide per-adapter, or move to a shared retry helper if a pattern emerges).

Out of scope (today)

• Image post-processing (cropping, watermarking).
• Cost-tracking (lands with the Replicate adapter, since only API backends bill).
• Multi-image n>1 per request — backends that support it can expose it via BackendOpts; the framework doesn't have a first-class field yet.