Backend Platform for AI Image Generation

A backend platform for AI image generation with LoRA style adapters, running entirely on a local machine using Firebase emulators.

Architecture

Publisher Client
      │  writes generation_requests doc (CREATED)
      â–¼
Firestore Emulator
      │  onDocumentCreated trigger
      â–¼
Cloud Function ──── GET /v1/config/{user_id} ────► Config Service
      │  POST /generate
      â–¼
Inference Server (FastAPI + LCM Diffusion)
      │  updates doc status: PROCESSING → DONE / FAILED
      â–¼
Firestore Emulator

Prerequisites

• Node.js 20+
• Python 3.11+
• Java (required by Firebase emulators)
• Firebase CLI: npm install -g firebase-tools

Setup

# Copy and configure environment variables
cp .env.example .env
# Edit .env and set a secure API_KEY

# Install all dependencies
.\setup.ps1

Running

.\start.ps1

This runs all tests first, then opens three service windows:

Service	URL
Config Service	http://127.0.0.1:3000
Firebase Emulators (UI)	http://127.0.0.1:4000
Inference Server	http://127.0.0.1:8000

Once all services are up, publish generation requests:

cd publisher
npm start

Environment Variables

Variable	Description	Default
API_KEY	Shared secret between Cloud Function and Inference Server	—
INFERENCE_SERVER_URL	Base URL of the inference server	http://127.0.0.1:8000
CONFIG_SERVICE_URL	Base URL of the config service	http://127.0.0.1:3000
FIREBASE_PROJECT_ID	Firebase project ID (use demo-* for emulators)	demo-local
AUTH_EMULATOR_HOST	Firebase Auth emulator host:port	127.0.0.1:9099
FIRESTORE_EMULATOR_HOST	Firestore emulator host:port	127.0.0.1:8080

Running Tests

# Cloud Function
cd functions && npm test

# Config Service
cd config-service && npm test

# Inference Server
cd inference-server
.\venv\Scripts\Activate.ps1   # source venv/bin/activate on Linux/macOS
pytest tests/

---

API Reference

Config Service — http://127.0.0.1:3000

GET /v1/config/{user_id}

Returns a LoRA style configuration for the given user. Results are cached per user for 60 seconds.

Response 200 OK:

{
  "lora_url": "https://huggingface.co/vislupus/SD1.5-LoRA-Loving-Vincent-Style/resolve/main/vg_style_v1-000048.safetensors",
  "lora_weight": 0.9,
  "updated_at": "2026-04-28T10:00:00.000Z"
}

Response 404 Not Found — no configuration available for that user.

---

Inference Server — http://127.0.0.1:8000

All endpoints require the header:

Authorization: Bearer <API_KEY>

POST /generate

Generates an image from a text prompt with an optional LoRA adapter. Updates the Firestore document status throughout the lifecycle (PROCESSING → DONE / FAILED).

Request body:

{
  "doc_id": "firestore-document-id",
  "prompt": "a forest cabin in winter, oil painting style",
  "lora_url": "https://example.com/lora.safetensors",
  "lora_weight": 0.8
}

lora_url and lora_weight are optional. If omitted, the base model runs without a LoRA adapter.

Response 200 OK:

{
  "image": "<base64-encoded-PNG>"
}

The generated image is also saved to inference-server/outputs/{doc_id}.png.

Error responses:

Status	Reason
401 Unauthorized	Missing or invalid Authorization header
500 Internal Server Error	Generation failed; Firestore doc updated to FAILED with an error field

---

Status Lifecycle

Status	Set by
CREATED	Publisher Client
QUEUED	Cloud Function
PROCESSING	Inference Server
DONE / FAILED	Inference Server

Project Structure

.
├── config-service/     # Express — GET /v1/config/:user_id
├── functions/          # Firebase Cloud Function (v2) — onDocumentCreated
├── inference-server/   # FastAPI — POST /generate
├── publisher/          # One-shot script — writes requests to Firestore
├── firestore.rules     # Firestore security rules
├── firebase.json       # Firebase emulator config
├── setup.ps1           # Install all dependencies
└── start.ps1           # Run tests then start all services