Diffused Lemon - AI Image Generator

A 3-layer AI image generation application using Lemonade server.

Architecture

1. Backend: Lemonade server (OpenAI-compatible)
2. Middleware: Python service with aiohttp
3. Frontend: Single-file HTML/JS

Setup

Prerequisites

• Python 3.8+
• Lemonade server running with image generation models (SD-Turbo, SDXL-Turbo, etc.)
• Lemonade server running with fluxassistant model for prompt expansion

Setting up fluxassistant in Lemonade Server

The prompt expansion feature requires the fluxassistant model. Import it from Hugging Face:

1. Open Lemonade Server's web interface
2. Go to Models → Import Model
3. Use this Hugging Face model: https://huggingface.co/mradermacher/Llama-3.2-3B-Fluxed-uncensored-GGUF
4. Name the imported model: user.fluxassistant
5. Wait for the model to download and load

Alternatively, use the API:

curl -X POST http://localhost:8000/api/v1/models/import \
  -H "Content-Type: application/json" \
  -d '{
    "source_url": "https://huggingface.co/mradermacher/Llama-3.2-3B-Fluxed-uncensored-GGUF",
    "model_name": "user.fluxassistant"
  }'

Installation

# Install middleware dependencies
pip install -r middleware/requirements.txt

Environment variables override config.json values. The config.json file is automatically loaded from the middleware/ directory.

Variable	Description
LM_SERVER_URI	Lemonade server URL
LM_STORAGE_DIR	Image storage directory
LM_LOG_FILE	Log file path
LM_LOG_LEVEL	Logging level
LM_AUTH_TOKEN	API authentication token
LM_DEFAULT_MODEL	Default image generation model
LM_DEFAULT_SIZE	Default image size
LM_PROMPT_ASSIST_MODEL	Model for prompt expansion
LM_PROMPT_ASSIST_SYSTEM_PROMPT	System prompt for expansion

Configuration

Edit middleware/config.json or use environment variables:

Variable	Description	Default
server_uri	Lemonade server URL	http://localhost:8000
storage_dir	Image storage directory	/path/to/storage
log_file	Log file path	/path/to/logs.json
log_level	Logging level (DEBUG, INFO, WARNING, ERROR)	INFO
auth_token	API authentication token	null
default_model	Default image generation model	SD-Turbo
default_size	Default image size	512x512
prompt_assist_model	Model for prompt expansion	fluxassistant
prompt_assist_system_prompt	System prompt for expansion	See config.json

Environment Variables

Environment variables override config.json values:

• LM_SERVER_URI - Backend server URL
• LM_STORAGE_DIR - Image storage directory
• LM_LOG_FILE - Log file path
• LM_AUTH_TOKEN - API authentication token
• LM_DEFAULT_MODEL - Default image generation model
• LM_DEFAULT_SIZE - Default image size
• LM_PROMPT_ASSIST_MODEL - Model for prompt expansion
• LM_PROMPT_ASSIST_SYSTEM_PROMPT - System prompt for expansion

Example config.json

{
  "server_uri": "http://localhost:8000",
  "storage_dir": "/home/ckuethe/diffused-lemon/storage",
  "log_file": "/home/ckuethe/diffused-lemon/logs.json",
  "log_level": "INFO",
  "auth_token": null,
  "default_model": "SD-Turbo",
  "default_size": "512x512",
  "prompt_assist_model": "fluxassistant",
  "prompt_assist_system_prompt": "Expand the user's simple image generation prompt into a detailed, descriptive prompt that would help generate a more interesting image. Keep it concise but evocative."
}

Usage

Basic Workflow

1. Open http://localhost:8080 in your browser (middleware serves the frontend)
2. The UI shows server connection status in the header
3. Select a model from the dropdown (auto-populated from Lemonade server)

Generating an Image

1. Enter a simple prompt (e.g., "mountain landscape")
2. Click Prompt Assist to expand it using fluxassistant
3. Edit the expanded prompt if desired
4. Adjust parameters (size, steps, seed)
5. Click Generate to create the image
6. The image appears in the main display area
7. Click Download to save the image

Advanced Features

• Prompt Assist: Click to expand simple prompts into detailed descriptions using fluxassistant
• Image History: View all previously generated images in the grid below
• Click thumbnails: View full-size images from history
• Random Seed: Generate random seeds with the dice button
• Model Preference: Your selected model is saved to localStorage

Parameters

• Model: Image generation model (SD-Turbo, SDXL-Turbo, etc.)
• Prompt: Text description of the image you want to generate
• Size: Image dimensions (256x256, 512x512, 768x768, 1024x1024)
• Steps: Number of denoising steps (4-50, higher = more detail but slower)
• Seed: Random seed for reproducibility (leave empty for random)

Storage

Images and metadata are stored in:

storage/
├── images/      # Generated PNG files
│   └── 2024-04-01_12-30-45.png
└── metadata/    # JSON metadata files
    └── 2024-04-01_12-30-45.json

Metadata Format

Each image has a corresponding JSON file with:

{
  "filename": "2024-04-01_12-30-45.png",
  "prompt": "A serene mountain landscape at sunset...",
  "model": "SD-Turbo",
  "size": "512x512",
  "seed": 12345,
  "steps": 4,
  "cfg_scale": 1.0,
  "timestamp": "2024-04-01T12:30:45Z",
  "prompt_assisted": true,
  "original_prompt": "mountain landscape"
}

API Endpoints

The middleware exposes the following endpoints:

• GET / - Main frontend page (HTML)
• GET /health - Server health check
• GET /models - List available models
• POST /prompt-assist - Expand a prompt using fluxassistant
• POST /generate - Generate an image
• GET /images - List image history (query param: ?limit=50)
• GET /images/{filename} - Get a specific image

Features

• Prompt Assist: Uses fluxassistant to expand simple prompts into detailed descriptions
• Image History: View and download previously generated images
• Responsive UI: Works on mobile and desktop
• JSON Logging: All operations logged in JSON format
• LocalStorage: Saves model preference and recent images
• Base64 Images: Images displayed directly in browser without intermediate files
• Single Server: Frontend served by middleware (relative URLs work from any host)

Code Style

Python code should be formatted with Ruff:

# Install ruff (if not already installed)
pip install ruff

# Format all Python files
ruff format middleware/

# Or format a specific file
ruff format middleware/server.py

Running the Server

The middleware server serves both the API and the frontend UI:

# From project root, run as module
python3 -m middleware.server

# Or directly (requires path setup)
python3 middleware/server.py

The server starts on port 8080 by default. Open http://localhost:8080 in your browser.

Running from another machine

To access the server from another machine on your network, bind to your machine's IP address or 0.0.0.0:

# Use your machine's IP (e.g., 192.168.1.100)
python3 middleware/server.py --host 192.168.1.100

# Or bind to all interfaces
python3 middleware/server.py --host 0.0.0.0

Then access from another machine: http://192.168.1.100:8080

Debugging

Use the --verbose or --debug flags for detailed logging:

# Verbose logging
python3 middleware/server.py --verbose

# Debug logging (most detailed)
python3 middleware/server.py --debug

# Combined with custom host/port
python3 middleware/server.py --host 0.0.0.0 --port 8080 --debug

Troubleshooting

Server won't start

• Check that Lemonade server is running and accessible
• Verify server_uri in config.json points to the correct Lemonade server URL
• Check log file for error messages

No models found

• Ensure Lemonade server has image generation models loaded
• Check that models have image in their labels or sd/flux in their IDs

Prompt Assist fails

• Verify that fluxassistant model is loaded in Lemonade server
• Check the system prompt configuration in config.json

Images not saving

• Verify storage_dir exists and is writable
• Check that the middleware process has permissions to write to the storage directory

License

This project is in the Public Domain.

I make no claim of copyright to any part of this project. It was generated using AI prompting with opencode and Qwen3 Coder Next. All code and content is freely available for any use without restrictions.

Credits

This section is the only part of this project actually written by me, @ckuethe. Everything else is the product of AI prompting.

I used Lemonade on an AMD Ryzen AI Max 395; opencode and Qwen3 Coder Next for code generation.

Thanks to HAL Heavy Duty on youtube for reminding us to say "Good job, buddy. Good job." when your machines and tools do what you want.

Thanks to @technigmaai for this wiki page on how to get Strix Halo to use more VRAM than the BIOS would otherwise allow.

Vincent Gourbin and mradermacher for the Llama model tuned to elaborate on vague image prompts.

Massive Attack for "Dissolved Girl", and gemma3-4b-FLM for deciding that "Diffused Lemon" would be a good name.

Examples

See examples/ for sample outputs:

Desktop Mode	Mobile Mode