GitHub - Lywald/SceneSynth: AI-Powered Hierarchical Scene Graph Generator for 2D Level Design

4 min read Original article ↗

SceneSynth

AI-Powered Hierarchical Scene Graph Generator for 2D Level Design

Transform natural language into explorable, nested world structures

Demo A Demo B

Powered by Gemini + Nano Banana (gemini-2.5-flash-image)

What is SceneSynth?

SceneSynth is a desktop application that generates semantic scene graphs from text prompts using AI. Describe a world, and watch it materialize as an interactive node graph. Each node can be expanded into its own detailed sub-graph, enabling infinite hierarchical world-building—from continents down to individual rooms.

Then bring your scenes to life with Nano Banana (gemini-2.5-flash-image)—Google's native image generation model—which renders your semantic graphs into vivid scene illustrations.

NEWS: Working on SceneSynth WEB, for no-coders to use this as a website, easy to use -- URL Will be Here SOON (star the project and stay in touch for January 2026)

SceneSynth Web

Key Features

  • Natural Language to Scene Graph — Describe your world in plain text; AI generates a structured graph
  • Recursive Drill-Down — Double-click any node to expand it into a detailed child graph
  • Semantic Positioning — Nodes are placed meaningfully (sky elements at top, underground at bottom)
  • Graph Modification via Chat — Refine your scene with instructions like "Add a tavern near the market"
  • Nano Banana Rendering — Transform your graphs into stunning visuals with gemini-2.5-flash-image
  • Multiple Art Styles — Fantasy, photorealistic, pixel art, watercolor, anime, and custom styles
  • Full Project Persistence — Save/load entire hierarchies including rendered images

Nano Banana: AI Scene Rendering

SceneSynth leverages Nano Banana (gemini-2.5-flash-image), Google's native multimodal image generation model via Vertex AI, to render your semantic graphs as actual scene artwork.

Semantic Scene Graph Nano Banana Render
Graph 1 Render 1
Graph 2 Render 2
Graph 4 Render 4

How it works:

  1. Your scene graph (nodes, relationships, spatial layout) is serialized to JSON
  2. A screenshot of your graph canvas is captured
  3. Both are sent to Nano Banana with your chosen art style
  4. The model generates a cohesive scene illustration respecting the semantic layout

Available Styles:

  • Top-down 2D Game Art
  • Fantasy Illustration
  • Photorealistic
  • Pixel Art (16-bit)
  • Watercolor
  • Anime/Manga
  • Dark Fantasy
  • Isometric
  • Custom prompts supported

Rendered images are stored with each graph and persist across save/load cycles.

Prerequisites

Important: SceneSynth requires Google Cloud services for AI functionality.

Requirement Purpose
Google Cloud Account Access to Vertex AI services
Vertex AI API Must be enabled in your GCP project
Gemini API Key For graph generation (Get one here)
GCP Project ID For image rendering via Vertex AI
gcloud CLI For authentication

Authentication Setup

# Install gcloud CLI if needed, then authenticate:
gcloud auth application-default login

Installation

# Clone the repository
git clone https://github.com/yourusername/SceneSynth.git
cd SceneSynth

# Install dependencies
pip install -r requirements.txt

# Run the application
python main.py

Dependencies

  • Python 3.10+
  • PyQt6
  • google-genai
  • networkx
  • pydantic

Configuration

  1. Launch SceneSynth
  2. Go to Edit → API Settings
  3. Enter your Gemini API Key (for graph generation)
  4. Enter your GCP Project ID (for image rendering)
  5. Click Save

Quick Start

  1. Generate a World — Enter a prompt like "A medieval fantasy kingdom with diverse regions" and click Generate
  2. Explore the Graph — Pan, zoom, and select nodes to view their details
  3. Drill Down — Double-click any node to generate a detailed sub-graph
  4. Modify with Language — Use the prompt panel to add, remove, or change elements
  5. Render Scenes — Click "Render Scene" to generate an AI visualization
  6. Save Your Work — File → Save Project to preserve your entire hierarchy

How It Works

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  Text Prompt    │ ──▶ │   Gemini LLM    │ ──▶ │  Scene Graph    │
│  "A forest..."  │     │  (Structured)   │     │  (Nodes/Edges)  │
└─────────────────┘     └─────────────────┘     └─────────────────┘
                                                        │
                                                        ▼
                                               ┌─────────────────┐
                                               │  Interactive    │
                                               │  Canvas (PyQt)  │
                                               └─────────────────┘

SceneSynth uses structured output from Gemini to generate semantically meaningful graphs. Each node contains:

  • Name and description
  • Semantic type (location, landmark, character, etc.)
  • Spatial coordinates for meaningful layout
  • Expandability flag for drill-down capability

The hierarchy manager maintains parent-child relationships between graphs, enabling seamless navigation through nested worlds.

Project Structure

SceneSynth/
├── main.py                 # Entry point
├── config.py               # App configuration
├── core/
│   ├── graph/              # Scene graph data structures
│   ├── llm/                # AI providers and prompts
│   └── state/              # Application state management
└── ui/
    ├── main_window.py      # Main application window
    └── widgets/            # UI components

License

MIT License — Feel free to use, modify, and distribute.

Built with PyQt6 and Google Gemini