Comic Trip Agent

The Comic Trip travel agent is an application that analyzes your travel photos and transforms them into a vibrant pop-art comic strip experience. It not only stylizes your pictures but also intelligently identifies the location and provides nearby points of interest.

The application is built using the Agent Development Kit (ADK) for Java, which allows us to define specialized agents, equip them with specific tools, and compose them into complex, multi-step workflows.

The web backend of the application is built with the Quarkus framework. Quarkus receives the uploaded pictures and sends them in parallel to the Comic Trip agent for processing.

How It Works

The core intelligence and orchestration of the application are handled in src/main/java/comictrip/ComicTripAnalyzer.java.

Agent Structure and Orchestration

The image analysis and generation process is driven by multiple specialized AI agents, using different Gemini models, orchestrated both sequentially and in parallel:

1. picture_analyzer_agent (LlmAgent)

   • Model: gemini-3-flash-preview
   • Task: Analyzes the uploaded picture to extract a detailed description of its contents and identifies the probable location where it was taken.
   • Output: Returns a structured JSON payload containing the description and location.

2. comic_illustrator_agent (LlmAgent)

   • Model: gemini-3.1-flash-image-preview (aka 🍌 Nano Banana 2)
   • Task: Takes the extracted description and location, and transforms the original photography into a pop-art comic panel featuring thick black outlines, color drops, halftone textures, and a vintage palette.
   • Callbacks: Utilizes an afterModelCallback to intercept the multimodal LLM response, extract the generated image bytes, save the file locally (for debugging purpose), and persist it as an artifact.

3. points_of_interest_agent (LlmAgent)

   • Model: gemini-2.5-flash
   • Tools: GoogleMapsTool
   • Task: Uses the location identified by the first agent to list points of interest (POI) within a kilometer radius, providing a name and description for each location using Google Maps data.

4. poi_and_commic_flow (ParallelAgent)

   • Task: Orchestrates the comicCreatorAgent and poiGoogleMapsAgent to run simultaneously. Since both agents depend on the output of the initial analysis but are independent of each other, running them in parallel significantly improves the overall processing time.

5. main_flow (SequentialAgent)

   • Task: The root flow that strictly orders the execution: it first runs the comic_trip_agent to gather the necessary context, and subsequently triggers the poi_and_commic_flow.

ADK Core Concepts

• App and Plugins: The entire agent graph (mainFlow) is wrapped inside an App container named comic_trip_app. It utilizes a LoggingPlugin for execution observability and debugging.
• Runner and Session: An InMemoryRunner is configured to execute the App. It manages the state and context across different agent steps using an InMemorySessionService. The context is organized and isolated by a SessionKey based on the specific app, user, and trip ID.

Storage Architecture

The application uses two primary storage mechanisms on Google Cloud to handle structured data and media artifacts:

1. Google Cloud Storage (GCS)

   • Usage: Stores the generated pop-art comic image files.
   • Implementation: Fully integrated into the ADK Runner configuration using GcsArtifactService. It handles the upload of generated artifacts to the comic-trip-picture-bucket. During the comicCreatorAgent's callback phase, the image is passed to callbackContext.saveArtifact(...) to safely upload the generated media to GCS.
   • Directory Structure: The ADK structures artifacts hierarchically based on the app name, user, session (trip ID), and artifact name. The resulting path in the bucket looks like this:

     gs://comic-trip-picture-bucket/comic_trip_app/comic_trip_user/{tripId}/{imageId}.png/0
     

2. Firestore Database

   • Usage: Stores the structured metadata of the generated trips.
   • Implementation: Managed by TripService.java. It persists trip data to the comic-trip Firestore database inside the trips collection. The stored documents include the trip's title and an array of picture objects containing their textual descriptions, recognized locations, generated image URLs (pointing to the GCS bucket), and the nearby points of interest gathered by the agents.
   • Data Structure: Each document in the trips collection represents a single trip (identified by {tripId}) and is structured as follows:

     {
       "tripId": "x7bY9p",
       "title": "GALATA GOLD: ISTANBUL RISING",
       "pictures": [
         {
           "tripId": "x7bY9p",
           "image": {
             "name": "https://storage.googleapis.com/comic-trip-picture-bucket/.../{imageId}.png/0",
             "mimeType": "image/png"
           },
           "details": {
             "description": "A view of the iconic Galata Tower...",
             "location": "Galata Tower, Istanbul, Turkey"
           },
           "pointsOfInterest": "* Torre Galata...\n* Kamondo Stairs..."
         }
       ]
     }

Deployment

For instructions on how to build and deploy the Comic Travel Agent application to Google Cloud Run from source, please refer to the Deployment Guide.

---

License

This project is licensed under the Apache License, Version 2.0.

Disclaimer

This is not an official Google project.