We will be undergoing planned maintenance on January 16th, 2026 at 1:00pm UTC. Please make sure to save your work.

Inspiration

Every photo in our camera rolls carries a feeling we didn’t finish —
a mood, a moment, a small unresolved story.

They say indie devs make games to process their lives.
But what if anyone could generate a tiny game
to finish a moment from theirs?

We wanted to build something that felt like an impossible kind of magic —
a tool that doesn’t just symbolize memory,
but turns real memories directly into playable worlds.

A room becomes a level.
A pet becomes a hero.
A door becomes a portal.
A feeling becomes a game loop you can complete.

That idea became GameMagical Project pic2game.


What It Does

GameMagical Project pic2game transforms any photo — a pet, a bedroom, a hallway, a messy desk — into a complete 8-bit platformer you can play instantly in your browser.

The Pipeline

1. Upload any image
A real iPhone-style photo, a sketch, a screenshot — anything meaningful.

2. Nanobanana → 8-bit spritesheet
The photo becomes a pixel-art spritesheet with cleanly sliced frames:

  • player character
  • platforms
  • moveable objects
  • tiles
  • environmental sprites

3. Gemini v3 interprets the memory and designs the game
Google Cloud Gemini v3 determines:

  • hero vs. background
  • platforms vs. decoration
  • emotional tone
  • level geometry
  • hazard placement
  • goal location

Gemini also generates:

  • a music-generation prompt for ElevenLabs
  • a final emotional epitaph for the completion screen (voiced via ElevenLabs TTS)

4. A retro engine constructs a winnable level
The result is a 10–30 second emotional micro-platformer where you:

  • jump across objects from your environment
  • navigate platforms formed from your memory
  • reach a glowing portal shaped by the scene
  • play through adaptive emotional music
  • complete the emotional loop

5. Your memory becomes a shareware-style link
Each generated game becomes its own page:

  • No downloads
  • No accounts
  • Just a shareable micro-world born from your life

A modern twist on classic shareware — but personal and infinite.


How We Built It

1. Nanobanana for spritesheets

We chose FAL Nanobanana for its stable pixel-frame consistency and clean black borders, making automated slicing far easier.

2. Gemini Flash 2.5 (Google Cloud) for semantic + emotional understanding

Gemini analyzes both the image and spritesheet to deduce:

  • which elements are playable
  • what forms platforms
  • where the hero should be
  • what emotional tone fits the memory
  • how to map mood → game logic

This gives each level the “intention” and emotional truth of the original moment.

3. Custom retro engine

A lightweight engine that:

  • animates sprites
  • handles collisions
  • performs auto-tiling
  • ensures solvability
  • generates a full level in under one second

4. Embedded sound + voiceover

ElevenLabs powers:

  • cinematic emotional narration
  • adaptive soft chiptune-style music influenced by Gemini’s emotional analysis

5. UI + Web Integration

A simple LAMP backend with a client-side canvas runtime ensures instant, zero-friction playability.


Challenges We Ran Into

  • Sprite inference from arbitrary photos
    Real images are noisy, emotional, and rarely orthographic.

  • Maintaining emotional truth
    Each memory needed to feel like itself, not like a generic generator.

  • Ensuring winnability
    Procedural levels often fail solvability, so we built a dedicated pass.

  • Avoiding uncanny results
    Pixel art became the perfect abstraction — compressing visual noise into emotional silhouettes.


Accomplishments We’re Proud Of

  • Turning real camera-roll photos into playable games in seconds
  • Creating genuinely emotional micro-experiences
  • Building a robust spritesheet → platformer pipeline across two AI models
  • Designing a memory-based level generator that always works
  • Producing cinematic pitch narration + adaptive music

What We Learned

  • Pixel art is the ideal emotional compression format
  • Every room can become a world
  • Every object can become a platform
  • Every unresolved moment can become a completed loop
  • People feel seen when their environment becomes playable
  • AI is most powerful when it amplifies meaning

What’s Next

  • A “Memory Library” for saving and sharing
  • Two-photo combined levels
  • A “Time Capsule” mode with embedded text
  • Exportable ROM-style microgames
  • Multiplayer emotional worlds
  • A surreal “Dream Mode”
  • Full mobile app
  • Daily “Memory Quest” prompts

Tools Used

  • FAL Nanobanana — Image → 8-bit spritesheet
  • Google Cloud Gemini v3 — Scene interpretation, emotional logic, & music/TTS generation
  • Custom Retro Engine — Level generation & gameplay
  • ElevenLabs — Cinematic narration & adaptive music
  • Freepik — Icon generation
  • GitHub Copilot Agent Mode — Development assistance Sonnet 4.5, Gemini v3, GPT5
  • LAMP Stack
  • Canvas / Web Runtime

--

Raindrop + Vultr Integration

While GameMagical focuses on emotional transformation at the user level, its backend pipeline is designed for repeatability, provenance, and iteration using LiquidMetal AI’s Raindrop platform, with Vultr providing scalable execution for generation workloads.

Raindrop: Orchestration & Provenance Layer

We use Raindrop Smart Components to turn each generated game into a reproducible artifact rather than a one-off output:

  • SmartBuckets
    Store all generated artifacts per run:

    • input photo
    • intermediate spritesheets
    • generated level data
    • final playable game bundle
      This allows us to version every generated game and revisit or compare runs.
  • SmartSQL
    Track generation jobs and metadata:

    • source image
    • inferred emotional profile
    • generation timestamps
    • output URLs
      This gives us a clear lineage from memory → game.
  • SmartMemory
    Persist reusable “generation profiles” (e.g. tone, difficulty, pacing) so similar memories can produce consistent gameplay results over time.

  • SmartInference
    Convert high-level intent (the uploaded photo + inferred emotional context) into a structured generation plan that downstream systems execute deterministically.

Raindrop allows us to treat emotional game generation as a pipeline, not a black box.


Vultr: Scalable Execution Layer

Vultr-hosted compute is used to execute the heavier parts of the pipeline:

  • image processing
  • spritesheet generation
  • procedural level construction
  • asset bundling for instant browser play

By separating orchestration (Raindrop) from execution (Vultr), we can scale generation workloads while keeping the system inspectable and debuggable — critical for a creative AI product meant to feel magical and reliable.


Why This Matters

AI-generated experiences are often ephemeral.
By integrating Raindrop and Vultr, GameMagical ensures that every generated game:

  • has a clear provenance
  • can be regenerated or evolved
  • can be reliably shared and replayed

This makes emotional, memory-driven game generation shippable, not just delightful.

Built With

Share this project:

Updates