Diffuji

A diffusion-powered instant camera that turns half-real, half-dreamed moments into physical prints.

Built at TreeHacks 2026.

Inspiration

The world needs a bit more silliness. Instant cameras are fun, and there's something special about memorializing a moment through physical media. But what if these moments were part dream, part hallucination of reality?

Diffuji captures a photo and reimagines it through image-to-image diffusion models, then prints the result on thermal sticker paper — no ink required.

What It Does

Take a picture and choose from a range of filters. These aren't your typical color grades — they can completely reimagine the context the photo was taken in:

Mode	Description
ghibli	Studio Ghibli anime style
greek	Classical Greek marble statues
duck	Replace all people with ducks
gpu	PS2-era low-poly 3D graphics
thinker	Rodin's "The Thinker" sculpture
1846	Time-travel to 1846
1922	Time-travel to 1922
1955	Time-travel to 1955
1984	Time-travel to 1984
1999	Time-travel to 1999
business_card	Extract info into a business card layout
pricing	Estimate retail prices of visible items

The pricing mode was inspired by a teammate's experience working at a thrift store — snap a pic of an object and use Perplexity to search the internet for competitive prices, then print a reference-backed price tag.

How We Built It

Hardware

Component	Purpose
Raspberry Pi Zero 2W	Low power draw, WiFi, 512 MB RAM, $15
Arducam camera module	Image capture
TTL thermal printer	Inkless printing onto sticker paper
Rotary encoder	Switching between modes
Push button	Shutter
Tactile switch	Power
I2C OLED display	Displaying modes and settings
2x 18650 batteries + 3A 5V UPS	Power supply with voltage regulation
1000 µF capacitor	Guards against current spikes from the printer

The shell was designed in Fusion 360 and 3D-printed at the TreeHacks makerspace. Components were soldered together and mounted with hot glue and screws.

Software

On-Device (camera/)

A Python script running on the Pi that:

• Manages hardware inputs (rotary encoder, shutter button, power switch)
• Displays animations and state on the OLED screen
• Captures images and either processes them locally or sends them to the cloud
• Dithers images for thermal printing using ordered Bayer dithering with custom gamma and contrast adjustments
• Prints the final result

Cloud API (dispoapi/)

A FastAPI server hosted on Railway that serves as the camera's brain:

• Provides the camera with its available modes
• Routes images to AI providers for diffusion-based transformation
• Returns processed images as base64

AI Providers:

Provider	Description
OpenAI	GPT-image-1 via Responses API
Gemini	Google Gemini 2.5 Flash image generation
Modal	Flux Kontext on serverless H100 GPU
Perplexity	Sonar API for text-based search and pricing

Architecture

┌─────────────────────┐         ┌──────────────┐         ┌─────────────┐
│  Raspberry Pi Zero  │  POST   │   DispoAPI   │  route  │ AI Provider │
│  - Arducam          │ ──────> │  (Railway)   │ ──────> │ OpenAI /    │
│  - OLED display     │ <────── │              │ <────── │ Gemini /    │
│  - Rotary encoder   │  base64 │              │  image  │ Modal /     │
│  - Thermal printer  │         └──────────────┘         │ Perplexity  │
└─────────────────────┘                                  └─────────────┘
         │
         v
   ┌───────────┐
   │  Thermal  │
   │  Sticker  │
   │  Print    │
   └───────────┘

Challenges We Ran Into

• Paper jamming — the sticker paper would bunch up inside the printer. We found that peeling back some of the sticker backing before feeding it helped.
• Inconsistent style transfer — prompting the diffusion model to consistently apply a style was tricky. We also discovered that feeding images upside-down produces terrible results (too far out of distribution).
• On-device diffusion — running diffusion on a device with only 512 MB of RAM is painfully slow, pushing us toward the cloud API approach.

Accomplishments We're Proud Of

We built a sleek, functional product that was genuinely fun to use. We hope everyone at the hackathon enjoyed getting custom sticker prints!

What We Learned

• Dithering algorithms and how to make images look aesthetic on thermal paper
• Inference strategies for extremely low-RAM systems
• Hardware integration with the Raspberry Pi ecosystem (I2C, TTL serial, GPIO)
• Prompting diffusion models for consistent style transfer

What's Next

On-device diffusion. With emerging lightweight diffusion architectures and quantization techniques, the dream is to run the full pipeline on the Pi itself — no cloud required.

Project Structure

dispo/
├── camera/             # On-device Python scripts (Pi Zero 2W)
│   └── main1.py        # Hardware control, capture, dither, print
├── dispoapi/           # Cloud API server
│   ├── main.py         # FastAPI app, endpoints, AI provider routing
│   ├── modal_app.py    # Modal deployment — Flux Kontext on H100
│   ├── prompts.py      # Mode-to-prompt mapping
│   ├── test_api.py     # Test script
│   └── pyproject.toml  # Dependencies
└── README.md

Getting Started

See dispoapi/README.md for API setup, configuration, and usage instructions.