---
title: Vision Base
emoji: 👁️
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.18.0
python_version: '3.13'
app_file: app.py
pinned: false
tags:
  - track:backyard
  - sponsor:openbmb
  - achievement:tiny-titan
  - achievement:best-demo
  - build-small-hackathon
  - minicpm
  - openbmb
  - computer-vision
  - zerogpu
  - gradio
---

# 👁️ Vision Base

> **Four practical vision tools in one app — powered by a 1.3 B parameter model.**

Live demo: [build-small-hackathon/vision-base](https://huggingface.co/spaces/build-small-hackathon/vision-base)

---

## What it does

Most people don't carry a nutritionist, a chef, a technician, and a fortune teller in their pocket. Vision Base puts all four there — and does it with a model smaller than most podcast files.

| Tool | What you snap | What you get |
|---|---|---|
| 🔍 **Allergen Lens** | A food label | Every allergen flagged, dietary status, prep instructions, and a personal safety verdict |
| 🍽️ **Fridge Dinner** | Your open fridge | Three dinner ideas using only what's visible, plus a "use soon" warning list |
| 🔮 **Object Oracle** | Anything at all | A tarot-style mystical reading of that object's hidden essence |
| 🛠️ **What's That Error?** | An appliance error screen | The fault code decoded, root cause explained, and step-by-step fix — no manual needed |

---

## Why it matters

Every one of these is a real problem people reach for their phone to solve. Checking a label with an allergy is stressful. Staring at an error code at 10 PM is frustrating. Staring at a half-empty fridge every evening is a genuine daily dilemma. Vision Base solves all four in under 5 seconds — no API keys, no subscription, no data leaving your session.

---

## The model: MiniCPM-V 4.6

- **1.3 billion parameters** — qualifies for the Tiny Titan badge (≤ 4B)
- Built by **OpenBMB** — qualifies for the OpenBMB sponsor prize
- Runs on **ZeroGPU** — accessible to anyone, no GPU required from the user
- Handles JSON-structured extraction, multi-image input, and free-form creative generation — all in the same weights

The model loads once at startup, runs via `@spaces.GPU(duration=120)`, and offloads to CPU between calls to be a good ZeroGPU citizen.

---

## Tech stack

```
openbmb/MiniCPM-V-4.6   ← vision backbone (1.3B params, bfloat16)
Gradio 6.18              ← UI framework
Hugging Face ZeroGPU     ← A100 GPU on demand
spaces + transformers    ← inference wiring
```

Custom UI details: dark gradient hero header, per-tab color theming, shimmer skeleton loading animation, slide-up reveal on output, styled result cards — all in pure Gradio CSS and `gr.HTML`.

---

## Demo video

🎬 **[TODO: paste your demo video URL here — YouTube, Loom, or HF Video]**

*(Record a 60–90 second walkthrough: snap a food label, open fridge, mystery object, and appliance error — show all four tools in action)*

---

## Social post

📣 **[TODO: paste your social media post URL here — X/Twitter, LinkedIn, etc.]**

*(Mention `#BuildSmall`, link the Space, and show a screenshot or short clip of the app working)*

---

## Tracks & prizes this submission targets

| Category | Why we qualify |
|---|---|
| **Track: Backyard AI** | Four practical everyday tools solving real user problems |
| **OpenBMB Sponsor Prize** | Core model is `openbmb/MiniCPM-V-4.6` — OpenBMB's own multimodal vision model |
| **Tiny Titan Badge** | Entire app runs on 1.3B parameters — well under the 4B cap |
| **Best Demo Badge** | Polished four-in-one app with video and social storytelling |

---

## Running locally

```bash
git clone https://huggingface.co/spaces/build-small-hackathon/vision-base
cd vision-base
uv venv .venv && source .venv/bin/activate
uv pip install -r requirements.txt
python app.py
```

> Requires a CUDA GPU locally. On the Space, ZeroGPU handles GPU allocation automatically.