openegiz Al-Farabi KazNU
AR MENU · BR24992975
Research project · BR24992975

Dish photo → 3D →
augmented reality

A 3D model generator for AR menus: a customer photographs a dish and a few minutes later places it on their own table in augmented reality — right in the browser, with no apps to install.

One photo in · a model and a QR code out · all on your own hardware

Digital Egiz Lab Al-Farabi KazNU OpenEgiz · open-source
Problem

3D content for AR is expensive and slow

🎨

Manual modeling

A 3D artist or photogrammetry for every dish — too slow and expensive for a menu that changes all the time.

☁️

Cloud services

Pay per generation, network latency, and dependence on someone else's API that may get pricier or shut down.

🔒

Data privacy

Product photos have to be uploaded to someone else's infrastructure — not everyone is comfortable with that.

Solution

The whole pipeline on a single device

The customer uploads a single photo, the system removes the background, generates a textured 3D model, embeds the real size and publishes the result. In return — a QR code and a link to view it in AR.

  • →No apps to install — opens in the phone's browser
  • →Data never leaves the premises
  • →A one-time hardware cost instead of paying per generation
Generation stages
Five stages of processing one request
How it works

From a photo to an AR model

01

Background removal

The BiRefNet neural network cuts out the dish and centers it.

seconds
02

3D generation

TRELLIS.2: structure → shape → texture.

the main stage
03

Real size

The dish's physical size is embedded into the model.

milliseconds
04

GLB export

Remeshing, simplification, baking of 2K PBR textures.

tens of seconds
05

Publishing

QR + preview, model upload to the website.

seconds

The finished files are published to the site's media library — the links and QR codes are permanent and keep working even when the generator is switched off.

Sample result

Pizza: photo → model → QR

Source photo of the pizza
Source photo
3D model: color and normals
3D model — color and normals
QR code to the AR viewer
QR → view in AR

Scan the QR code and a 3D model opens with an AR button. Put the pizza on the table at life size right through the phone camera.

Architecture

How the system is built

System architecture
FastAPI → Redis → Celery + TRELLIS.2 → WordPress · all on a single edge device
Hardware

A compact AI supercomputer on the desk

All generation runs on a single device NVIDIA DGX Spark (GB10 Grace Blackwell chip) — the size of a small box, with unified CPU/GPU memory.

128 GB of unified memory Blackwell GPU 20-core ARM CPU Edge · no cloud
128 GBunified CPU/GPU memory — the model and the processing live side by side
100%of the data stays in place — photos never go to the cloud
Technology

What it is built on

🧠

3D generation

TRELLIS.2 (Microsoft) — flow-matching image-to-3D. The background is removed by BiRefNet.

⚙️

Server

FastAPI + queue Celery + Redis: photo intake, background generation, real-time status.

📱

Viewing and AR

glTF/GLB, a component by <model-viewer> and WebXR — 3D and AR right in the browser.

Open standards · open-source application layer

Features

What makes the solution practical

  • →Life size in AR — the dish appears exactly as big as it is in real life
  • →No apps — the QR code opens the model in a regular browser
  • →Permanent links — models live on the website independently of the generator
  • →Fault tolerance — a job is not lost even during a power outage
  • →Model catalog — builds itself automatically on the website
  • →Two modes — quick preview and high quality
Results

Already running in production

~5 minfast mode (512) — previews and bulk menu upload
~20 minhigh quality (1024) — detailed “showcase” dishes
1 photoall that is needed from the user

The system is deployed at digitalegiz.kz: a customer uploads a photo through the generator page and gets a ready 3D model with a QR code and an AR button.

Applications

Not just food

🍽️

Restaurants and cafés

AR menu: the guest sees the dish and its size before ordering.

🛍️

Retail and marketplaces

3D product cards from a single photo.

🏛️

Museums and education

Exhibits and teaching models in augmented reality.

The architecture is the same — only the storefront and the meaning of “size” change.

Project BR24992975

Try your own dish in AR

One photo — and you have a 3D model with a QR code that can be seen in augmented reality right in the phone's browser.

Open digitalegiz.kz → Source code on GitHub

Digital Egiz Lab · Al-Farabi KazNU · openegiz

Slide overview · BR24992975

←→ navigation O overview F screen
01 / 12