Print-ready 3D models
from photos and words

Polyzon Maker runs on your own computer: 16 open-weight AI models, CAD, OpenSCAD and 101 ready-made tools. Free, no account, no subscription.

  • Free
  • Open source, MIT
  • Windows and Linux, Mac soon
  • English and Russian
The library: models Polyzon Maker made, with pictures and dates

What comes out

Real models from Polyzon Maker. Each card shows the engine and the time on an RTX 5070 Ti, 23 GB RAM, Windows 11.

AI models that
work today

Each one installs with a single button inside the app and then runs offline. This list comes from the catalog of the app itself, version 0.3.2.

Photo to 3D7

One photo or picture in, a 3D model out.

  • TRELLIS.2 Microsoft Research, trellis.cpp by Piotr Wilkin

    The most detail, on almost any graphics card.

    • Tested
    • 83 s
    • from 4 GB of video memory
    • NVIDIA, AMD, Intel
    • MIT
  • Pixal3D Tencent ARC, trellis.cpp by Piotr Wilkin, GGUF by vegax87

    Keeps closer to the photo, a fine-tune of TRELLIS.2.

    • Experimental
    • from 4 GB of video memory
    • NVIDIA, AMD, Intel
    • MIT
  • Clean, closed shapes in half a minute.

    • Tested
    • 32 s
    • from 8 GB of video memory
    • NVIDIA
    • Tencent Hunyuan Community
    • Not for the EU, the UK or South Korea
  • Hunyuan3D-2 Tencent

    The full model: more detail than Mini, wants more video memory.

    • Experimental
    • from 8 GB of video memory
    • NVIDIA
    • Tencent Hunyuan Community
    • Not for the EU, the UK or South Korea
  • Hunyuan3D-2.1 Tencent

    The newest open Hunyuan shape model: cleaner forms, slower than the turbo ones.

    • Experimental
    • from 10 GB of video memory
    • NVIDIA
    • Tencent Hunyuan 3D 2.1 Community
    • Not for the EU, the UK or South Korea
  • TripoSG VAST AI Research

    Fine shape detail on a good NVIDIA card.

    • Tested
    • 45 s
    • from 7 GB of video memory
    • NVIDIA
    • MIT
  • TripoSR Stability AI, Tripo AI

    A quick draft on any computer, no graphics card needed.

    • Tested
    • 19 s
    • from 3 GB of video memory
    • NVIDIA
    • Works without a graphics card
    • MIT

Several photos to 3D4

Shots from several sides in, the real shape of the object out.

  • MapAnything Meta and Carnegie Mellon University

    Several photos from any sides: rebuilds the real shape, in real size.

    • Experimental
    • from 8 GB of video memory
    • NVIDIA
    • Apache-2.0
  • Pi3 Yifan Wang and others

    Several photos from any sides, no order needed. Weights are for non-commercial use.

    • Experimental
    • from 8 GB of video memory
    • NVIDIA
    • CC BY-NC 4.0
    • Non-commercial use only
  • VGGT Meta and University of Oxford

    Several photos from any sides: a popular, fast rebuild of the shape. For non-commercial use.

    • Experimental
    • from 8 GB of video memory
    • NVIDIA
    • CC BY-NC 4.0
    • Non-commercial use only
  • Hunyuan3D-2mv Tencent

    Three photos in turn: front, left, back.

    • Experimental
    • from 8 GB of video memory
    • NVIDIA
    • Tencent Hunyuan Community
    • Not for the EU, the UK or South Korea

Picture from words2

They draw a picture from your words, and the picture becomes 3D.

  • Z-Image Turbo Tongyi-MAI

    Makes a picture from your words, then a model from the picture.

    • Tested
    • 20 s
    • from 4 GB of video memory
    • NVIDIA, AMD, Intel
    • Apache-2.0
  • FLUX.2 klein 4B Black Forest Labs

    A faster picture from your words.

    • Tested
    • 9 s
    • from 4 GB of video memory
    • NVIDIA, AMD, Intel
    • Apache-2.0

Assistant3

They write CAD and OpenSCAD code, look at the result and fix it.

  • Gemma 4 E4B Google, GGUF by Unsloth

    Small, sees pictures. For 6 GB of video memory.

    • Experimental
    • from 6 GB of video memory
    • NVIDIA, AMD, Intel
    • Apache-2.0
  • Gemma 4 12B Google, GGUF by Unsloth

    Writes better code, sees pictures. For 8 to 12 GB of video memory.

    • Experimental
    • from 8 GB of video memory
    • NVIDIA, AMD, Intel
    • Apache-2.0
  • Qwen3.6 27B Alibaba, GGUF by Unsloth

    The strongest here, sees pictures. For 16 GB of video memory and more.

    • Experimental
    • from 16 GB of video memory
    • NVIDIA, AMD, Intel
    • Apache-2.0
  • Any GGUF model from Hugging FaceFor the assistant: search right in the app and pick a size that fits your card.
  • New models through pluginsA plugin can add an AI model the app does not run yet.

Times measured on an RTX 5070 Ti on Windows 11 at the default detail. «Tested» means the model has made models on Windows. On Linux every model is still experimental.

Everything from an idea
to a file you can print

Model from a photo

Give it a photo or a picture and get a 3D model back. You choose: a quick draft, a clean closed shape, or the most detail.

  • Hunyuan3D-2 mini Turbo32 s
  • TripoSG45 s
  • TRELLIS.2, 51283 s
  • TripoSR on the processor35 s

Times on an RTX 5070 Ti, 23 GB RAM, Windows 11.

Dragon figurine from the front, a model made from a text prompt

Model from words

Describe what you want. A local model draws a picture, and the picture becomes 3D.

Straight to the printer

Checks the surface, closes holes, sets the size in millimetres and saves STL or 3MF. Your slicer opens with one button.

Armoured goblin with a poleaxe, the model after print preparation
80 × 52 × 42 mm

CAD for parts and scenes

Build parts and whole scenes of several models in code or with the mouse. Handles, colours, 17 examples, STL and STEP.

CAD: a helicopter built in code, its list of parts and the tools

OpenSCAD

An editor with highlighting, a live view, parameter sliders, templates and your own programs.

OpenSCAD: the code of a mug, the live view and parameter sliders

An assistant that does the work

Say it in words or show a photo. A local model picks the generator, writes a plan, builds the model in CAD or OpenSCAD, looks at the result and fixes it.

For AI agents

The «AI agent» button gives the ready command for Codex, Cursor, Gemini CLI, Qwen Code, Kimi and others. The agent builds in CAD and OpenSCAD and sees what came out.

Plugins

A plugin adds what is not there yet. It runs apart from the app, with only the rights you gave it.

Yours stays yours

Photos, descriptions and models never leave your computer.

  • Anonymous reports only if you agree
  • Updates are signed
  • A hidden AI mark in every generated file

No cloud and no subscription. Your photo stays on your computer, and what comes out is a file you can print.

Three steps to a print

  1. Install and check

    Install the app. The first run measures your computer in 15 seconds and shows which models will work and how long they take.

    First run: the app checks the processor, memory, disk and graphics card
  2. Make a model

    Download a model that fits, once: it takes several gigabytes. Then give it a photo or a few words.

    A new model: a description in words, and the AI models for 3D and for the picture
  3. Get it ready

    Set the height in millimetres, check the surface, and save STL or 3MF, or open the model in your slicer.

    A mug from OpenSCAD before printing: size, check, STL, 3MF and the slicer

101 tools that need
no graphics card

Ready-made models you adjust to size, and utilities for models and printing. All of it works at once, with nothing to download.

Download Polyzon Maker

Free, no sign-up.

The models for generation download inside the app and take several gigabytes. Polyzon Maker does not run on phones, you need a computer.

Every version with its checksums, and the source code, is on GitHub.

Questions

Is it really free?

Yes. Polyzon Maker is free, the code is open under the MIT license, and you do not need an account. Each AI model keeps its own license, which the app shows in the list of models.

Do I need a graphics card?

For generation, it helps a lot. TRELLIS.2 starts at 4 GB of video memory and 8 GB of RAM and runs on NVIDIA, AMD and Intel; Hunyuan3D-2 mini Turbo asks for 8 GB of video memory. Without a card you can use TripoSR, a quick draft model: about 35 seconds on the six-core Intel processor in our test. The other models take over an hour on a processor. The tools, CAD and OpenSCAD need no graphics card.

How much disk space does it take?

The Windows installer is 12 MB. AI models download separately, only the ones you choose: 5 to 18 GB per model, including everything it needs to run.

How long does one model take?

It depends on your computer. On an RTX 5070 Ti with 23 GB of RAM on Windows 11 we measured 32 seconds for Hunyuan3D-2 mini Turbo, 45 seconds for TripoSG, 83 seconds for TRELLIS.2 at 512, and 9 to 20 seconds for a picture from words. Yours will differ: the check on first run gives you an estimate.

Do my photos go anywhere?

No. Photos, descriptions and models stay on your computer. The app goes online for downloads you start, for an update check at startup, and for anonymous reports if you agreed to them. Those reports hold no pictures, descriptions, file names or folders.

Can I print a model right away?

After preparing it. Generation rarely gives one solid surface, so the app checks the model, repairs it and tells you the result: closed, or not yet. Wall thickness and self-intersections are not checked yet, so look at the model in your slicer before printing.

What files do I get?

STL and 3MF for printing and the original as GLB, plus STEP from CAD. The app finds the slicers you have, such as PrusaSlicer, OrcaSlicer, Bambu Studio or Cura, and opens the model in them. It does not send G-code and never starts a print by itself.

Does it run on Mac and Linux?

There are builds for Windows and Linux now, and the Mac version is on its way. All our measurements were made on Windows, so AI models on Linux are still marked experimental.

Are there any limits by country?

The licenses of the Hunyuan3D models (2 mini Turbo, 2, 2.1 and 2mv) do not apply in the EU, the UK and South Korea, and the app says so before installing them. The Pi3 and VGGT weights are for non-commercial use only. The other models have no such limits.

How do I connect an AI agent?

Press «AI agent» in CAD, OpenSCAD or the assistant: it has the ready command or a button for Codex, Cursor, Gemini CLI, Qwen Code, Kimi and others, and shows who connected. The API listens on 127.0.0.1 only and needs a key, so nothing on the network can reach it.

Can I extend it?

Yes, with plugins. A plugin is a .zip that can add a page, panels and commands, a theme or a new AI model. Before installing you see its permissions and accept all of them or none, and a plugin’s AI model runs in a sandbox of its own on Windows.

See what your
computer can make

The app is free, and the check takes about 15 seconds.