TURRRBO — a stylegan instrument

✦ what this is

Not a prompt box. A place to dig.

This is not a text-to-image tool. You are not typing prompts and getting illustrations. You are navigating a mathematical space that a neural network learned by looking at thousands of real images — and pulling out what lives there.

The results are weirder. More specific. More yours.

✦ what's inside

5 pretrained models, each a different visual territory

Synthetic Portraits

70,000 human faces collapsed and remixed into synthetic identities. Coherent at mid settings. Dissolves into anatomy-horror at the extremes.

Museum Wreckage

Trained on portrait paintings from the Metropolitan Museum of Art. Painterly surfaces, visible brushwork, art-historical drift.

Haunted Schematics

LSUN Churches dataset. Architecture with recombinant structural logic — arches and spires merged in ways that feel almost recognizable.

Animal Forms

AFHQ cats. Feline anatomy that abstracts into fur-texture fields under pressure.

Automotive Drift

LSUN Cars. Chrome logic, strange reflections, body lines that almost make sense.

✦ controls that actually mean something

Every parameter maps to a real place in the model

Seed

Each integer is a different identity, a different location in the model's learned world.

Truncation ψ

How far from the average to go. Low = safe. High = strange.

Split truncation

Push pose and texture in different directions simultaneously.

Noise strength

Flood the surface with stochastic detail.

Style mixing

Let two seeds share one image — coarse structure from one, fine texture from another.

✦ 9 style routes

Named artistic modes, not sliders you have to guess at

Ghost Contour

Thermal Hallucination

Overexposed Memory

Broken Chroma

TV Wreckage

Cartoon Corrosion

Structure Transplant

Fine Transplant

✎ CLIP text guidance (optional)

Type a description and the latent vector optimizes toward it. Not a prompt engine — a steering mechanism. Effective prompts describe visual qualities, not subjects.

✦ 8 built-in templates

Starting points that demonstrate what the models can do

Raw Output

Average Collapse

Full Chaos

Structure Swap

Split Truncation

Noise Flood

Ghost

✦ fit check

Who this is for — and who it very much isn't

this is for you if —

You're an artist bored of the same AI aesthetic.
You're a designer who wants source material that doesn't look like everyone else's source material.
You're a photographer or collage artist looking for synthetic textures, faces, and forms to work with.
You're an experimenter who wants to understand how these models actually behave — not just what they produce when you ask nicely.
You find glitches, artifacts, and model failures more interesting than clean photorealistic output.

this is not for you if —

You want photorealistic images from text prompts — use Midjourney or Stable Diffusion instead.
You want an image that could have come from anywhere. TURRRBO is for images that feel like they came from somewhere specific — a model with character, trained on a particular dataset, pushed into territory it was never meant to go.

✦ technical requirements

What it needs to run

macOS 12 Monterey or later (Apple Silicon M1/M2/M3 recommended)

Memory

8 GB RAM minimum, 16 GB recommended for the 1024px models

Storage

~4GB base version, unlimited with custom modules

Setup

No Python, no terminal, no setup

Network

No internet connection required after install

✦ the fine print, up front

A note on licensing

⚠ non-commercial, on purpose

TURRRBO uses pretrained model weights from NVIDIA Research. These weights are licensed for non-commercial use only. That means TURRRBO is free — and must stay free.

This is not a limitation. It's the whole point.

✦ no funding round required ✦

Go dig around in the latent space.

Free. Offline. Yours to keep, for as long as your machine will run it — which, unlike most AI companies, is not a trick question.

Get it free on Gumroad →!!