AI is a node. Not a panel.
Generation sits inside the graph, wired to the thing it feeds. That one architectural choice is the difference between art direction and a slot machine.
Everyone has the models. Almost nobody has them in production.
That's the actual problem, and more models won't fix it. A model produces one output when you ask it nicely. Production needs the same output again — on brand, in every format and language, on a deadline, approvable and deliverable.
AI as a panel
A box off to the side. You type, it returns something, you download it and carry it into the real work by hand. Press it again and you get something else. Nothing upstream feeds it; nothing downstream depends on it.
AI as a node
An operator with typed inputs and outputs, sitting in the graph between the things it connects. It takes your geometry, your type, your brand colours, your data — and what it makes flows onward into the rest of the work.
Five things you get from the wire.
It's art-directable, because it has inputs
A generation node is fed by real things — a model you built, a layout you composed, a palette from your brand system. You're directing it with your own work rather than describing your work in a sentence and hoping.
It's repeatable, because it's part of a graph
The same graph run twice does the same thing. Change one parameter upstream and everything downstream — including the generated parts — updates coherently. That's what makes a set of two hundred variants possible at all.
It composes with everything else
Generation isn't a separate stage you round-trip through. It sits next to geometry, materials, signals, data, state and the document — so a generated texture can drive a real material on real 3D that ends up in a real page.
It can be operated by someone who isn't you
Because the AI step is wired into a system with defined inputs, an author can expose just the parts that should vary. The person running it gets a form. They can't wander into a prompt and take the brand somewhere it shouldn't go.
It leaves a record of what you decided
A graph is a written-down set of decisions — this drives that, this varies, this never does. Documented human creative direction is the strongest authorship position AI-assisted work can hold. US law protects human-authored expression (D.C. Cir. 2025; cert denied 2026) — this isn't a guarantee of ownership, but it's a claim prompt-only output can't make at all.
MLOP is a family, like everything else.
Inference isn't a special case bolted to the side of the engine. It's one of the ten operator families, with the same colour coding, the same strictly typed wires and the same per-frame cook as geometry or textures.
Per-model generators
Dedicated nodes per model rather than one vague "generate" box — so each exposes the controls that model actually has, instead of a lowest-common-denominator prompt field.
Editing and conditioning
Depth, edges, references and masks are inputs like any other. Condition a generation on geometry you built rather than describing that geometry in words.
Live diffusion
Generation that cooks per frame, so it can react to input, signals and live data — the difference between an AI image and an AI system.
It builds the graph with you — and can't invent nodes.
There's a second kind of AI here, and it's the one that does the wiring. Describe what you want and the assistant composes it: adding nodes, connecting them, setting parameters, shaping behaviour.
A closed registry
The model can only use node types the engine actually has. It cannot invent a node, a callback or a global that doesn't exist.
That's the difference between an assistant that helps and one that produces confident nonsense you then spend an hour debugging. If it suggests something, that something is real.
It speaks MCP
The same tools the assistant uses are exposed over the Model Context Protocol — so any agent that speaks it can drive the editor, build graphs and read outputs.
Brief in one, work in another, run a graph by name from either. The canvas and the agent are the same instrument.
We orchestrate models. We don't train them.
Training foundation models is a capital race that obsoletes itself every few months. We'd rather be the best place to use whatever exists.
Model-agnostic by construction
Providers are transport. Models are swappable without re-architecting your graph, so when a better one ships you point at it — and when one degrades or changes terms, you aren't stranded.
Open weights, so the cost curve works for you
Open models get cheaper and better without us doing anything. We ride that curve rather than paying to push it, which is also why we can afford a free tier that isn't crippled.
Runs where you choose
Hosted by default, so nobody needs a GPU to start. On your own hardware when the work can't leave the building — the case that matters for client IP under NDA. See motion design studios for how that's set up.
No lock-in on the way out
A scene compiles to an ordinary web page and a workflow hands you ordinary files. The graph is yours, and so is what it makes.
What people ask about the AI.
Which models can I use?
Any open model we can wire, and the next one. We deliberately don't lead with a logo wall — the list dates in a season, and the point is the mechanism rather than the inventory. The current set is visible in the editor.
Do I need my own GPU?
No. Free and Individual plans run hosted inference, so you can start on a laptop. Running on your own hardware is an option for teams whose client work can't go to a public service, not a requirement for everyone else.
Do you train on my work?
Hosted tiers may use graph structure to improve the assistant, and that's disclosed in the terms rather than buried. A contractual no-training guarantee is available for enterprise. Local and self-hosted inference never phones home, by construction.
Can I use what it generates commercially?
Yes, on a paid plan — Free is non-commercial. Individual models carry their own terms, which is one reason we favour open weights. For the authorship question, see the fifth point above: a human-authored graph is documented creative direction, which is the strongest position AI-assisted work can hold.
Is the assistant going to build my whole project?
It'll do a lot of the wiring, and it's genuinely fast at the tedious parts. It won't decide what's good. A graph is a record of decisions, and the decisions are still yours to make.
What if I don't want to use the AI at all?
Then don't. MLOP is one family of ten, and plenty of work here uses none of it. The engine is a procedural engine first.
Wire it up and see the difference.
Free to start, nothing to install. The gap between a panel and a node is obvious about ninety seconds after you've used one.
Working with a team?
Tell us where to reach you and we'll set up a walkthrough.
No spam. A real reply from a human.