The problem we are organised around
Generative models got good at a specific thing very quickly: producing one striking image from one description. That capability is real, it is genuinely new, and an enormous amount of tooling has been built to expose it — a box, a button, a result, a feed.
Then people tried to do work with it, and the shape of the problem changed. A campaign is not one image, it is forty that have to carry the same product. A scene is not one shot, it is twelve that have to hold the same character under the same light. A pitch is not a frame, it is a sequence a client will ask you to change on Thursday without it falling apart.
At that point the bottleneck stops being image quality and becomes continuity. And continuity is not a model capability you can wait for — it is a structural property of how the work is organised. If each frame is generated independently from a re-typed description, they will drift, no matter how good the model gets, because nothing in the system records that they were supposed to be related.
That is the observation the whole product is built on. Nova Studio does not try to be the tool with the best single output. It tries to be the tool where the relationships between outputs are the thing being edited.
How the product answers it
Four decisions that follow from taking continuity seriously.
A graph, not a feed
A project is a connected graph. Drag from one node to another and the output of the first becomes an input to the second — that is how a character reference reaches every shot that needs it, and how changing the reference changes all of them at once instead of none of them.
Providers behind one seam
Models sit behind adapters with a single interface. Adding one adds a node; it does not teach the canvas anything about that vendor, and it does not strand the work you made with the last one. The field moves quickly and your projects should not be hostage to which model was best the month you started.
Runs that are accounted for
Every generation is a job with a state — claimed, running, settled. A failure returns the credits it reserved. Nothing is charged twice, and a sweeper exists specifically to make sure a job that died mid-flight does not quietly keep your money.
History attached to the frame
Every result keeps the settings and inputs that produced it. "Which version did the client approve" and "what did we change between these two" are questions the document answers, not questions you answer from memory.
None of these are features in the marketing sense. They are the consequences of one decision — that the unit of work is a project rather than a generation — and each of them costs something. A graph is more to learn than a prompt box. Adapters are slower to add a model than calling its API directly. Accounted runs mean a settlement path that has to be correct under failure, which is most of the difficulty in the platform.
We think those costs are worth paying, because the alternative is a tool that demos well and cannot be used to deliver anything.
What we hold to
These are operating rules, not values on a wall. Each one has cost us something to keep.
- Say what is true
- Sample output is labelled as sample output, in the interface, every time. A provider that is not connected says so rather than failing quietly. Interface copy that promises a capability the build does not have is treated as a bug and fixed like one. This is why the pages on this site keep telling you what the product does not do.
- Your work is yours
- What you make is yours to use commercially. We do not train on your projects. The credits page tells you what a run costs before you press it, not after.
- Reversible over clever
- Undo covers structural edits. Deleting is confirmable. A decision that cannot be walked back gets a deliberately slower path — the extra click is the point, not an oversight.
- Degrade, do not break
- A missing environment variable turns off a capability and explains itself. It does not fail the build. A deployment must not be the first place a missing key is discovered.
- Built where it is used
- Half the team sits in the market where most of our early users are. The Chinese interface is written in Chinese rather than translated into it, and the two versions of a page are held to describing the same page by a test.
- Guests first
- The studio opens without an account. Guest runs return bundled samples — not as a crippled trial, but because a cookie anyone can mint by clicking a button is not an identity that can be billed or rate-limited, and pretending otherwise would be how the bill gets run up by somebody else.
The company
| Detail | |
|---|---|
| Product | Nova Studio — the creation surface people work in. |
| Platform | NovaOS — the layer underneath that routes runs to models, meters them and settles them. |
| Company | BOYU Labs, registered in Singapore. |
| Where | Singapore and Hangzhou as offices; the rest of the team remote. |
| Stage | Early. Small enough that everyone who joins is visible in the product within a month. |
| Contact | hello@novastudio.world reaches a person, not a queue. |
Two hubs is a deliberate choice rather than an accident of where people happened to live. The centre of gravity for creative AI adoption is not in one place: the model research and the enterprise buyers cluster in one set of markets, and an enormous share of the people actually making things every day are in another. A company that only sits in one of them ends up building for the half it can see.
What we are not
The fastest way to waste your time would be to let you find these out later.
Not a model lab
We do not train foundation models. We route to them, evaluate them and build the structure around them. When a model is bad at something, we will tell you rather than claim our wrapper fixed it.
Not a finishing tool
Grade, sound, conform and the last mile of brand colour happen in the tools you already have. Nova Studio is built to hand off cleanly, not to replace your pipeline.
Not a stock library
The samples bundled with guest mode illustrate the interface. They are labelled, and they are not a content offering.
Not finished
Long-form continuity is unsolved, by us and by everyone. The roadmap says what we are working on, the status page says what actually shipped, and neither one carries a date we cannot keep.
Questions we get
Do you train on my projects?
No. Your projects are not used as training data, and that is not a setting you have to find — it is the default and the only behaviour.
Can I use the output commercially?
Yes. What you make is yours. Model providers have their own terms about the outputs of their models, and where a specific provider imposes a restriction that matters we surface it rather than burying it.
What happens to my work if a model I used is removed?
The graph survives. Nodes hold prompts, settings, inputs and their previous results, so a project remains readable and re-runnable against a different provider. This is the main practical reason for the adapter layer.
Is there a free tier?
Guest mode opens the whole studio without an account and runs on bundled samples, so you can learn the tool without spending anything. Real generation costs money because it costs us money; the pricing page has the numbers.
Who is behind this?
BOYU Labs, a small team in Singapore and Hangzhou. If you want to talk to someone, hello@novastudio.world is read by people who build the product.