Visual production inside your product, without building it

The constraint

Platforms want to offer visual editing to their users. Getting there today means assembling it: one vendor for enhancement, another for staging, a third for object removal, each with its own API, its own pricing model and its own idea of what good looks like. The result is a feature that costs more than it should, behaves inconsistently across tools, and produces output your users can tell was generated.

Quality is the harder half. Generic image models were not trained on property interiors, so they get windows, verticals and empty rooms wrong in ways a buyer notices. Ours were trained on eight years of real listing production, millions of data points and before-and-after pairs from shoots we ran ourselves, which is why the output holds up on a listing rather than only in a demo.

Cost is the other half. Token and credit models make your margin move every time a user edits something twice, and nobody can price a feature they cannot forecast. A fixed price per image means you can put a number in your own pricing page and keep it.

What we built

The studio interface, running on the same API this page describes.

One panel, every model.

Enhancement runs down the right rail: perspective correction, Tonecraft, Auto privacy, Blue sky. Below it, Generative edit opens into four modes on the same call - Declutter to empty a room, Mess cleaner to remove clutter only, Indoor staging to furnish, Outdoor staging for garden and terrace. Delivery packages the output in the format the user publishes to, social formats included. Everything in one surface instead of four tools, and the full list of models and parameters is in the API documentation.

Automatic where it can be, manual where it matters.

Enhancement runs without being asked. Generative edit is user-initiated, because only the person selling the property knows what should go and what should stay.

The pipeline is visible.

Detect, generate, finish, package, with elapsed time and progress shown as it runs. Users tolerate a wait they can see; they abandon a spinner they can't.

Built for embedding, not just for us.

The same panel exposes an SDK snippet and an MCP connection, because we built it as a reference implementation of the API rather than as a closed product.

Compliance built in, not bolted on.

Auto privacy handles faces and plates, and AI-generated visuals carry the transparency label the EU AI Act requires. We had to build that anyway. You don't.

What we learned

Every client has their own idea of what good looks like.

One house style is too dark for one agency and too flat for the next. Features have to flex to each client’s taste, configured per account, or you spend your time arbitrating between them.

Per-account defaults matter more than per-image controls.

Users configure once and then stop thinking about it. Sliders on every image go unused.

Show the before.

Every edit needs a visible original to compare against, or users don't trust the output enough to publish it.