Case study

Twelve Lumens: building AI virtual staging that respects the photograph

Twelve Lumens turns listing photos into furnished, decluttered rooms in one consistent house style, for estate agents, photographers and homeowners. We built it in 2026 on the same foundation as our other products, with an image pipeline designed around one rule: the room stays the room.

Product
AI virtual staging SaaS
Launched
2026
Stack
Rails 8.1, Hotwire, OpenAI
Languages
5
Status
Live
The Twelve Lumens editor comparing an original living room photo with a Scandinavian-style render, with version history and an edit prompt
The challenge

The challenge: staging is slow, retouching is inconsistent

Physical staging is expensive and slow. Outsourced retouching is hard to keep consistent from one room to the next. Agents want photos that sell a home without misrepresenting it.

We wanted one property workspace where you upload the photos, declutter them, apply a house style, mark the finals and share a page with buyers. With billing that counts images instead of mysterious credits, and a clear label on everything AI has changed.

Our approach

Our approach

Every AI task in Twelve Lumens is a Generation record with its own class, running as a background job with live progress in the interface. Before anything is rendered, a vision model analyses the photo. The image model then gets a prompt assembled in a fixed order, so results are predictable and the architecture stays as photographed.

  • A vision pass on every upload: room type and number, how furnished it is, collages and existing AI labels
  • Prompts built from fixed rules first and the user's own guidance last, applied only when it doesn't conflict
  • Originals always kept; every render is a new version on a timeline
  • A disclosure label and IPTC metadata on every export

From a full shoot to sorted rooms

Agents upload the whole shoot at once, in JPEG, PNG, WebP or HEIC up to 20 MB. HEIC is converted, EXIF data is stripped and the files go into private storage. A vision model then sorts each photo into one of 21 room types and keeps the numbering consistent across upload batches, so the second bedroom stays the second bedroom.

Photos are shrunk to 1024 pixels before analysis to keep the cost per photo low. The same pass writes a description of every room in five languages, which later feeds the page buyers see. Dutch addresses are enriched with data from the public BAG and Kadaster registries, with the source of every field recorded.

A Twelve Lumens property with its photos sorted into bathroom and bedrooms, each with a description written by the vision model
Uploaded photos sorted by room, each with a description written by the vision model.

Staging that respects the photograph

Rendering runs on OpenAI's gpt-image-2 through its image edits endpoint, at high quality, with PNG masters in 1K or 2K. The prompt builder stacks its sections in a fixed order: wording for the room type, a hard rule that the architecture stays as photographed, a rule that a kitchen stays a kitchen, rules per room, the style brief, sliders and locks, and only then the user's own guidance.

Decluttering empties a lived-in room while walls, windows, built-ins and the camera angle stay where they are. On exteriors it removes cars, bins and clutter but leaves the neighbouring houses alone. We build the same pipeline into AI photo editing and virtual staging software for other platforms.

The same room staged in a warm luxury style, with beams, windows and mantel unchanged
A cluttered canal-house room with moving boxes and stacks of books
Before After
Declutter and stage in one pass. Example images, generated for illustration.

One house style for the whole property

A property gets one pinned house style, so every room looks like it belongs to the same home. Six presets (Modern, Scandinavian, Luxury, Industrial, Mid-century and Coastal) cover most needs. For everything else, a custom style starts from a description and optional reference photos: a vision model reads the references, a language model writes the style spec, and the app renders a moodboard and a sample room.

In the editor, sliders for furniture density, warmth and decor and locks for windows, floors and walls give control without writing prompts. Edits by prompt create new versions instead of overwriting old ones, and any version can be marked as the final.

One empty living room in six versions: empty, mid-century, Scandinavian, luxury, industrial and coastal
One empty room, five styles. Example images, generated for illustration.

Labels and exports you can publish

Every render carries IPTC metadata marking it as made with AI, and libvips bakes a disclosure bar into the image, which is always on for free accounts. Paid plans can design their own disclaimer templates, translated into five languages by a language model. We explain the approach in our guide to labelling AI-generated images.

Exports arrive as a ZIP organised by style, aspect ratio and room, in JPG or PNG, cropped to any of nine aspect ratios. Buyers get a showcase page per property: public, password-protected or off, with before and after views, room descriptions and a listing description in five languages, written with the instruction never to invent facts.

The Twelve Lumens export dialog with JPG and PNG options, a watermark setting and a standard disclosure template
The export dialog, with the disclosure template applied to every image.

Built to rank in five languages

Twelve Lumens publishes about 600 indexable pages in English, Dutch, German, French and Spanish: 164 local landing pages across 13 countries, guides for eight property portals, style and room guides, comparisons and a blog. Every city page is written natively for its own market, with real neighbourhoods and housing stock.

Sitemaps with hreflang and images, structured data, llms.txt files and social cards are all generated from the same registries, and tests enforce title and description lengths and follow every redirect. It's the setup behind our multilingual and programmatic SEO work.

The Dutch city page 'Verkoopstyling in Amsterdam' on Twelve Lumens, with local details about canal houses and IJburg
One of the local landing pages, written in Dutch for the Amsterdam market.
Results

Twelve Lumens in numbers

indexable pages in five languages
~600
room types recognised automatically
21
local landing pages in 13 countries
164
automated tests
1,900+

Stack

A Rails application with a Hotwire interface, background jobs on Solid Queue and images in a private Cloudflare R2 bucket. Public showcase images are served by a Cloudflare Worker that only accepts keys signed by the app. It deploys with Kamal to European servers.

  • Ruby on Rails 8.1
  • Hotwire
  • ViewComponent
  • Tailwind CSS
  • PostgreSQL
  • Solid Queue
  • OpenAI gpt-image-2
  • libvips
  • exiftool
  • Cloudflare R2
  • Cloudflare Workers
  • Stripe
  • Kamal
What we learned

What we learned

Keep providers behind your own classes

The first image pipeline ran on Fal's nano-banana. Because every AI task was already its own Generation class, moving to OpenAI's gpt-image-2 meant replacing those classes rather than rebuilding the product around them.

Free tiers have edges

Images were first resized on the fly by Cloudflare Image Resizing, until its free monthly quota of unique transformations ran out. The app now renders its own WebP variants and lets Cloudflare cache them at the edge, so visitors still get small, cached images without a quota to watch.

Test the metadata, not just the pixels

WebP looks like the obvious export format, but exiftool silently drops our IPTC disclosure tag when it writes WebP, and property portals don't accept it anyway. So exports are JPG or PNG, and every file that leaves the system carries its label.

Want staging or AI photo editing in your platform?

Tell us which photos, which edits and for whom. You'll get an honest view on quality, running costs and how to integrate it.

martin@balk.dev

Typically a reply within one business day.

Useful to mention

  • What you want to build, in a few sentences
  • Who is going to use it
  • Any deadline or launch date
  • Tools and systems it should connect to