The wall that grows itself is actually a slice at 1.2 meters high

You open the page and see a floor plan. Black lines, room names, dimension annotations, like a drawing printed out by an architect. Scroll down, and the walls rise up out of the paper; the camera drops to door height and walks from room to room. Before the camera reaches each room, it's a line drawing; only when it arrives are materials applied.

This is Casa Serra by Samnang Aing, a 946-square-meter courtyard house set in the hills of Sintra. When Meng To shared it, he said it's a new path that mixes architecture, web design, and 3D, and that Opus 5.5 paired with three.js is a bit of a beast.

The original author wrote on Twitter that this demo got a combined 60K views overnight on Threads and LinkedIn, and he said that had never happened to him before.

The floor plan is not an image

This is the easiest place to misread. That "drawing" is not an image at all; it is the actual 3D model, sliced at 1.2 meters high and viewed from directly above through a very long lens. Scrolling moves the cut plane upward, and the walls grow along with it. The drawing and the building are the same thing; there are no two sets of assets.

All geometry is generated in code; Blender was not used. The textures are CC0 scans from Poly Haven, about 39MB on laptop and 13MB on mobile. The stack is three.js, GSAP ScrollTrigger, and Lenis; it's plain files with no build step. On devices that can't run it, you see a still frame plus a walkthrough video.

This approach is worth noting: visually it looks 2D, underneath it's 3D, and between them there is only a clipping plane and an orthographic telephoto lens. What it saves is not just work, but also the hassle of keeping a "2D drawing" and a "3D model" in sync.

The cost is 70 prompts and two weeks

The README says it plainly: there was no single prompt. The whole project took about 70 prompts, two weeks, and four complete versions, using Claude Code plus Opus 5.5. The starting point was a short video of an old game website he liked; the floor-plan idea only emerged on day four.

The repo includes a starter prompt. The author tested it himself with a session that knew nothing about the new project: first ask it to give a few directions for the building layout and camera path, wait for the answer, then build in checkpoints. The first checkpoint does only three things: the drawing, the walls rising, and walking to the first room. About an hour later he got a result, and what ran was another demo, Casa Larga.

The author says the last paragraph of that prompt is as important as the idea itself: without that paragraph, you'll get a finished product you didn't choose.

Two real questions in the comments

Someone asked how to control the order in which the walls rise. Meng To replied that this prompt gives you the foundation; you need to change the specific order yourself.

Another comment that hit the point: a netizen said this kind of scroll narrative works because a floor plan is a model people already have in their heads; each scroll only adds one layer, so people don't get lost. The real difficulty is the reduced-motion fallback: the non-moving version has to tell the same story.

Another comment was quite practical: when architects use this to pitch clients, it closes deals much faster than static renderings.

Current status

The repo labels itself work in progress, early stage, and the experience is best on laptop. The release build's scripts are minified. Whether it runs on mobile depends on your device.

Following this line of thinking, someone in the comments posted a Chinese floor-plan editor that switches back and forth between 2D and 3D, places furniture at real millimeter dimensions, and can export the plan as JSON. If you actually want to build something similar, that furniture library and dimension table will save you more effort than the demo itself.