Building Visualization

Overview

A digital twin of a real building can be hundreds of thousands of components across dozens of floors. The tools people want are simple — drag a box through a wall to see the pipes inside, select floor 8 and ghost everything else — but the obvious implementations all scale with the size of the building, which is exactly backwards.

The approach

Clip volumes write their transforms into a single Material Parameter Collection, once per frame and only when something changed. Every clippable material reads that collection and evaluates the spatial test per pixel on the GPU, feeding Opacity Mask. The CPU never looks at the building: one mesh or four hundred thousand, the work is identical.

The catch is that a per-pixel cut has to run in each material's shader. But meshes and materials aren't the same count — a 400,000-component building typically shares a few dozen materials, and injecting into a parent material covers all its instances. A Python editor script does the injection in one call, with a dry run first.

What didn't work

The natural design — one material function every material calls — was built first and failed silently: connections made inside a material function through Unreal's Python API don't persist, so the graph saves with every node present and every wire missing. A single Custom HLSL node was next; it compiled and cut nothing, with one input binding impossible to inspect from outside. The current version builds the graph from native material nodes only, which removes every opaque link.

Status

This is work in progress, and the repository documents it honestly. Verified: the collection carries the right values, the generated mask is correct (routed to Base Colour it draws a crisp box exactly inside the volume), and one real cause was found and fixed — UMaterial::GetBlendMode() reports Opaque for a Masked material when bCanMaskedBeAssumedOpaque is set. Still open: geometry inside the volume is drawn on some pre-existing materials even with a literal zero on Opacity Mask, while a freshly created material clips correctly. The next step is a controlled comparison of the two materials' compiled shaders. Clip volumes are currently axis-aligned.