Demonstrating Continuous Virtual Geometry for On-the-Fly Mesh Simplification

Hey everyone, I've been experimenting with the concept of virtual geometry over the past few weeks and wanted to share some exciting results.

What is virtual geometry?

• It’s an approximation of real geometry where the viewer cannot tell the difference between the "real" geometry and its approximation.

This approach is similar to LOD, but it's continuous. In typical LOD systems, you have fixed levels (e.g., LOD0, LOD1, LOD2) where each level reduces the triangle count. Here’s a picture:

The idea is to simplify the mesh on the fly by removing triangles that are smaller than, say, 1 pixel.

Below are several screenshots demonstrating the process. In the images, the left-hand side shows the “real” source geometry, while the right-hand side displays the “virtual” geometry using a debug shader. The source geometry contains exactly 100,000 triangles. At the initial resolution and camera angle, the virtual geometry also maintains 100,000 triangles, but as the camera moves away, triangles get dropped progressively.

For example:

~60,000 triangles

A bit further, ~10,000 triangles are rendered

And at a greater distance, only around 1,000 triangles are visible

This means that drawing 100 instances of these geometries utilizes the same GPU effort as drawing the original mesh (for example, 1000 triangles * 100 = 100,000 triangles).

The continuous nature of this LOD approach prevents popping effects, ensuring that the mesh's appearance remains consistent from any angle or distance, even though the actual triangle count is often a fraction of the original.

While the technique isn't completely new—it has roots in methods like the ROAM algorithm used for terrain rendering—my approach is more flexible, working with generic static meshes. However, this flexibility comes with additional engineering challenges, including the need for special pre-processing to create a custom data structure for real-time optimization. The pre-processing step currently takes around 2 seconds for the displayed geometry, but it enables near-instantaneous run-time optimization.

This is still a work in progress, with many improvements planned before it's production-ready. I thought the progress so far was cool enough to share. Enjoy!