Throughout the history of 3D modeling, surfaces have been defined explicitly. From each vertex, edge, and face of a polygon mesh to the control points and weights of a NURBS curve, "storing the surface" is something you actively do. The surface is the construct — the core output of the modeling tool. Johan Holwerda's WebGPU SDF Editor, published on Reinder Nijhoff's RenderQueue, does something fundamentally different: it never stores any surface at all. What it stores are distance functions — continuous mathematical functions that, for every point in space, answer the question "how far are you from the nearest surface?" The surface itself is merely the zero-crossing of this function, a byproduct of computation, a remainder of mathematical derivation.

This architectural inversion cascades. Because the surface is never stored, it must be re-extracted every frame — through Marching Cubes or Surface Nets algorithms that approximate the zero-crossings of the continuous function into discrete points or triangles. Every surface you see is the sediment of real-time computation, not a prior definition. Deeper still: Holwerda's renderer abandons even triangles. Surface Nets vertices are rendered directly as pixel-sized points — because in a space of 2^32 grid cells, the lookup table needed to find neighboring vertices for triangle construction would require 4 billion entries, which exceeds available memory. The point cloud is what remains when the triangle construct is chiseled away by memory constraints. And temporal anti-aliasing (TAA) fills the gaps between points by blending sub-pixel-jittered results across frames — the final image is the temporal remainder of point accumulation over time.

The chisel-construct cycle manifests here in at least three layers. First: polygon modeling is the construct of 3D form-making; SDF chisels away explicit geometric representation, leaving pure mathematical distance as the remainder — the surface is never owned, only implied. Second: triangle rendering is the construct of real-time graphics; memory constraints chisel away triangle indices, leaving isolated points as the remainder — point-cloud aesthetics are not a design choice but a forced consequence of constraint. Third: WebGPU itself is the remainder of the browser's security model chiseling away native GPU access — spatial partitioning, octree splitting, marching cubes, 1024-direction ambient occlusion shadow maps, all running as compute shaders in a browser tab, each step negotiating with sandbox boundaries.

The naming gap is almost comically wide. Is it a CAD tool? (No engineering constraints, no manufacturing export, no parametric history.) A sculpting app? (No brushes, no clay metaphor, no subdivision surfaces.) A shader toy? (It has a full hierarchical scene graph with groups and boolean operations.) A math visualizer? (It produces actual 3D objects you can rotate, scale, and light.) A game engine component? (No physics, no scripting, no game loop.) Five naming systems fail simultaneously. It lives on Reinder Nijhoff's RenderQueue — an experimental rendering playground by a Dutch creative coder — as a guest contribution. One person's experiment, neither paper nor product, neither demo nor tool. The logic is still growing.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics