InkField

Everyone chases Gaussian splats toward photoreal. I wanted the opposite: to read a scan as drawing.

A splat scene isn't a smooth mesh. It's thousands of little oriented blobs -- Not spheres, but ellipsoids, each with a direction and an elongation baked into its covariance. Most stylized renderers ignore that and fall back on a brightness-to-ASCII gimmick. Boring! InkField uses the structure instead: neighboring splats are assigned glyphs and strokes that match the real eccentricity and orientation of the geometry, so the marks describe the surface rather than just its brightness.

But something was missing as a DP...so I gave the renderer a camera a cinematographer would recognize: focal-length presets, an aperture, a focus distance, a focus pull. The twist is that defocus is never a pixel blur. It's a property of each mark, meaning marks grow, soften, thin out, and bloom into bokeh discs as they leave the focus plane.

Most NPR tools think like illustrators. InkField also thinks like a camera.

The lens model carries the controls found on set: a 24 / 35 / 50 / 85 mm focal preset, an aperture in stops, and a focus distance. A focus-peaking preview tints in-focus marks so I can confirm where the plane sits.

An animatable camera rig drives two deterministic path. An Orbit turntable and a Dolly in with play, scrub, duration, and speed. Scrubbing to a given progress always lands on the same pose, which is exactly what makes a rack-focus or a dolly push repeatable. It's the same virtual-production logic I built into CineWave's digital twin, pointed at a renderer instead of a robot arm.

The whole point: give an indie creator a real lens vocabulary over a scanned object, and let them pull focus as a drawing decision rather than a post-blur.

Under the marks is a graphics pipeline, built on a stack of genuinely beautiful research I borrowed for a deliberately silly purpose.

Splats project to screen as oriente ellipses (EWA splatting). The major axis of each gives a line direction; because line directions have no sign, they're accumulated as double-angle vectors so averaging is actually correct. For surface-locked looks, a background worker recovers per-splat normals straight from the stored covariance, makes their signs consistent with an MST, and smooths a tangent line field. It's the same math as a 2-RoSy direction field, with topologically unavoidable defedcts that the tracer detects and routes around.

Flow mode traces evenly spaced streamlines through that field and stamps discrete glyphs along them, using tonal art maps so darker regions only ever add marks, which is what makes relighting flicker-free. The depth-of-field math lives as pure, Three.js-free, Node-tested modules, so the live view and the offline export share one implementation.

The full derivations, includingcovariance projection, the nematic field, surface-frame recovery, streamline placement, tonal hatching, and the determinism story are written up in the technical notes, with citations.