Flow Field is the generative line’s flagship — a GPU particle system that advects hundreds of thousands of particles through a vector field and renders them as points, dye, or trail-feedback ribbons. It runs the simulation entirely on the GPU with WebGL2 float-texture ping-pong: positions, age, and seed live in texels, advected each frame in a fragment pass and drawn instanced. A float-tier probe and a graceful 2D-canvas fallback keep it honest on weaker hardware — never a blank canvas.

The moat is the field sourcing. Beyond curl/simplex noise, the field can be derived from an image (a luminance Sobel gradient, so particles trace a photo’s contours) or from an SDF text mask (an exact Euclidean distance transform that confines particles into letterforms) — the kind of effect that’s genuinely hard to reproduce from a CodePen snippet. Cover-fit is handled so a source image maps cleanly across any container aspect ratio. And it speaks audio: with the shared @caustics/audio-core contract, field: 'audio' lets a Sonogram instance modulate the flow — bass swells the swirl, treble shimmers — a no-op on silence.

It is built milestone by milestone against a stable surface and carries the full Caustics lifecycle: a pointer force (attract / repel / swirl) on fine pointers, film grain in the composite, fourteen presets across points / trails / dye / lines, OKLab color ramps, IntersectionObserver pause, reduced-motion handling, WebGL context-loss recovery, DPR capped at 2, and an idempotent destroy() that releases the float targets and programs. Ships for Svelte 5, React 18+, and vanilla JS through one mount().