Verified performance captures#

This page holds the pictures and the measured evidence behind the rest of the section. Everything here is a record of one machine on one day. None of it is a target, a guarantee, or something to compare your own machine against directly.

The benchmark panel#

Note

There is deliberately no screenshot of the metric grid after a run.

Those tiles show a live frame rate, so any capture of them records how busy the capture machine happened to be, not what Cellucid does. A trial capture on a loaded build machine read “FPS 2” at 100,000 points — a true statement about that machine and a badly misleading one about the software. The metric names, units and meanings are written out instead in Benchmark tools, where they can be checked against the interface and do not decay.

Acceptance samples (historical, one machine)#

The whole window after the Benchmark panel generated one million points of the Model surface pattern: the synthetic point cloud on the canvas and the panel's presets, point count, pattern and embeddings controls in the sidebar; the live readout is omitted. — One million points of the **Model surface** pattern, made by **Load synthetic data**, beside the panel that made them. The live readout is omitted: it would report the machine that took the picture, not Cellucid.#

The scripted 1440 × 1000 acceptance series recorded the following samples on that machine and that build:

Browser engine	Points	Generation	FPS sample	Frame-time sample
Chromium	10,000,000	1,592 ms	18	56.33 ms
Firefox	10,000,000	1,838 ms	18	55.76 ms
WebKit	10,000,000	1,735 ms	19	53.74 ms
Chromium	20,000,000	3,082 ms	8	123.00 ms

Every row loaded its pattern exactly once, published the exact requested point count, returned WebGL error 0, and emitted no browser error or warning.

Warning

These figures predate the renderer changes of 2026-08-01 and have not been re-measured since.

Two of those changes move the numbers in opposite directions depending on the scene. Hidden points are now rejected before rasterisation, and synthetic generation moved off the main thread. A synthetic acceptance run draws every point it generates, so the first change should barely touch these particular rows; the generation column is the one most likely to have moved.

Treat the table as a historical acceptance record. If you need a current number, run the benchmark yourself — Benchmarking methodology and metrics explains how to make the result comparable.

Compare measurements only on the same machine, browser build, viewport, and dataset. Numbers from different engines are different benchmarks.

Layout shapes#

Cellucid renders any number of views: the live view plus any number of snapshots. Two and three views are laid out as a single row; four or more as a grid of ⌈√n⌉ columns (four as 2×2). That difference matters more than the count, because point size follows the height of the pane a point is drawn in.

The whole window in Grid compare with the sidebar open: two equal panels to the right of the sidebar show the Pancreas 3D embedding from the same linked camera — view 1 coloured by cell_type, view 2 recoloured by S_score on a Viridis scale — each under a title chip naming its embedding and field. The sidebar shows Compare views with the two view rows, each marked linked and Orbit, View 2 shows UMAP · 3D, Keep view, Link cameras, Layout Grid compare and Clear, and Coloring & filtering with S_score in the Continuous obs select. — Two views share the full canvas height and split the width to the right of the sidebar equally, so each point keeps its full size and the pair draws every cell twice — roughly twice the point shading of a single view. This is the most expensive layout per unit of insight — tune in one view first, then split.#

The whole window in the dark theme on the Grid (dark) background, in Grid compare with the sidebar open: two equal panels show the Pancreas 3D embedding coloured by cell_type from the same linked camera, each under a title chip. The sidebar shows Visualization with Theme Dark and Background Grid (dark), and Compare views below it with the two view rows. — Theme and background apply to every panel of the comparison layout alike, and neither choice changes the cost of the frame.#