Skip to content
m4bwav edited this page Nov 15, 2026 · 3 revisions

Why does the Sierpinski carpet measure 1.81 and not 1.89?

Grid (Kronecker) fractals read 0.05 to 0.07 low when the box sizes are not powers of the grid factor, because the counts swing log-periodically. On boxes of size 3^j the carpet reads exact; the tests record both. Read the local_slopes line and the verdict, not the single number.

What does "descriptor only" mean?

The fit covered under two decades of box size. Below that, a box-counting slope describes the picture but is not a trustworthy dimension. Render larger to get more decades.

Is the Mandelbrot area exact?

No. Pixel counting at 512 px and 10 000 iterations gave 1.5071 against the reference 1.5066. More pixels expose more boundary points that have not escaped at that iteration count, so the error does not fall evenly with size.

Why float64 and not the GPU?

In the benchmark behind the design, float32 already failed at a view width of 1e-4. NumPy float64 is the reference. The browser route (render --route browser, the explorer) is a float32 accelerator that hands views too deep for it back to NumPy; past float64's own limit, Mandelbrot and Julia go to perturbation (render --deep).

Does it use AI to draw fractals?

Stills are exact math. Videos are exact too unless you pass --ai (off by default since 0.2.2). With --ai and a local ComfyUI answering, a second video marked interpreted is written next to the exact one, which is always kept. See Video and audio.

Will a video trigger seizures?

Every video is checked against the WCAG general and red flash thresholds before it is reported, and audio damps flashes as it renders. A failing video exits 4 and is not shipped. The check is a threshold rule, not a medical guarantee.

Clone this wiki locally