The cached artwork is Cellularity d_0020 by Jonathan McCabe. Its native image is 59,049 × 59,049 pixels. Broad smooth forms contain fine stripes, spots, and sharply different neighboring textures.
McCabe’s description of Cellularity specifies four floating-point values per pixel, neighbor diffusion, and polynomial reactions influenced by the values’ history. This is distinct from his earlier multiscale Turing model. A licensed Cellularity description also mentions an expanding space.
Independent Turing scales: each scale has its own radius, inhibition range, reaction strength, preference, regional response, and color. The active layers compete on one shared pattern field. Separate history channels preserve their local influence. Smooth averaging uses three box-filter passes to approximate a Gaussian; Square box keeps the earlier kernel.
Regional reactions: two concentrations, two history fields, and evolving regional fields that vary feed and kill rates across the image. Different regions can develop different kinds of cells. This is an experimental construction, not McCabe’s recovered polynomial map.
Color: Independent layers blends a separately colored ramp for each active layer using the retained layer histories. Multichannel spectrum maps several state dimensions independently into RGB. Original single palette keeps the earlier mapping for comparison. Layer colors can be edited directly; color seeds now generate new ramps instead of only rotating a short list.
Inflation: an optional resampling step expands structures about the center while the simulation adds new detail. It runs on a fixed grid, so this is an approximation to a growing domain.
Playback speed is remembered on this device. It is separate from saved simulation states. Pausing or switching tabs clears the playback backlog, so returning does not cause a burst of catch-up steps.
New rules depend on the current field. Use Restart seed for a controlled comparison. Changing model or resolution starts over. Binary .cell checkpoints contain sixteen floats per pixel. Version 2 JSON files retain all their values; version 1 imports approximate missing histories. Equal seeds are reproducible on the same backend; floating-point differences between GPUs and the CPU can diverge over time.
Use + and − in Layers to add a scale or remove the selected one, up to twelve. Remaining histories and colors are kept. Regional reactions use these as color families.
Randomize perturbs the current settings. Choose an amount, affected controls, a seed, and the chance of adding or removing one layer. The seed advances after each click. Restart with new noise is optional.
The canvas follows the window without restarting. Its square field fills the window with a centered crop; PNG export includes the entire square.
Resolution accepts any integer size within the selected memory budget and device limits. Applying a new size restarts. GPU memory usage is roughly 384 bytes per pixel, plus the display and temporary resources. Resizing retains the old field until the new allocation succeeds. Browsers cannot report reliable free GPU memory, so a very large budget can still cause device loss or a tab crash.
PNG and .cell exports process rows in batches. Browsers with a Save File picker write directly to disk; other browsers limit the download buffer to 128 MiB. Large simulations still keep their coupled fields in memory. Use Cancel export to stop a file before it is committed.
The exact polynomial map, history mechanism, coefficients, color mapping, seed, growth schedule, and iteration count for d_0020. Neither mode claims a pixel-for-pixel reconstruction. Optional relief shading is a display effect added here.