Demos

In the browser, nothing to install

▶ Open the gallery

The gallery lists demos that run entirely in your browser (Chrome, Edge or Safari 26, for WebGPU and service workers). Images are read from their public buckets and every chunk on screen is computed in the page when the viewer asks for it; nothing is uploaded or precomputed. Each card also shows the chunkmirage serve command that serves the same from Python, for Neuroglancer, Fiji, napari or a dask script.

demo	what is computed per chunk	data
Fly brain templates	the moving brain resampled through a field solved on your GPU	OME-Zarr RFC-5 examples
EASI-FISH rounds	an affine found, a field solved, finer fields fitted where you zoom	Janelia EASI-FISH, rounds 1 and 2
Los Angeles fires, any reader	burn severity (dNBR) by normalized_difference from two Sentinel-2 passes, before and after the Palisades and Eaton fires; read by your pick of Neuroglancer, OpenLayers (as GeoZarr) or GDAL itself in WebAssembly (as zarr v2), which writes a georeferenced GeoTIFF to download	Sentinel-2 L2A, 2 January and 1 February 2025
Six tiles stitched by RANSAC	interest points in every overlap, matches, RANSAC and the global fit (rerun as you move a setting, with inliers and rejected matches drawn), then each fused chunk: chunkmirage.stitching, as stitch:// runs it	BigStitcher-Spark's stitching example, a larval fly CNS
Organelle contact sites	the mitochondria and ER predictions thresholded at 128 (mitochondria labelled), and contacts then label on the two, stacked and flipped into the EM's frame	OpenOrganelle jrc_hela-2
Single mRNA molecules	spots on both FISH channels	Janelia EASI-FISH, fly central brain
Follow nuclei through two days	not chunks but lineages: each nucleus double-clicked is followed frame by frame by its overlap with the next frame's nuclei (chunkmirage.tracking, one call a frame), reading only a box around it from each; when one collapses into mitosis and is lost, the new nuclei that appear near where it was (labels nothing covered the frame before) are guessed to be its daughters and followed too, a guess the segmentation cannot confirm; volumes plotted as frames come in; the nuclei followed served as a segmentation of their own, made per chunk from the bucket's (each lineage one segment in every frame, where the bucket numbers nuclei afresh in each), served again as tracking fills in frames the viewer has	Allen Institute for Cell Science, Dixon et al. 2024: hiPS colony time-lapse, lamin B1 and nuclear segmentation
A 3-D fractal to zoom into forever	the Mandelbulb itself, synthetic://mandelbulb, computed by chunkmirage's synthetic source in the page's workers (nothing is read): a slice to zoom into, finer levels iterating more, and the bulb volume rendered, finer levels loading as you zoom into it (a 3-D panel's depth of 2 view heights and 512 samples: Neuroglancer picks one level for the whole depth)	computed: an array 2^28 voxels across, 21 levels
The Mandelbulb as a solid surface that sharpens as you zoom	a multi-resolution mesh: four levels of detail (the bulb at 256³ to 2048³), each node marching cubes over a chunk of synthetic://mandelbulb, Draco-encoded and padded, made when Neuroglancer asks for its bytes; zoom in and finer nodes replace the ones in view	computed: levels 17 to 20
Shackleton crater's rim in 3-D	a terrain mesh of NASA's 5 m elevation (two triangles per 20 m cell), each fragment made when fetched, beside the elevation as stored	LOLA south-pole elevation, Shackleton rim
Where to land at the Moon's south pole	hillshade (the sun's direction and height at sliders, computed again for the tiles on screen) and slope (ground under a chosen slope coloured by the map) on 5 m elevation, any of 26 sites; drawn by OpenLayers, not Neuroglancer, reading the page's chunks as GeoZarr, fitted to the site so everything on screen is computed	NASA's LOLA south-pole elevation maps (Barker et al.), cloud-optimized GeoTIFFs
Hurricanes' cold wakes	diff along time: each day's sea temperature minus the day before's; beside it the temperature itself, shifted from kelvin to °C by scale	NASA MUR sea-surface temperature, 2002 on, 0.01°
Gulf Stream fronts	gradient along latitude and longitude, smoothed over 3 km first (sigma): two channels, coloured by the direction warmer water lies in and brightened by how fast it warms; the op returns only the interior it can compute, as a model does	NASA MUR sea-surface temperature, 15 February 2018
A solar flare (Python only)	diff along time on NASA's SDO images of the sun; the gallery shows its command, since NASA's bucket allows no browser page to read it	SDO machine-learning dataset, AIA 171 Å

The pipeline demos run chunkmirage's own Python in the page: Pyodide (Python compiled to WebAssembly, with numpy, and scipy only for pages whose ops import it) loads the package's ops and chunkmirage.fused, the code a server's pipeline stage runs, in a few web workers (4 to 7 s, once). Meanwhile the page opens the data and starts the viewer, whose first requests wait for Python; a package an op imports unannounced is loaded when it first does. One reader worker reads the images in TypeScript, as the Python sources do (OME-Zarr through zarrita, N5 itself, xarray-written zarr with its coordinates and CF packing, stack:// and flip://), decoding each store chunk once for the page; the page hands a worker the block a chunk needs and gets the chunk back. tests/test_fused.py checks that a chunk computed this way equals the pipeline's. On an RTX 2080 Ti workstation the contact-sites view filled in about 40 s, a spot or contact block taking 0.05 to 0.13 s in Python; chunks the viewer gave up on before their turn are dropped unrun, as the server drops them. Sources the workers compute themselves (synthetic://) skip the reader: a worker describes the source's levels, and generates each padded block where it computes the chunk. A demo is an entry of web/src/cards.ts: views (each a pipeline spec: source, select, ops, chunks) and a Neuroglancer layout, or for a map demo (map.html) OpenLayers layers, styles and sliders. The stitch page (stitch.html) runs chunkmirage.stitching's steps in the same workers, one call each, and serves its fused volume as a view whose chunks it computes itself; the track page (track.html) runs chunkmirage.tracking's step a frame at a time on boxes its reader reads at each timepoint (a read can pick its own entry of a source's other axes, {t: 12}). The pages share the engine (web/src/engine.ts), which serves the views as OME-Zarr and as GeoZarr.

The hurricanes card shows the date at the viewer's position, and what the storms were doing that day, beside the viewer (which shows time only in seconds); its play button steps a day a second. It reads MUR, NASA's daily 1 km sea temperature of every ocean (6443 × 17999 × 36000 values), which has one resolution in 65 MB tiles of 5 days × 18° × 36°: it opens zoomed in on Katrina's track inside one tile, one download per 5 days, and zoomed out to the globe the viewer would fetch every tile at full resolution. Its contact-sites card starts from OpenOrganelle's predictions rather than its published segmentations, which are stored in 512³ blocks (268 MB decoded) at full resolution; examples/contact_sites.py --segmentations uses those from Python, next to OpenOrganelle's published contact sites.

From Python

script	shows	needs
examples/contact_sites.py	the contact-sites demo served by Python, recomputing at a prompt; --segmentations starts from OpenOrganelle's segmentations and shows its published contact sites too	uv sync --extra all
examples/fish_spots.py	the spots demo served by Python, thresholds at a prompt	uv sync --extra all
examples/hurricane_wakes.py	the hurricanes demo served by Python, any date and place, the lag at a prompt	uv sync --extra all
examples/model_at_one_resolution.py	the shape of a live-inference op: a stand-in network that reads 8 nm voxels and writes two channels of 16 nm ones, served over OpenOrganelle's HeLa EM with the coarser levels downsampled from what it made (ops at one resolution)	uv sync --extra all
examples/solar_flares.py	the running difference of NASA's SDO images of the sun (no browser access to that bucket, so Python only), on the X1.6 flare of 10 September 2014	uv sync --extra all
examples/fly_brain_registration.py	the fly templates through their published OME-Zarr 0.6 transformations (scene://)	uv sync --extra all
examples/track_nucleus.py	a nucleus and its descendants followed through the colony time-lapse (tracking.lineage), volumes per frame written as CSV (the track page's work, from Python)	uv sync --extra all
examples/register_demo.py	a deformable registration solved as the source opens, re-solved as you type settings	uv sync --extra all --extra gpu (or --extra cpu)
examples/swirl_demo.py	a procedural deformation and its field, animated on a time axis	uv sync --extra all
examples/demo.py	one small volume served in every format at once, with a shared cache	uv sync --extra ops
examples/quickstart.py	the library in thirty lines: a cached blur, a threshold, a server	uv sync --extra ops

Each prints a viewer link; see Getting started for the basics.