Luxar demo datasets: ShareAlike (CC BY-SA 4.0)

Precomputed demo datasets for Luxar whose upstream sources carry ShareAlike terms. Why these are a separate record ShareAlike propagates to derivative works. A Gaussian-splat fit of a CC BY-SA volume is a derivative of that volume, so it must itself be released under CC BY-SA 4.0. Keeping these files apart from the permissively licensed ones means a user who takes the other record does not unknowingly acquire a ShareAlike obligation. How to use these files Each file is a .gsplats.zarr.zip archive that the Luxar demo suite fetches on first run and caches under ~/.cache/luxar/ /. To fetch one directly: from luxar.demos import ensure_dataset paths = ensure_dataset("gsplats_opencell_map4") # downloads, verifies, returns paths To open one in the viewer without writing code: luxar gsplat view opencell_map4_ch0.gsplats.zarr.zip luxar demo # browse every demo that uses these files The zebrafish timelapse is a single 4D node rather than a bundle of per-frame stores: one archive holding four coarse-to-fine levels, each progressively streamed. Every file is pinned by SHA-256 in data_manifest.json, and the fetch helper verifies the checksum before use; a mismatched download is quarantined rather than returned. How the Gaussian-splat files were produced Each source volume was fitted with the Luxar gsplat pipeline (luxar gsplat fit) on a GPU. These are lossy representations built for visualisation and interactive exploration; they are not a substitute for the source imagery in quantitative analysis. Choosing the splat count. The number of Gaussians K was calibrated per dataset by Noise2Self blind-spot cross-validation (luxar gsplat cal), and then checked by refitting at the candidate count and scoring the reconstruction against the whole source volume. The second step is not a formality. The cross-validated estimate is computed over a five percent mask of held-out voxels, and on these datasets it disagreed with the direct measurement in both directions: for one dataset it favoured half the splat budget that measurement showed to be better, and for another it predicted a gain of about 1.5 dB where refitting delivered under 0.3 dB. Where two candidate budgets fell within roughly a decibel of each other, the choice was made from renders of both at matched magnification, because at that margin the metric and the eye disagree. Additional Gaussians capture real structure and photon noise alike, and peak signal-to-noise ratio rewards the two equally. Level of detail These files carry level-of-detail structure so a viewer can show something immediately and refine it as the camera approaches, instead of waiting for a whole fit to arrive. The mechanism used here is: Progressive (additive) ladders. A leaf is stored as an ordered series of chunks whose union is the complete fit, so any prefix is a valid, dimmer rendering of the whole object. Every file in this record carries one. Both datasets here are small enough to load whole, so neither carries the coarse replacement levels that the larger objects in the companion record use. Reconstruction quality Two figures are reported per file, and both are stamped inside the archive itself rather than only quoted here: PSNR over the whole volume, in decibels. Foreground PSNR, over only those voxels above an Otsu threshold of the source volume, with the fraction of voxels that threshold selects. The foreground figure is the honest one for sparse microscopy and astronomy data. Where a volume is mostly empty, a global PSNR is dominated by correctly reconstructed emptiness and can look excellent while the structure of interest is poorly represented. A low foreground figure is also not automatically a defect: on a noisy channel it partly measures photon noise that a sum of Gaussians declines to reproduce, which is desirable behaviour that this metric charges for. Where that applies it is noted against the dataset. To read the stamps from a downloaded file, so that nothing here has to be taken on trust: luxar gsplat info .gsplats.zarr.zip Each archive also records the grid and element type of the volume it was fitted from, which is what makes the compression ratio quoted against it reproducible rather than asserted. The files in this record were computed between 2026-08-16 and 2026-08-24; each archive records its own build timestamp, readable with luxar gsplat info. The date is stated here rather than in the structured dates field because Zenodo stores a date range only when its start and end differ, and inventing a second date to satisfy that would misreport when the work was done. Contents filesizeelementslevel of detailPSNR / foreground (dB) opencell_map4_ch0.gsplats.zarr.zip0.8 MB59,6094 ladder rungs27.4 / 20.8 opencell_map4_ch1.gsplats.zarr.zip0.8 MB59,4784 ladder rungs24.8 / 18.0 zebrafish_4d.gsplats.zarr.zip19.2 MB1,270,2334 coarse-to-fine levels, each progressively streamednot scored Every figure above was measured from the file this record holds. “Elements” is what a viewer draws at full detail: splats summed over the finest level of each part, or rows for a coordinate table. Foreground PSNR is restricted to signal voxels and is the number to judge a fit by — a global PSNR on a mostly-empty volume largely measures how well the fit reproduces empty space. A timelapse bundle is scored on one representative frame, not the whole series; “not scored” means no quality figure was recorded at fit time and none is claimed here. gsplats_opencell_map4 (1.5 MB, cc-by-sa-4.0). Source: OpenCell (Biohub) MAP4. Attribution: Cho, N. H. et al. OpenCell: Endogenous tagging for the cartography of human cellular organization. Science 375 (2022). CC BY-SA 4.0. gsplats_zebrafish (19.2 MB, cc-by-sa-4.0). Source: Zenodo 1211599 (zebrafish embryo, confocal laser-scanning). Attribution: Aanstad, Pia. Time-lapse (4D) volumetric fluorescence microscopy image sequence of a living zebrafish embryo. Zenodo (2018). CC BY-SA 4.0.