Scroll geometry: lasagna fields, spiral fit surfaces, m7 surface predictions and dinovol slabs

Geometry computed from public Vesuvius Challenge CT scans with the community's open models and code, so you can reuse it without running a GPU job yourself: one villa lasagna field, villa spiral fit surfaces, m7 surface predictions and two dinovol embedding slabs. Everything here is Model output, not a reading. No file holds ink labels, ink maps, hand-made annotations or CT voxels, and no surface here has been checked by hand.

All scans are Vesuvius Challenge open data. None of the scrolls here is on the prize eligibility list (Sourced fact: villa prizeEligibility.json at deb1b695, read 3 Nov 2026, and rechecked on villa main, 11 Nov 2026).

What is in it

Paths are KIND/SCROLL/VOLUME/.... VOLUME is the open-data volume id; -L2 means level 2 of that volume.

The volumes are at https://vesuvius-challenge-open-data.s3.us-east-1.amazonaws.com/SCROLL/volumes/ (PHerc1667's 20231117161658 volume is at https://data.aws.ash2txt.org/samples/PHerc1667/volumes/).

The files

Lasagna field. tNNN/PHerc0172-lasagna-tNNN.tar holds one part of the field as written by villa's lasagna-preprocess predict3d (OME-Zarr groups PHerc0172_cos, _grad_mag, _nx, _ny and the manifest PHerc0172.lasagna.json); tNNN/meta/ is a copy of each part's small metadata files. To use the field, unpack all five tars into one folder, then copy PHerc0172.lasagna.json from the field's folder over the one the tars wrote: it is the joined manifest that lists all five crops. Only z 9,024 to 10,560 was predicted; a few edge chunks of parts t003 and t004 reach past z 10,560. The parts' sha256:

Spiral fit surfaces. Each seedN-VARIANT/SCROLL_zA-B_sN.tifxyz/spiral-checkpoint.tifxyz/ is one tifxyz surface (x.tif, y.tif, z.tif and meta.json), as written by villa's spiral fit export. Some coordinate files are stored zstd compressed as x.tif.zst (and y, z); zstd -d gives the original TIFF. Coordinates are voxels of the volume and level in the folder name. The variants:

• A: villa's spiral fit with its default settings, 30,000 steps, on a villa lasagna field of the same scan.
• A-repeat: A again with the same seed, to show run to run spread.
• AV: villa's spiral fit with flow-gradient smoothing alone.
• E, G and Q: villa's spiral fit given relative windings (villa's relative_windings.json) made by the rule of Ringleader 1.1.0's ladders.py (page): E with round 1 ladders, G with round 2 grown ladders, and Q with G's plus-one control, which should make the fit worse. The ladder files are not included.
• G-pilot and Q-pilot: a second, separate run of G and Q with the same seeds.

They are here as surfaces to reuse or compare. This card makes no claim that one variant is better.

m7 surface predictions. Each box folder holds pred.zarr, an OME-Zarr (zarr v2) group with the full shape of the scan level and levels 0 to 5, where only the box's 192-voxel chunks plus a halo of one chunk on each side are stored; values are 0 or 255 (m7's foreground probability at threshold 0.2). Folders named crops/ also hold tNN/prob.npz with prob (uint8, 256 x 256 x 256, m7's foreground probability times 255, no threshold) and origin (the cube's z, y, x origin). tiles.json in each folder lists every box, the cube origins and the run settings.

dinovol slabs. zarr v2 groups with dense (int8, z x y x x x 64) and dense_scale (float16, z x y x x): value = dense * dense_scale, then each vector has unit length. Index i along an axis covers voxels [origin + 4 i, origin + 4 i + 4) of the level named in the folder, with origin z 3,360 (PHerc0009B) or 1,568 (PHercMANBp) and y, x origin 0. Outputs whose block has no nonzero voxel are zero vectors.

How it was made

• Lasagna field: villa at commit db3dbea26185a9e09ab997bfa21a2a00477f0d56, lasagna-preprocess predict3d, tile size 192, with checkpoint model_current.pt of scrollprize/lasagna at revision de45f01147adea9f4ad0f621b1ccd885a1f03760 (sha256 4cc0c40a846b296769b9f7982470559078662a1a0b0e1f74ff41ed9341c0d600); the run's inference.json records the villa checkout as having local changes (repository.dirty).
• Spiral fits: villa's spiral fit at commit e2254ebe0168a232e5ba7b94997c749b4e6a4544, eager mode, 30,000 steps, seeds 1 to 3, on villa lasagna fields of the same scan, around a center line made with the Center Finder (PHerc0172) or the open-data umbilicus (PHerc0332).
• m7: scrollprize/surface_m7_nnunet at revision cd83ccf7e1d58b70bd9074a9a7e158cd2c894f20, run with villa at commit e0bbb8b40a2db58b1d71864f286eb85717e59e64 (the plans' normalization, a 192-voxel halo, threshold 0.2).
• dinovol: scrollprize/dinovol_8um_ps192_102126, checkpoint teacher_step_640000.pt (sha256 2354c493f050070e17ea0cda5cd3c1a6ebdde19a70a9c28bc9d368b80faabb86), villa's dinovol code at commit 4ea26d9cbfc1778c68e0875c21c1f670922deba1, robust normalization with fixed volume-wide statistics, a 192-voxel tile lattice at stride 96.
• All runs used Hugging Face Jobs. The files were copied server side from the runs' working storage, byte for byte.

License and credit

The scans are Vesuvius Challenge open data, licensed CC BY-NC 4.0; everything here is derived from them and carries the same terms. Please credit the Vesuvius Challenge and its scanning partners for the scans, the villa authors for the lasagna, spiral fit, m7 and dinovol models and code, and this bucket. Made by Chase Hendrick, Independent Researcher.