City-scale 3D Gaussian splats converted analytically, at native texture resolution, from textured photogrammetry city meshes. Berlin: all 719 survey tiles of the textured city mesh are converted, 22.0 billion Gaussians at ~4 cm per splat (the source texel size), over a 26 × 28 km footprint.



Brandenburger Tor from Pariser Platz (street level)	Reichstag from Platz der Republik (street level)

Berliner Dom from the Lustgarten (street level)	Fernsehturm from Alexanderplatz (street level)

Olympiastadion and Westend	All 719 converted tiles, top-down, north up

All renders are of the splats themselves (no mesh), drawn with an out-of-core LOD renderer: per view it frustum-culls tiles, picks each tile's level by distance (full resolution within 250 m of a street-level camera), and alpha-composites far-to-near passes when a view exceeds one GPU's budget (the street views are 120–200M Gaussians each, on a 24 GB GPU).

Source: the textured Berlin mesh tiles from TUM's MovingDrone dataset.

Releases

folder	area	Gaussians	files
`berlin/`	Mitte (Berliner Dom, Museumsinsel, Fernsehturm), 1.46 × 1.46 km, 17 tiles	517M	4 chunks
`berlin_city/`	full city, 719 tiles, ~103 km²	22.0B	139 chunks (234.9 GB with LODs) + 1/16 and 1/256 LOD tiles

`berlin/` layout

berlin/
  scene.usda        # one Xform per chunk with a `splat:file` asset path + `splat:gaussianCount`
  manifest.json     # CRS, origin, frame, per-chunk counts and bounds, source tiles
  chunks/chunk_XXX.spz

Chunks. Splats from all source tiles are merged and then re-split. Each split cuts a block at the median Gaussian count along its longer horizontal axis, and splitting repeats until every chunk is at most 10 GB of uncompressed data. At 56 B per Gaussian (14 float32 values: position, scale, rotation, opacity and degree-0 colour) that is up to 178M Gaussians per chunk, so each chunk fits a single-GPU VRAM target when loaded. scene.usda places the chunks, and manifest.json gives each chunk's bounds for streaming or culling.

Frame. All values are in metres and Y-up, with x = east, y = up and z = south. They are given relative to origin_utm_enh in manifest.json, which is in ETRS89 / UTM 33N (EPSG:25833). To convert back: E = x + o[0], N = -z + o[1], H = y + o[2]. The same origin is stored in scene.usda's customLayerData.

`berlin_city/` (full city)

Same chunking rule and frame convention, at city scale. SPZ positions only cover ±2 km, so every chunk (and every LOD tile) is stored relative to its own origin, and scene.usda places it with a float64 translate relative to one city origin (manifest.json: city_origin_utm_enh). l1.usda / l2.usda reference per-tile LODs at 1/4 and 1/16 texture resolution (≈1/16 and ≈1/256 of the Gaussians) for overviews and streaming.

Format. SPZ v3 files with degree-0 spherical harmonics. Positions are 24-bit fixed point with 12 fractional bits, which gives 0.24 mm precision over ±2 km.

Conversion

Each Gaussian is placed analytically on the textured mesh surface by the splataverse mesh-to-splat converter. There is no optimisation step and no rendering loss. Each texel becomes a flat Gaussian on its triangle, coloured with the texel's colour.

Licence and attribution

The source meshes come from the MovingDrone dataset (Dhaouadi et al., TUM). Check the original dataset's terms before redistributing. Terms for the textured mesh layer had not been confirmed when this was published. The other geodata in that dataset (LoD2, DSM, orthophotos) is listed under dl-de/by-2-0.