Size versus quality

Every compressed splat format covered in this part reaches its size reduction through some combination of the same three levers: quantisation, dropping spherical-harmonics detail, and sorting the data so a generic codec compresses it better. Knowing what each lever actually costs in quality, rather than treating “compressed” as one undifferentiated trade-off, is what turns a size target into a specific, defensible export decision.

Quantisation: fewer bits, mostly invisible, sometimes not

quantisationStoring a value at lower numeric precision than it was computed at, for example rounding a float32 position to 16 bits or mapping a scale to one of a fixed set of codebook entries, to shrink storage size at the cost of some accuracy. means storing a value at lower precision than the trainer optimised it at, a float32 position rounded to 16 bits, a scale mapped to one of 256 codebook entries rather than kept exact. SPZ quantises positions to 24 bits and SH coefficients to a configurable 4-8 bits per band; SOG goes further, splitting positions across two 8-bit image channels and mapping scales, colour and higher SH bands through learned or fixed codebooks. Niedermayr, Stumpfegger and Westermann’s academic compression work reports up to 31x reduction on real scenes using sensitivity-aware vector clustering, quantising the attributes a scene is least sensitive to more aggressively than the ones it’s most sensitive to, rather than applying one flat bit depth everywhere. That’s the general shape of every quantisation scheme worth using: not “fewer bits everywhere,” but fewer bits where the eye won’t notice and more precision kept where a Gaussian’s exact position or shape actually matters to the silhouette or a specular highlight.

Dropping SH bands: the biggest lever, and the easiest to overuse

Spherical harmonics store view-dependent colour, and higher bands are the most expensive part of a raw PLY: degree-3 SH is 45 of a Gaussian’s roughly 59 stored floats. Dropping to degree 0 keeps flat diffuse colour only and removes essentially all of that cost, which is exactly what the .splat format does by omitting SH entirely. Research on where higher bands actually earn their keep, including work behind adaptive SH schemes, has found that many Gaussians, particularly smaller ones with limited view-dependent variation, contribute little from their highest-degree coefficients even during training, so a blanket assumption that degree 3 is always necessary doesn’t hold up well under scrutiny. But “little” isn’t “none”: a glass facade, chrome fitting or wet stone surface is precisely the content whose specular response lives in those upper SH bandsThe successive groups of spherical-harmonics coefficients, band 0, band 1, band 2 and so on, each adding 2l+1 coefficients per colour channel; band 0 gives flat diffuse colour and higher bands add increasingly directional, specular-like variation at increasing storage cost., and dropping them uniformly across a scene softens or flattens exactly the details a client is most likely to notice on a hero material. The practical version of this lever isn’t “keep or drop degree 3” as a scene-wide switch; it’s recognising that a format like SPZ, which quantises every coefficient rather than collapsing them into a shared palette, keeps more of that detail than SOG’s centroid-based higher-band encoding does, at the cost of a larger file.

Spatial sorting: making the data compress-friendly before compressing it

Sorting is the lever that doesn’t touch a single value’s precision at all: it just reorders which Gaussian sits next to which in storage; SOG uses Parallel Linear Assignment Sorting to arrange Gaussians into a 2D grid where neighbouring cells hold similar attribute values, and other academic compression work orders Gaussians along a Morton, or Z-order, curve for the same reason, so that spatially close Gaussians end up close together in the serialised data. Either approach turns a scene’s inherent spatial coherence, nearby surfaces tend to share similar colour, scale and orientation, into something a generic codec can exploit: WebP compresses a smoothly-varying image far better than scattered noise, and run-length or predictive encoding both do better on a sorted sequence than a randomly-ordered one. This is why SOG’s compression ratio against raw PLY is noticeably better than a quantisation scheme applied to the Gaussians in their original, unsorted training order: the sorting step is doing real, separate work, not just packaging the same savings differently.

Practical budgets for the web

For a scene meant to load over a mobile connection rather than sit on a LAN, a rough target worth working back from is comfortably under 30MB for broadband delivery and closer to 15MB if a meaningful share of traffic is mobile, with total Gaussian count kept under a few million if mobile viewing matters at all. Those are targets to hit through format and quantisation choice first, PostShot’s Studio tier and most open trainers can also be told to cap Gaussian count or densification aggressiveness during training, rather than through cropping SH bands uniformly across an otherwise untouched scene as a last resort.

In practice

The order to reach for these levers in is spatial sorting and quantisation first, since a well-sorted, quantised SOG export costs almost nothing in perceived quality for most architectural scenes, and SH-band reduction last, and selectively, only on regions or exports where the size budget still isn’t met after the first two levers have done their work. If a hero material genuinely needs the specular detail higher SH bands carry, that’s a case for exporting that asset via SPZ, whose per-coefficient quantisation keeps more of that detail than SOG’s shared-palette higher-band encoding, not a case for abandoning compression altogether.