3DGS Preset Config

Background

A single configuration cannot cover every 3DGS scene. Data precision, file size, GPU memory, device performance, and quality requirements vary by scenario, so each integration should start from a suitable configuration preset.

This page summarizes common 3DGS file formats, packType choices, presets, and key parameters that can be tuned on top of a preset.

Quick Choice

Choose a preset from the business constraint first, then tune only the key parameters. Avoid changing precision, sorting, and blur parameters all at once.

Scenario Goal	Starting Point	Watch
Best quality and strong user hardware	Max Quality	packType, packHighPrecisionEnabled, highPrecisionAttachEnabled
Large scenes that break at low precision	Quality First	compressed, high-precision merge, maxStdDev
Lower-end devices with acceptable quality	Performance First	super-compressed, detailCullingThreshold, maxPixelRadius
Extremely large scenes or very low-end devices	Extreme Performance 0	repackEnabled, sortMinDuration, more aggressive precision compression
Source data is sog and the goal is larger scene scale	Extreme Performance 1	sog, precalculateEnabled, GPU memory usage

3DGS File Formats

Format	Size	Render Quality	Implementation Notes
ply	100%	Good	Highest source precision and largest file size.
supersplat compressed ply	30%, about 17% after gzip	Good	Splits data into chunks of 256 splats, likely spatially grouped like ksplat. center, quat, scale, and rgb are compressed by min/max, rescale, and quantization. SH can be compressed to u8; observed data is about 5 bit.
spz	10%	Average	Keeps relatively high precision for splat data, especially center, so sharpness loss is lower. SH precision is low and can cause visible color shifts in fine-detail scenes.
splat	14%	Average, not universal	Drops shN during compression. Layout: center.xyz (f32), scale.xyz (f32), color.rgba (u8), quat (u8), for 32 bytes total.
ksplat	20%-30%	Depends on compression level	Level 0 is uncompressed, level 1 is 16 bit, and level 2 is 8 bit. It spatially clusters splats for local coordinate compression, roughly similar to compressed ply.
sog	5%	Average	Applies PLAS sorting to center, scales, quats, and sh0(rgba), then uses min/max and quantization. shN uses k-means clustering with centroids and labels to restore data while reducing size. Images tend to be blurrier.

packType

packType controls the precision of objects produced while parsing splats. Different options trade quality, size, memory, and performance.

Compressed

Field	Precision
position	f32 (3)
scale	f16 (3)
quat	f16 (4)
color & alpha	f16 (4)
shN	s_11_10_11

Compressed favors image quality and data precision. Use it for high-quality output, large scenes, or scenes that produce artifacts with lower precision.

SuperCompressed

Field	Precision
position	f16 (3)
scale	u8 (3)
quat	u8 (4)
color & alpha	u8 (4)
shN	sh1 (sint5), sh2 and sh3 (sint4)

SuperCompressed favors file size, memory, and GPU memory. Use it when resources are tight, devices are lower-end, or performance is the priority.

Sog

Sog is for sog source data. It has the smallest size but can look blurrier. Prefer it when the source format is sog, when shN is absent, or when extreme scene scale is the main goal.

Preset List

Preset	Suitable Scenario
Max Quality	Highest visual quality requirement and strong user hardware.
Quality First	Large scenes, such as cities, that can render incorrectly at low precision. Requires reasonable hardware.
Performance First	Lower-end machines.
Extreme Performance 0	Very low-end machines or extremely large scenes.
Extreme Performance 1	Very low-end machines or extremely large scenes when the source data is sog. Prefer this over Extreme Performance 0 when the condition is met, because it can open larger scenes.

Max Quality

Quality First

Performance First

Extreme Performance 0

Extreme Performance 1

Custom Configuration

Presets cannot cover every scene. In practice, choose the closest preset as a starting point, then tune a small number of key parameters. For scene-specific tuning, import the scene into [EGS GS Viewer] and compare the result.

Parameter	Purpose	Recommendation
packType	Produces objects with different precision during parseSplats.	Use sog when there is no shN and the format is sog. Use super-compressed when memory or GPU memory is tight. Use compressed when high data precision is required.
packHighPrecisionEnabled	Enables high-precision data merging.	Determines the final data precision used for rendering. Usually enabled for compressed; evaluate per scene for sog.
precalculateEnabled	Enables spherical-harmonic calculation.	Enable it when the data has no shN, saving performance and GPU memory.
repackEnabled	Enables repack behavior.	A performance optimization for large scenes, usually used with sortMinDuration. It can improve performance by about 50%-100%, with extra GPU memory cost.
highPrecisionAttachEnabled	Enables a high-precision render attachment.	Consider enabling it when the scene shows ripple-like artifacts, or when quality is more important. It increases GPU memory usage.
normalizedFalloff	Enables normalized Gaussian falloff.	Most scenes show little difference. Do not enable it unless maximum quality is required.
preBlurAmount / blurAmount	Controls blur parameters.	Non-AA training usually uses 0.3 / 0; AA training usually uses 0 / 0.3. Other values are generally not recommended.
focalAdjustment	Adjusts splat spread scale.	2 is closer to the reference result.
detailCullingThreshold	Approximate detail culling.	Usually in [0, 4]. Setting it to 1 often has minimal visual loss; the performance gain depends on scene detail.
maxPixelRadius	Maximum screen-space radius covered by a Gaussian.	Default is 1024; recommended range is [128, 1024]. Too small can break the scene visually.
maxStdDev	Maximum standard deviation of Gaussian spread.	Should be between sqrt(5) and sqrt(9). Larger values cost more performance but improve quality; sqrt(8) is usually a practical quality/performance midpoint.
sortMinDuration	Minimum interval between sorting operations.	Usually used with repackEnabled. Common value is 16 * n, where n is no greater than 10.

normalizedFalloff Comparison