Two captures, one honest answer
Splat pipelines lose things quietly. A converter drops the spherical harmonics, a compressor clamps the extreme scales, a cleaner takes out more than you meant, a training run converges somewhere different. This tool puts two captures side by side and tells you what actually changed.
The visual half
Both captures are drawn in linked viewports: orbit, zoom or pan either side and the other follows, because a side-by-side comparison from two different angles tells you nothing. Each side draws the splat centres as coloured points rather than a full Gaussian render — that works at two million splats, works without WebGL 2, and is honest about what it is showing.
The numeric half
A full diff of the two captures: splat count, spherical-harmonic degree, file size, estimated GPU memory, bounding-box diagonal and centre, median opacity, near-transparent count, median scale, degenerate splats, and average colour. Every row shows both values and the difference.
The part most comparisons miss
Summary statistics can match perfectly while the two scenes sit in completely different places. So the tool also measures the geometry directly: it takes a sample of capture B, looks each splat up in a uniform grid built over capture A, and reports the mean, median and 95th-percentile distance to the nearest splat — in scene units and as a fraction of the scene diagonal.
Near zero means the same geometry. A few percent of the diagonal usually means a coordinate-convention mismatch, which the Gaussian Splat Coordinate Converter fixes. Splats with nothing inside the search radius are counted separately rather than being allowed to skew the average.
This is a sampled, one-directional nearest-neighbour distance, not a Hausdorff or Chamfer distance. Everything runs in your browser; neither capture is uploaded.
Tiny Online Tools







