Lab

Gaussian Splats

This lab targets a paper-first Image-GS workflow: anisotropic 2D Gaussians are initialized from image structure, optimized against sampled pixels, progressively densified from reconstruction error, and organized into a BSP shell hierarchy for fast random-access inference.

InputUpload, drop, or paste an imageNew input cancels the active run, drops the current GPU session, and rebuilds the pipeline for the new source.

Source

No image loaded

The uploaded source image appears here.

Fitted Gaussians

Idle · Step 0/50,000

Generate a run to fit the image with anisotropic Gaussians.

Residual Heatmap

GPU residual energy is visualized here at checkpoint cadence.

Residual error appears here once optimization starts.

Checking WebGPU support…

Run Progress

Load an image to initialize a paper-first Image-GS optimization run.

ETA

—

Recent Improvement

0.0% / 1k steps

Overall Progress0/50,000 steps

Next Densify—

Status

Idle

Densify Milestone

0/10

Gaussians

0/8,000

Avg Step Ms

Current Loss

0.0000

Best Loss

0.0000

Step

0/50,000

Densify Events

Mean Error

0.0000

Hot Tiles

Avg Step Ms

GPU Readback Ms

BSP Leaves

Avg Shell Gaussians

0.0

Inference Error

0.0000

Device

Unavailable