All tools

Image → 3D

Early version

Photo → Gaussian Splat

Convert a photo into a Gaussian-splat scene and view it in WebGL.

The problem

Gaussian splatting is the new standard for photoreal 3D capture, but the toolchain is intimidating and GPU-bound. Students want to understand and play with splats without a CUDA pipeline.

What this does

An early bridge: the on-device depth point cloud is converted into a Gaussian-splat file (.splat / .ply) — each point becomes a small coloured gaussian — and rendered in a real-time WebGL splat viewer. It's an approximation, not a trained splat, but it makes the format tangible and shareable.

Try it

Loading tool…

What it does

This takes the coloured point cloud from the depth pipeline and writes a Gaussian-splat file: every point becomes an isotropic gaussian with a position, a small scale (from local point spacing), an opacity, and a colour stored as the splat format expects. The file is then loaded into an in-browser splat renderer so you can orbit it.

Why it's useful

  • Get hands-on with the .splat / gaussian .ply format without a training pipeline.
  • Produce a shareable splat from a single image for quick studies.
  • Understand what splats are before committing to the heavy capture workflow.

Honest limits (early / approximation)

  • This is not a trained 3D Gaussian Splatting reconstruction — it's a point cloud re-expressed as gaussians. No view-dependent shading, no densification.
  • Quality is bounded by single-image depth.

The real pipeline (roadmap)

Real single-image-to-splat models — Flash3D, Splatt3R, pixelSplat, LGM — predict full gaussian fields but need a GPU. Multi-image capture uses the original INRIA 3DGS / nerfstudio + gsplat training. This artifact documents that path while giving you something to hold today.

Built on