Show HN: Gaussian Splat of a Strawberry

Original Article ↗ Hacker News Discussion ↗

Capture Setup & Workflow

  • Strawberry splat was shot from ~90 viewpoints, each with 88-image focus stacks (≈7,920 photos) using a full‑frame mirrorless body, 180mm macro, LED lighting, and bluescreen.
  • Capture takes ~20 minutes thanks to a fast camera and a motorized focus rail plus rotary disk.
  • Photographer uses a computer only for setup, test stacks, and triggering rotation, then records to memory card during the main shoot.

Macro, Microscopy & Focus Stacking

  • Several commenters want to combine microscopy and Gaussian splats; macro attempts at 2× and 5× magnification highlight depth‑of‑field becoming extremely shallow.
  • Focus stacking is used to avoid training on blurry areas; without it, the optimizer would reproduce out‑of‑focus blur.
  • Some speculate that focus cues could be integrated into training instead of pre‑stacking; references are made to “DoF‑Gaussian” and more advanced camera models.

Artifacts, Quality & Visual Characteristics

  • The strawberry’s bottom looks “rotten” or missing; this is attributed to incomplete capture/occlusion from the mounting hardware.
  • Interior and underside look wrong or translucent; identified as reconstruction artifacts rather than true material properties or refraction.
  • Commenters note Gaussian splats look great at normal distances but break down when you get very close or into unseen regions, producing dreamy, foggy, or painterly degradation that some find artistically appealing.

Performance, Data Size & Compression

  • Web viewer runs smoothly even on some mobile devices; others report GPU load, low FPS, or crashes (especially on mobile Firefox/Safari or when WebGL is unavailable).
  • Splats can be large: examples include tens to thousands of megabytes, contrasted with smaller polygonal models of similar subjects.
  • Compressed formats (e.g., SOG, SPZ) with LOD support are discussed as partial answers to size concerns.

Tools, Formats & Ecosystem

  • Various training tools are mentioned: PostShot, Slang‑Splat, LichtFeld, KIRI Engine, and others; VRAM limits influence tool choice.
  • Some want more approachable workflows; camera‑pose tracking and software UX are cited as barriers.
  • Apple’s “ml‑sharp” model, generating splats from a single image, is noted as promising but heavy (multi‑GB weights, high VRAM use).

Applications, Animation & Future Directions

  • People imagine uses in games, sim racing, Google Maps/Earth–style navigation, concerts, and video stabilization/“decropping.”
  • Animated or 4D splats are an active interest; commenters debate how far one can go toward skeletal‑style deformation, dynamic lighting, and relighting.
  • There is speculation about generative models that could synthesize splats from prompts, possibly merging diffusion, NeRFs, and 3DGS.

Licensing & Access

  • Licensing language (CC BY but with optional attribution in the description) is debated; some see it as effectively a waiver rather than strict CC BY.
  • Some users report needing to loosen script/host blocking to see more than a blurred thumbnail.