A macro photographer captured a strawberry from 90 perspectives with 88 focus-stacked images each, trained into a Gaussian splat via slang-splat and COLMAP.
Key Takeaways
Shot on a Nikon Z8 full-frame with a Laowa 180mm macro lens, f/7.1, ISO 100, against a bluescreen with LED lighting for clean separation.
Each of the 90 camera positions used 88 focus-stacked images to overcome macro depth-of-field limits, yielding sharp detail across the full subject.
Training used slang-splat (github.com/MichaelMoroz/slang-splat) and COLMAP for reconstruction; result is 22.94 MB, hosted on SuperSplat.
Released CC BY 4.0 with attribution appreciated but explicitly not required, making it freely usable for training data, demos, or tooling tests.
Hacker News Comment Review
Commenters flagged a legal ambiguity: CC BY cannot formally waive the attribution requirement, so the “no attribution needed” language is informally generous but not strictly enforceable.
Dynamic lighting for Gaussian splats is an open question; current splats bake lighting at capture time, and relightable splats remain an active research area with no production-ready solution.
Apple’s ml-sharp (github.com/apple/ml-sharp) was raised as a related tool that generates splats from a single image in ~30 seconds on M1 Pro, useful for shallow VR parallax though it breaks under large viewpoint changes.
Notable Comments
@Tade0: notes Gaussian splats degrade gracefully into a “dreamy” blur rather than hard LoD cutoffs, framing it as an aesthetic property of the format.
@ovenchips: PlayCanvas creator, dryly observing the engine built for games in 2011 is now rendering strawberries in 2026.
