Since last time, I've replaced the video-to-3D pipeline with single-image Gaussian Splatting. A photo now generates in ~10 seconds on a Mac Mini, down from 15-120 seconds for video. You can compare the two approaches—video-based meshes: https://mukba.ng/discover/#videos vs image-based splats: https://mukba.ng/discover/#images
Benchmarks for SHARP between Mac Mini M4 and Macbook Pro M4: https://mukba.ng/blog/2026-01-21-sharp-apple-silicon-benchma...
I've also optimized rendering speed: I've chunked and zstd'd the output so it can start rendering as soon as the first packet comes in.
Why food? I started with Structure from Motion on video, and food was a good constraint because it doesn't move. People look terrible when they blink or shift mid-capture. Now that doesn't matter as much.