Macro Gaussian Splats (danybittel.ch)

Macro Gaussian Splats demonstrates a practical pipeline for making high-fidelity, free-viewpoint 3D “photographs” of macro subjects by combining Gaussian splats (view-dependent, blurry ellipsoids optimized like a neural model) with traditional macro photography techniques. The author solved the core problem—extremely shallow depth of field in macro shots—by using focus stacking (16 images per stack), shooting at f/18, and automating capture with a rotary disk, a WeMacro focus rail, and a boom-mounted camera tilt to collect 111 perspectives (1,776 source photos) across slightly more than a hemisphere. Camera poses were recovered in COLMAP, images were color-corrected and masked, and Postshot was used to train the splat representation; final models needed only minimal retouching. This work is significant because it fuses classic macro optics with modern neural rendering, enabling highly detailed, view-consistent 3D insect models that preserve fine hair and texture. Practical insights include the trade-off between aperture and diffraction, the necessity of minimizing photos per stack to keep capture times reasonable, and hardware limits (e.g., DSLR buffer slowdown). The project page (superspl.at) showcases the results, and a cluster fly splat model has been released under a CC BY license for reuse in commercial or non-commercial projects with attribution.