Macro Insects as 3D Gaussian Splats via Focus Stacking
The author builds macro 3D ‘Gaussian splats’ of insects by overcoming shallow depth of field using focus stacking. A scripted rotary capture with 111 viewpoints (1,776 images) and a Tamron 90mm setup feeds a pipeline of COLMAP, color/mask cleanup, and Postshot training. The final splats are viewable online, and a cluster fly model is released under CC BY.
Key Points
- Gaussian splats reconstruct view-dependent, photoreal 3D from photos via optimization, ideal for detailed subjects like insects.
- Macro depth of field is a major obstacle; focus stacking (16 images per view at f/18) produces fully sharp inputs.
- Capture setup: insect on rotary disk, camera tilts on a boom; 111 viewpoints, 1,776 images over ~4 hours with automated rail.
- Processing pipeline: batch focus stacking → COLMAP for camera poses → color correction/masking → Postshot training → minor cleanup.
- Results are published online, with a cluster fly model released under CC BY for free reuse with credit.
Sentiment
The overall sentiment of the Hacker News discussion is overwhelmingly positive and highly appreciative of the author's innovative work. While constructive feedback and technical discussions about current limitations for certain applications (like game development) and website usability were present, these were generally framed as areas for future research or improvement rather than criticisms of the presented achievement. The engagement demonstrated a strong interest in the technology's potential.
In Agreement
- The visual quality of the macro splats is amazing, incredibly clean, and impressive, particularly the fine detail on insect textures.
- The real-time rendering performance of the 3D models on various devices, including mobile phones, is highly impressive and indicates Gaussian splats are not expensive to render.
- The file sizes of the generated splat models are remarkably small.
- The author's decision to release a model under a Creative Commons (CC BY) license is greatly appreciated.
- The innovative combination of macro photography and Gaussian splatting is seen as a powerful and novel approach, with significant potential for educational visualization and other applications.
- The results are remarkable given the relatively modest hardware setup and effort involved in their creation.
Opposed
- Current Gaussian splats have significant limitations for typical video game development due to baked-in, view-dependent lighting (not true PBR), lack of rigging/animation support, and difficulty with editing and recomputing global illumination.
- The website's dark text on a dark grey background violates accessibility guidelines for contrast, making it unreadable for some users.
- The 3D models are difficult or "completely unusable" on mobile devices for some users, experiencing issues like locking to a southern pole view and erratic rotation, particularly noted with Firefox on Android.
- Artifacts like "ghost legs" and transparency can occur in the models, often due to tracking issues or subject movement during the long photographic capture process.
- Performing Structure-from-Motion (SfM) on heavily out-of-focus macro images, as would be required for modeling bokeh directly, is a significant technical challenge.