Gaussian splatting in 3D and 2D
3DGS (Kerbl et al., 2023): per-scene 3D Gaussians; project to the image plane, sort by depth, alpha-composite near to far; optimise against photographs; adaptive density control (clone, split, prune).
C(x) = Σi ci αi(x) Πj<i (1 − αj(x))
| Work | Combine | Placement | Carries |
|---|---|---|---|
| 3DGS | Alpha composite | Per scene | RGB |
| GaussianImage | Summation, 8 parameters | Per image | RGB |
| Image-GS | Not on abstract | Per image, error-guided | RGB |
| Instant-GI | Not on abstract | Network then fine-tune | RGB |
| This project | Softmax sharing | Unrolled clustering and Adam, any count | 384-d DINOv3 features |
Key numbers (abstracts): 3DGS at least 30 fps at 1080p; GaussianImage at least 3x less memory and 5x faster fitting than INRs, 1500-2000 fps; Instant-GI training time down by up to one order of magnitude; Image-GS 0.3K MACs per pixel.
Defence answer. “Image-plane Gaussians exist for RGB. I use them for a frozen DINOv3 feature field, solve the features in closed form, share locations by softmax and predict placement with an unrolled clustering-and-Adam model that serves any count from 16 to 256.” Caveat: split and merge arms were tested on 100 to 200 images with tuned settings.
Not to confuse: 2DGS (Huang et al.) is surfels in a 3D scene.