Tutorials and paper breakdowns on 3D Gaussian Splatting — implementation details in PyTorch, rendering internals, and the research worth reading.
The map of the field, and where every other post fits on it.
How feed-forward 3DGS went from 64×64 synthetic objects to metric-scale scenes from a single photograph in under a second — Splatter Image, pixelSplat, MVSplat, DUSt3R, NoPoSplat, VGGT, Depth Anything 3 and SHARP, told as six assumptions being removed.
Reconstructing a scene in one forward pass, without per-scene optimization.
Two papers, one benchmark, opposite philosophies: probabilistic depth against plane-sweep cost volumes, the transplant experiment that settled it, and what each taught the field.
1.2 million metric-scale Gaussians from one photograph in under a second: the architecture, the two training tricks behind it, how to run it on a Mac, and what its license actually allows.
One prediction per pixel — a depth and a ray — and an unmodified DINO transformer beat VGGT by wide margins on pose and geometry. The representation, not the architecture, is the story.
How one transformer predicts camera poses, depth, point maps and 3D tracks in a single forward pass — architecture, training, results, and why it replaced the specialized pipelines.
An introduction to feedforward 3DGS: Splatter Image implemented from scratch in PyTorch, including per-pixel Gaussian decoding and the output-head initialization that makes training stable from step zero.
Reptile meta-learning applied to coordinate networks, so a NeRF adapts to a new scene from a single view in a fraction of the usual iterations.
Where 3DGS genuinely wins, where NeRF still holds an advantage, and why the honest answer in an interview is more nuanced than the headline.
Why splatting is fast, and how to make it faster.
Tile-based rasterization, the debt 3DGS owes to decades of GPU pipeline design, and where the time actually goes during a render.
Cutting the number of tiles each Gaussian touches, for faster rendering at identical image quality.
The engineering details that papers skip.
Writing an autograd.Function with a manual backward that recomputes rasterization weights on the fly, instead of storing them.
Bitwise-identical gradients, dramatically different VRAM. Why tile-parallel backprop is what puts 1080p and 4K training on a 24GB card.
What goes into the model, and what you can read back out of it.
Densification grows past a million Gaussians, and PSNR still drops when the initial point cloud is thin. An experimental look at why, and what to do about it.
High-quality depth maps out of a trained model, with no retraining, no regularizers and no changes to the pipeline.
Where this is all going.
Genie-class models generate every frame as you move; reconstruction builds geometry you can keep. What separates the two paradigms, who ships which, and why they now meet at 3D Gaussians.
Why the next wave of AI needs both a 3D representation of the world and a fast generative model, and what that means for what is worth learning now.
Want to truly understand 3D Gaussian Splatting—not just run a repo? My 3D Gaussian Splatting Course teaches the full pipeline from first principles in PyTorch only (no C++, no CUDA). You’ll learn initialization, rasterization, backward passes, training loops, and how to experiment with recent papers.
Explore the Course →I share practical posts on 3DGS, NeRF, PyTorch implementations, research breakdowns, and the engineering details that usually get skipped in papers.
Subscribe →