Hundreds of millions of Gaussians in real time, by turning them into pixel-sized points
A SIGGRAPH 2026 paper scales Gaussian splatting to enormous scenes with a deceptively simple trick: sample opaque, pixel-sized points from each Gaussian and splat them with 64-bit atomics. The thread wanted the one thing the paper doesn't quite sell — why splatting at all.
The trick, stated simply
Gaussian Point Splatting (ACM TOG / SIGGRAPH 2026, by Joris Rijsdijk, Christoph Peters, Michael Weinmann and Ricardo Marroquim) attacks the scalability wall of 3D Gaussian splatting. Instead of rasterizing each Gaussian the usual way, it samples pixel-sized, opaque points from the Gaussians and splats them to a framebuffer using 64-bit atomics. The genuinely hard problem they solve is statistical: "how many points should be splatted for a Gaussian or how they should be distributed to achieve the desired opacity." Paired with hierarchical frustum and occlusion culling to spread work evenly across millions of GPU threads, the method "renders hundreds of millions of Gaussians in real time," with the authors claiming the only differences from original Gaussian splatting are "slight noise and differences in aliasing."
The thread keeps asking "why," not "how"
Even an admiring thread circled one unanswered question — the case for splatting at all:
"While the outputs look nice, I'm not sure that I could distinguish them from a nice ray-traced scene. What I think I'm missing is the 'why?' of splatting." — a reader asking for the payoff
That's the right challenge to a technique that's been on HN's front page repeatedly without a crisp "here's what it unlocks." Others pulled at adjacent threads: one wondered when the first AAA game ships rendering its world this way instead of as traditional 3D, reaching back to Ecstatica, a 1994 game with an idiosyncratic ellipsoid-based renderer. Another asked how it stacks against mesh splatting, arguing triangles still represent sharp features better than Gaussians do.
A smaller lament was almost poetic: someone trying to learn the original 1990s point-splatting technique complained that Gaussian splatting has "completely overtaken the search results," burying the ancestor method it's named after.
A scaling win looking for its killer app
The contribution is real and specific: a clean way to push splat-based rendering into the hundreds-of-millions-of-primitives regime in real time. What the thread correctly identifies as missing is the demand-side story — the application where Gaussian splatting beats a good ray tracer so decisively that studios switch. The math now scales. The "why" is still, mostly, a render of a pretty scene and a promise.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Do transformers really need three separate projections? A study cut one and barely noticed
A new ablation paper tests whether attention needs distinct Q, K and V projections — and finds that tying key and value together costs almost nothing while halving the KV cache. The thread liked the result and roasted the notation.