From 3D representations to practical, controllable media systems.

NeuralScene Bench

Before optimizing Gaussian Splatting, I needed to understand what advantage the representation actually offered.

I compared Splatfacto and Nerfacto TCNN under the same Nerfstudio parser, held-out split, image scale, camera treatment, and nominal 5K iteration budget. Splatfacto reached higher held-out reconstruction quality on both Bonsai and Garden.

The quality gap is also visible in held-out reconstructions. Fine foliage, thin geometry, and local texture structure remain closer to the reference in the tested Splatfacto outputs. I treat runtime throughput separately because repeated evaluation showed that evaluator FPS is implementation and execution-condition sensitive.

FastGS Training Efficiency

Once the representation looked promising, the next question was how expensive it was to create.

3DGS grows its scene representation through densification. More Gaussians can add detail, but they also increase training time, memory use, and checkpoint size. I tested whether all of that growth was necessary.

Bonsai tolerated an earlier densification stop extremely well, reducing training time by about 29% while maintaining, and in this run slightly improving, held-out quality. Garden preserved the cost reduction but paid a measurable quality penalty.

The useful result was not a universal recipe. It was evidence that training efficiency is scene dependent.

SplatStream Lab

Faster training still did not solve the delivery problem.

The Bonsai scene remained large after training. That shifted the question from optimization time to rate-distortion: how much of the representation could be removed or stored at lower precision before held-out quality collapsed?

A moderate operating point, 25% pruning with FP16 storage, reduced the experimental entropy-coded payload from 185.43 MiB to 70.04 MiB, a 2.65× reduction, while retaining 27.62 dB PSNR. More aggressive INT8 quantization caused a sharp quality cliff.

I treat that failure as evidence. It identifies the boundary of a simple compression strategy and motivates more structured coding approaches.

Ray-Scene Acceleration

Compression addresses what must move through a system. The next layer is what happens when scene data must execute efficiently.

I implemented brute-force traversal, a median BVH, a binned SAH BVH, and a uniform-grid baseline from scratch in C++17, then instrumented the actual geometric work performed per ray.

The binned SAH BVH reduced average triangle tests from 99,972 to 12.8 per ray and produced a 444× wall-time speedup in the clustered benchmark. The uniform grid became useful negative evidence because it was slower than brute force for this distribution.

4D Human Deformation Locality

Once Gaussian representations become articulated humans, efficient rendering is not enough. Small edits also need to stay controllable.

I audited pretrained HUGS digital humans to separate expected kinematic propagation from learned nonlocal coupling. The primary left-wrist probe was replicated across Seattle, Parkinglot, and Jogging, then a left-elbow z +10° generalization test was predeclared before inspecting its displacement result.

Across three independently pretrained checkpoints and two tested joints, all six joint × checkpoint diagnostic cells passed. There were 36 nested pose diagnostics in total. Replacing learned deformation weights with a subject-specific SMPL-derived K=6 target removed at least 99.7086% of the tested contralateral response. Selectively removing the learned perturbed-branch components eliminated 100% of the tested contralateral response in every nested pose test, with a global minimum removed-mass correlation of 0.9689.

The statistical unit remains the pretrained checkpoint, n=3. The two joints and 36 pose tests are repeated diagnostics within those models, not independent replications. I also keep failed assumptions and provenance corrections in the public research record instead of silently rewriting the experimental history.