3DGS Code Reviewer

Problem

3D Gaussian Splatting implementations often run, but subtle details cause failures. Common risks include incorrect alpha compositing sorting, misused SH coefficients, negative scale values that explode rendering, duplicated opacity sigmoid application, and missing tile culling that can slow kernels by several times. Engineers debugging dark output, color artifacts, unstable training, or OOM often need to reason about graphics, CUDA, and training pipelines at the same time.

How It Works

The skill reviews code from a graphics-engineering perspective rather than generic linting. It checks the rendering pipeline: whether splats are sorted by depth rather than distance, whether T_i = T_{i-1} * (1 - alpha_i) and color accumulation are correct, and whether early termination is safe. It also flags high-frequency pitfalls such as C0 coefficients, scale constraints, and background alpha handling, then suggests performance fixes like tile culling, reducing SH degree, sorting every N iterations, and gradient checkpointing. The output should explain the mathematical or technical reason behind each issue and offer a concrete local replacement, not just label code as buggy.

Boundaries

It is best suited for inspecting 3DGS code snippets in PyTorch/CUDA-style implementations, especially rendering kernels, training loops, or loss functions. It is not a substitute for a full engineering audit and cannot guarantee coverage of every framework version. For JAX, TorchScript, or custom operators, combine its findings with framework version checks and actual profiling results.