AI Agent Hub
Back to skills
3DGS Code Reviewer icon

3DGS Code Reviewer

Development Updated 2026.08.30

Paste the following prompt into your AI chat to install this skill:

Please follow https://skillhub.cn/install/skillhub.md to install @user_1e94615e/3dgs-code-reviewer-new into your AI assistant.

About this skill

Problem: “Runs” Does Not Mean Correct

3D Gaussian Splatting code touches sorting, alpha compositing, SH colors, scale/opacity parameterization, CUDA kernels, and training losses. Many bugs do not raise exceptions; they appear as black frames, flicker, edge noise, dim rendering, divergence, or slow training. Manual review can miss T_i accumulation, front-to-back order, tile culling, and near-plane clipping.

How It Works: A 3DGS-Specific Review Path

The skill reviews code as a graphics engineering specialist, focusing on:
- Rendering pipeline: depth sorting, alpha compositing, early termination, and SH coefficients
- Common pitfalls: double sigmoid opacity, negative scale, wrong SH basis, and 55+ known patterns
- Performance bottlenecks: missing tile culling, excessive CPU sorting, over-high SH degree, and other 5–10x costs
- Edge cases: near-plane clipping, low-alpha background SH, and camera-close Gaussians
It prioritizes correctness, performance, and style, and asks for the mathematical or engineering reason behind each issue while accounting for PyTorch/CUDA/JAX version differences.

Boundaries

Best for reviewing, debugging, and optimizing existing 3DGS rendering or training code, not for generating a complete project from scratch. If camera parameters, sorting logic, or loss definitions are missing, conclusions are limited.

Use Cases

  • Review a 3DGS rendering kernel to verify depth sorting, alpha compositing, and early termination.
  • Debug black frames, flicker, or dim color by checking SH, opacity, and scale parameterization.
  • Inspect the training pipeline and loss implementation to catch incorrect losses, gradient issues, or OOM.
  • Profile PyTorch/CUDA 3DGS code for tile culling, CPU sorting, and SH-degree overhead.

Best For

  • Graphics engineers building 3DGS renderers or SIBR viewers who need to audit CUDA kernel correctness.
  • Researchers tuning 3D Gaussian Splatting training stability who need to trace loss and parameterization bugs.
  • Backend or graphics engineers maintaining real-time large-scene rendering who need to cut tile, sorting, and SH overhead.
  • Code reviewers who need to check error-prone 3DGS alpha, SH, and scale implementations in pull requests.