跳到主要内容

Gaussian Splatting

Use for 3D gaussian splatting: train, edit, convert, serve.

Skill metadata​

SourceBundled (installed by default)
Pathskills/creative/gaussian-splatting
Version0.1.0
AuthorMerlin Agent
LicenseMIT
Platformslinux, macos, windows
Tags3dgs, gaussian-splatting, nerf, colmap, supersplat, gltf, novel-view-synthesis
Related skillsmanim-video, claude-design

Reference: full SKILL.md​

信息

The following is the complete skill definition that Merlin loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.

Gaussian Splatting (3DGS)

When to Use​

  • Reconstruct a 3D scene from photos/video (capture → COLMAP → train → edit → export)
  • Task mentions .ply / .splat / SPZ / KHR_gaussian_splatting assets, floaters cleanup, or SuperSplat
  • Asked to build a splat viewer/editor (Three.js, Babylon, self-hosted SuperSplat)
  • Don't use for plain point clouds (no photometric training), meshes, or NeRF training

What it is​

Explicit scene representation: millions of 3D Gaussians, each with 6 attribute groups — position μ (3 floats), rotation quaternion q (4), log-scale s (3), opacity α (1), and view-dependent color as spherical harmonics (SH degree 3 ⇒ 16 coefficients × 3 channels = 48). Covariance is never stored directly; it is composed Σ = R S Sᵀ Rᵀ so gradient descent always yields a valid positive semi-definite matrix. Scenes run ~100k Gaussians (SfM seed) → 1–5M after densification. No MLP: rendering is GPU rasterization at 100+ FPS; training 10–30 min on an RTX 4090 for a typical scene.

Pipeline (the 2026 production stack)​

  1. Capture — 20–50 photos / video frames; move slowly, parallax over every subject, avoid motion blur
  2. Poses — COLMAP or GLOMAP (SfM) ⇒ camera extrinsics + intrinsics + sparse point cloud
  3. Train — nerfstudio splatfacto / gsplat / inria gaussian-splatting / PostShot desktop
  4. Edit — SuperSplat (browser editor; a self-hosted build works — bundled build is a PWA with the spz/Niantic codec) — delete floaters, crop, segment
  5. Export — .ply (raw/portable) · .splat (PlayCanvas quantized) · glTF KHR_gaussian_splatting (Khronos 2026 RC) · OpenUSD 26.03 UsdVolParticleField3DGaussianSplat
  6. View — Three.js GaussianSplats3D, Babylon.js 9, Unreal, Cesium, Vision Pro

Rendering math (what makes it work)​

  • 3D Gaussian → 2D screen splat: Σ' = J W Σ Wᵀ Jᵀ (W = world-to-camera, J = Jacobian of perspective projection at the splat center — EWA-splatting first-order approx)
  • Per pixel, depth-sorted alpha compositing front-to-back: C = Σᵢ αᵢ Tᵢ cᵢ, Tᵢ = Π_{j<i}(1−αⱼ), αᵢ = opacityᵢ · exp(−½ dᵀ Σ'⁻¹ d)
  • Same volumetric equation NeRF integrates — hence equivalent quality, minus per-pixel MLP queries
  • Color at view direction d: evaluate SH basis at d, dot with learned coefficients (degree 0 = flat diffuse; degree 3 captures specular/sheen)

Training dynamics​

Photometric loss (L1 + D-SSIM), Adam, ~30k iterations. Adaptive density control every N iters:

  • clone: high positional gradient + small scale (under-reconstruction)
  • split: high gradient + large scale, into two smaller Gaussians (over-reconstruction)
  • prune: opacity below threshold → drop

Unconstrained params mapped through activations at render time (log-scale, logit-opacity, quaternion normalized) — the standard parameterization in every implementation.

Quickstart (this machine's environment)​

pip install nerfstudio gsplat            # Meta's gsplat CUDA rasterizer
ns-download-data example                  # or your own COLMAP'd capture
ns-train splatfacto --data <dir>          # 10–30 min on decent GPU

Executable proof: scripts/toy2d.py​

The 2D toy is now a runnable script (in-repo: skills/creative/gaussian-splatting/scripts/toy2d.py, commit cfd084c7; needs torch+numpy CPU only, ~30s):

python toy2d.py

Verified result (2026-10-02): step 199 mse 0.0043, 91.5% pixels within 0.2 of target, PASS — the formula chain (Σ=R diag(s²)Rᵀ → exp(−½dᵀΣ⁻¹d) → depth-sorted alpha blend → Adam) is executable evidence, not prose. Re-run after any touch of the skill's math sections.

SuperSplat (editor)​

SuperSplat is the browser editor for splat cleanup: floater removal, crop, segment, transform, then export .ply/.splat/SPZ/glTF. Options: the hosted app or a self-hosted build (it ships as a PWA — index.html, sw.js, spz codec bundle; any static server pointed at the dir works). Detection lesson: a Windows-side merlin install check can show NO-MERLIN when the install actually lives in the user's WSL — probe via wsl -e bash -lc before concluding absence.

Pitfalls​

  • Training needs CUDA; CPU-only = the 2D toy or viewer-only work (Three.js can render 2M+ splats client-side)
  • Baked lighting: 3DGS bakes illumination into SH — relighting needs GaussianShader-class variants
  • Memory: 1–5M Gaussians × 59 floats ≈ hundreds of MB/scene — quantize (.splat/SPZ) for web delivery
  • Secondary rays (reflections/refractions) are weak in vanilla 3DGS
  • COLMAP feature matching struggles on texture-less/reflective surfaces — add texture or masks

References​

  • Kerbl et al., SIGGRAPH 2023 "3D Gaussian Splatting for Real-Time Radiance Field Rendering" (the paper)
  • arXiv 2510.18101 — "From Volume Rendering to 3DGS: Theory and Applications" (tutorial survey)
  • IPOL 2025 "Gaussian Splatting: An Introduction" (1D→2D→3D pedagogical derivation)
  • ai-engineering.academy/learn/04-computer-vision/22-3d-gaussian-splatting (from-scratch code)
  • 3dgs.zip — exhaustive extension catalog