2nd/10_Wiki/Topics/AI_and_ML/WebSplatter (3D Gaussian Splatting).md

---
id: wiki-2026-0508-websplatter-3d-gaussian-splattin
title: WebSplatter (3D Gaussian Splatting)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [WebSplatter, 3DGS Web Viewer, Gaussian Splatting Web]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [3dgs, gaussian-splatting, webgl, webgpu, neural-rendering]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: TypeScript
  framework: WebGL2/WebGPU
---

# WebSplatter (3D Gaussian Splatting)

## 매 한 줄
> **"매 NeRF 의 후계자 — 매 3D scene 을 millions of anisotropic Gaussian 으로 표현, web 에서 60fps real-time render"**. 2023 SIGGRAPH (Kerbl et al.) 의 3DGS paper 가 NeRF 의 slow ray-marching 을 differentiable rasterization 으로 대체하면서 photorealistic 3D capture 가 commodity 가 되었다. 2026 현재 WebSplatter / antimatter15-splat / SuperSplat / gsplat.js 등이 매 browser 에서 native handling — Apple Vision Pro / Quest 3 immersive content 의 default format.

## 매 핵심

### 매 3DGS 본질
- **Representation**: 매 scene = N개 (보통 1M-10M) 의 3D Gaussian — 매 Gaussian 은 (position μ, covariance Σ, opacity α, SH color coefficients) 로 parametrize.
- **Rendering**: 매 differentiable rasterization — Gaussian 을 screen space ellipse 로 project 후 alpha-blend (front-to-back).
- **Training**: 매 SfM (COLMAP) sparse cloud 로 init → photometric loss + densification/pruning heuristic 으로 optimize (~30min on RTX 4090 for one scene).

### 매 NeRF 대비
- **Speed**: NeRF 매 second-per-frame, 3DGS 매 100+ fps (1080p, RTX 30 class).
- **Quality**: 매 PSNR 비슷 (Mip-NeRF 360 기준 27.4 vs 27.5), 매 sharper detail in foreground.
- **Editability**: 매 explicit primitive — 매 Gaussian 단위 select / delete / transform 가능 (NeRF 매 implicit MLP, edit 어려움).

### 매 Web 배포 challenge
1. **File size**: 1M Gaussian × 60 byte/Gaussian = 60MB raw. 매 SOG / SOGS / .ply quantize 로 ~5-10MB 까지 압축.
2. **Sort cost**: 매 frame 마다 view-dependent depth sort 필요 (correct alpha blend). 매 GPU radix sort 필수.
3. **Browser GPU**: WebGL2 매 instanced rendering hack 필요, WebGPU 매 compute shader 로 native sort.

### 매 응용
1. Real estate 3D walkthrough — 매 phone 으로 capture, 매 WebSplatter viewer 로 share.
2. E-commerce — 매 product 360 turntable.
3. VFX previz — 매 set scan 후 Unreal/Blender 로 import.
4. Cultural heritage — 매 monument digitize.

## 💻 패턴

### 1. Splat 파일 load (.splat / .ply)
```typescript
// gsplat.js style loader
import { Scene, Splat, SplatLoader } from "gsplat";

const scene = new Scene();
const loader = new SplatLoader(scene);
const splat: Splat = await loader.loadAsync(
  "https://cdn.example.com/scene.splat",
  (progress) => console.log(`${(progress * 100).toFixed(1)}%`)
);
console.log(`Loaded ${splat.data.vertexCount} Gaussians`);
```

### 2. Gaussian sort (WebGPU compute)
```wgsl
// depth-sort.wgsl — view-dependent radix sort
@group(0) @binding(0) var<storage, read> positions: array<vec4<f32>>;
@group(0) @binding(1) var<storage, read_write> depths: array<u32>;
@group(0) @binding(2) var<uniform> viewProj: mat4x4<f32>;

@compute @workgroup_size(256)
fn computeDepth(@builtin(global_invocation_id) gid: vec3<u32>) {
  let i = gid.x;
  if (i >= arrayLength(&positions)) { return; }
  let clip = viewProj * vec4(positions[i].xyz, 1.0);
  // negate so far → small key, near → large (back-to-front blending)
  depths[i] = bitcast<u32>(-clip.z / clip.w);
}
```

### 3. Splat instanced rendering (WebGL2)
```glsl
// vertex shader — project 3D Gaussian to 2D ellipse
in vec3 a_quad;        // unit quad corner [-1,1]
in vec3 a_center;      // Gaussian μ
in vec3 a_cov_a;       // covariance row 0
in vec3 a_cov_b;       // covariance row 1
in vec4 a_color;       // SH degree-0 + opacity

uniform mat4 u_viewProj;

out vec2 v_uv;
out vec4 v_color;

void main() {
  vec4 clip = u_viewProj * vec4(a_center, 1.0);
  // project 3D covariance to 2D screen space (Zwicker EWA splatting)
  mat2 cov2d = projectCovariance(a_cov_a, a_cov_b, clip);
  vec2 axis = computeMajorAxis(cov2d);
  vec2 offset = a_quad.x * axis + a_quad.y * perpendicular(axis);
  gl_Position = clip + vec4(offset, 0.0, 0.0);
  v_uv = a_quad.xy;
  v_color = a_color;
}
```

### 4. Quantize (SOG format, 2025)
```python
# self-organizing Gaussian — 90% size reduction
import torch
from sogs import SOGCompressor

splat = torch.load("scene.pt")  # raw 3DGS state
compressor = SOGCompressor(
    position_bits=16,
    scale_bits=8,
    rotation_bits=8,
    sh_bits=6,
)
compressed = compressor.compress(splat)
compressor.write("scene.sog", compressed)  # ~6MB instead of 60MB
```

### 5. Train custom scene (gsplat library)
```python
# Modern training pipeline (Nerfstudio + gsplat backend)
from nerfstudio.scripts.train import main as train

train([
    "splatfacto",  # gsplat-based pipeline
    "--data", "data/my_scene",
    "--max-num-iterations", "30000",
    "--pipeline.model.cull-alpha-thresh", "0.1",
    "--pipeline.model.densify-grad-thresh", "0.0002",
])
# Export: ns-export gaussian-splat --load-config outputs/.../config.yml
```

### 6. React + WebSplatter component
```tsx
import { Canvas } from "@react-three/fiber";
import { Splat } from "@react-three/drei";

export function SceneViewer({ url }: { url: string }) {
  return (
    <Canvas camera={{ position: [0, 0, 5], fov: 50 }}>
      <Splat src={url} />
      <orbitControls enableDamping dampingFactor={0.05} />
    </Canvas>
  );
}
```

### 7. Edit / mask Gaussians
```typescript
// Remove Gaussians inside bounding box (e.g., remove a person)
function maskOutBox(splat: Splat, min: Vec3, max: Vec3) {
  const keep: number[] = [];
  for (let i = 0; i < splat.data.vertexCount; i++) {
    const p = splat.data.getPosition(i);
    if (p.x < min.x || p.x > max.x || p.y < min.y || p.y > max.y) {
      keep.push(i);
    }
  }
  return splat.subset(keep);
}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Static scene capture (real estate, VFX previz) | 3DGS — speed + quality 둘 다 |
| Dynamic scene (humans, fluids) | 4D Gaussian Splatting (4DGS) or NeRF derivatives |
| Mobile / low-bandwidth | SOG / SOGS quantize → < 10MB |
| AR/VR (Vision Pro, Quest) | 3DGS native — Metal / Vulkan compute path |
| Editable scene | 3DGS (explicit) > NeRF (implicit MLP) |

**기본값**: gsplat (training) + WebSplatter/antimatter15-splat (web viewer) + SOG (compression).

## 🔗 Graph
- 부모: [[3D Gaussian Splatting (3DGS)]]
- 변형: [[NeRF]]
- 응용: [[Spatial_Computing|Spatial Computing]]
- Adjacent: [[WebGPU]]

## 🤖 LLM 활용
**언제**: 매 photo-realistic real-world scene 을 매 web/AR 에 deploy 하고 싶을 때. 매 capture-once-view-anywhere workflow.
**언제 X**: 매 procedurally generated content (game asset)는 매 mesh + PBR 가 여전히 우월. 매 dynamic deformable mesh, 매 physical simulation.

## ❌ 안티패턴
- **Quantize 안 하고 60MB raw 배포**: 매 mobile 에서 OOM. 항상 SOG/SOGS 거쳐야.
- **CPU sort**: 매 1M Gaussian sort 매 frame 매 100ms+. 매 GPU radix sort 필수.
- **Sparse SfM init 생략**: 매 random init 매 converge 안 함. COLMAP step skip 금지.
- **Aggressive densification**: 매 split threshold 너무 낮으면 매 30M Gaussians 폭발 → OOM.

## 🧪 검증 / 중복
- Verified (Kerbl et al. SIGGRAPH 2023, gsplat library docs, antimatter15 reference impl).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — 3DGS web viewer + SOG compression + WebGPU sort patterns |