本文へ移動
cccskills
無料GitHub で公開

3d-web-experiences

Create 3D web experiences: scene setup, models, materials, lighting, animation, scroll-driven scenes, and performance budgets. Use when adding 3D to websites beyond basic demos.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md5.2 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

3D Web Experiences

A practical guide to shipping 3D on the web — product viewers, hero scenes, scroll-driven storytelling, configurators — with an emphasis on the asset pipeline and performance budgets that separate smooth experiences from janky tech demos.

Overview

Web 3D = WebGL (or WebGPU) + a scene graph library (three.js dominates) + 3D assets + disciplined performance work. The hard parts are rarely the rendering code: they're asset weight (a 50MB model kills mobile), draw calls (hundreds of meshes = slow), and frame budget (16.6ms at 60fps, less on low-end devices).

Rule of thumb: design the experience for the weakest device you support, then enhance — not the reverse.

When to use

  • 3D product viewers and configurators (color/material/parts switching).
  • Animated hero scenes and scroll-driven 3D storytelling.
  • Interactive data visualizations in 3D.
  • Choosing model formats and compression.
  • Diagnosing 3D performance problems.

Core concepts

  • Scene graph. Scene → meshes (geometry + material) → lights → camera. Everything is a positioned node; groups organize parts.
  • Geometry. BufferGeometry of triangles. Polygon count is a budget: hero models 50–200k tris desktop, far less for mobile; background props much less.
  • Materials. PBR (physically based rendering): MeshStandardMaterial with color/roughness/metalness maps. Fewer unique materials = fewer shader programs = faster.
  • Lighting. Key + fill + rim, or environment maps (IBL) for realistic reflections. Real-time shadows are expensive — one shadow-casting light max on mobile, or bake shadows into textures.
  • glTF. The standard interchange format. Prefer .glb (binary) with Draco mesh compression and KTX2/Basis texture compression.
  • Animation. requestAnimationFrame loop with delta-time (never fixed steps); THREE.AnimationMixer for skeletal/glTF animations; lerp for smooth interpolation.
  • Scroll-driven scenes. Map scroll progress → camera path / object transforms. Use a smoothed (lerped) scroll value, never raw scroll events, to avoid judder.
  • Draw calls. Each mesh+material = at least one draw call. Merge static geometry, use instancing (InstancedMesh) for repeated objects (grass, bolts, stars).

Practical workflow

1. Prepare assets first.

# Typical pipeline: Blender -> glTF -> compress
# - Apply transforms, merge materials, bake AO/shadows where possible
# - Export .glb, then compress:
gltfpack -i model.glb -o model-packed.glb -cc   # meshopt compression

Target: hero model under 3–5MB, textures ≤2048px (1024 for mobile), KTX2 compressed.

2. Scaffold the scene.

import * as THREE from 'three';
import { GLTFLoader } from 'three/addons/loaders/GLTFLoader.js';
import { OrbitControls } from 'three/addons/controls/OrbitControls.js';

const renderer = new THREE.WebGLRenderer({ antialias: true, powerPreference: 'high-performance' });
renderer.setPixelRatio(Math.min(devicePixelRatio, 2));
const scene = new THREE.Scene();
scene.environment = await loadEnvironment(); // RoomEnvironment or HDRI for PBR reflections

3. Load with feedback. Show a real loading UI with progress (onProgress bytes) — 3D assets are heavy; a blank canvas reads as broken.

4. Animate correctly.

const clock = new THREE.Clock();
function tick() {
  const dt = Math.min(clock.getDelta(), 0.05); // clamp tab-switch spikes
  controls.update();
  mixer?.update(dt);
  renderer.render(scene, camera);
  requestAnimationFrame(tick);
}

5. Budget and measure. Use the browser's frame profiler + renderer.info (draw calls, triangles, geometries). Set budgets: e.g., <150 draw calls, <500k tris, <10MB assets for a marketing hero.

6. Degrade gracefully. WebGL2 check → fallback to a static image or video. Respect prefers-reduced-motion (pause auto-rotation/camera moves).

Common pitfalls

  • Uncompressed assets. Shipping raw 40MB FBX/OBJ exports. glTF + meshopt/Draco + KTX2 routinely cuts 80–95%.
  • Draw call explosion. A model with 300 separate meshes renders 300+ draw calls. Merge, instance, or atlas.
  • Pixel ratio blindness. setPixelRatio(devicePixelRatio) on a 3x phone = 9x pixels. Cap at 2 (or 1.5 for heavy scenes).
  • Shadow abuse. Multiple shadow-casting lights or huge shadow maps tank mobile GPUs. One directional light with tight shadow camera, or baked shadows.
  • Memory leaks. Geometries, materials, and textures not disposed on scene teardown (geometry.dispose(), material.dispose(), texture.dispose()). SPAs navigating between 3D views leak fast.
  • Z-fighting. Coplanar surfaces flicker — offset with polygonOffset or tiny positional gaps.
  • Blocking the main thread. Parsing a big glTF synchronously freezes the page. Load async, show progress, consider Web Workers for heavy prep.
  • No loading state. Users stare at black. Always: skeleton → progress → reveal.
  • Ignoring color management. Set renderer.outputColorSpace = SRGBColorSpace and use sRGB textures correctly, or colors look washed out/wrong.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Master 1Password: vaults, Watchtower, passkeys, SSH agent, CLI, and family/team administration. Use when getting full value from 1Password personally or administering it for others.

日本語の概要は準備中です。原文の説明を表示しています。

aicodedecode/awesome-muse-skills172026年10月10日 更新

Create 3D visuals with modeling, texturing, lighting, rendering, and optimization for web and product.

日本語の概要は準備中です。原文の説明を表示しています。

aicodedecode/awesome-muse-skills172026年10月10日 更新

Writing abstracts that get papers read — structured content, the 5-sentence core, and journal-specific constraints.

日本語の概要は準備中です。原文の説明を表示しています。

aicodedecode/awesome-muse-skills172026年10月10日 更新

Learn effectively from courses and academies: choosing programs, studying actively, and converting courses into skills. Use when investing time/money in structured learning.

日本語の概要は準備中です。原文の説明を表示しています。

aicodedecode/awesome-muse-skills172026年10月10日 更新

Audit designs for accessibility with WCAG checklists covering color, type, focus, motion, and content.

日本語の概要は準備中です。原文の説明を表示しています。

aicodedecode/awesome-muse-skills172026年10月10日 更新

Run ABM programs — target account selection, personalized campaigns, sales orchestration, and account-level measurement.

日本語の概要は準備中です。原文の説明を表示しています。

aicodedecode/awesome-muse-skills172026年10月10日 更新

aicodedecode のスキルをすべて見る

このスキルの問題を報告する