跳到正文
cccskills
免费在 GitHub 公开

3d-orbit-inspect-demo

Build a Three.js object viewer with a gentle camera lean on hover and a click-to-inspect mode. Visitors can rotate and zoom the object, then return it to its original position using mouse, touch, or keyboard controls.

中文简介正在准备中。当前显示英语简介。

原文查看原文说明

Make a 3D object lean around on hover and open a full inspect view on click. Hover is a damped, bounded camera lean; click lifts the piece off its bench into the hand, where you drag to turn it, coast, dolly with wheel or pinch, and put it back with Esc, click-away or a Close button. Use for product, object, joinery and portfolio pieces in Three.js where a hover orbit and click-to-full-view must not fight each other; use webgl-3d-object for a floating hero mesh with no inspect mode, and 3d-virtual-tour for a camera path through a place.

查看安装方法

适用场景

  • Add interactive inspection to a 3D product showcase
  • Let visitors rotate and examine joinery details
  • Add touch and keyboard controls to portfolio objects
  • Resolve conflicts between hover movement and clicks

包含的文件(6)

  • SKILL.md9.7 KB
  • agents/openai.yaml320 B
  • demo.html36.6 KB
  • preview-full.jpg144.1 KB
  • preview.jpg189.3 KB
  • PROMPT.md4.3 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

3D Orbit and Inspect

Three.js · WebGL. No OrbitControls, no custom shaders.

The mechanism: hover leans the camera a few degrees around the object with a gain that rises only while a raycast hits it, and click hands the object (not the camera) to a second render layer that lifts it to a fixed spot in front of the lens. The two modes share no state except one rule: the lean freezes whenever a click is possible or the piece is in the hand. Remove that rule and the target slides out from under the click.

Nearest skills: webgl-3d-object (a floating mesh, no inspect), 3d-virtual-tour (its "orbit inspection" is one mode of a walkthrough; the camera itself flies). Reach for this when there is one object and two gestures on it.

Reuse demo.html: the joint, the room, the lean, the lift and the return are all in one module. Copy updateLean, updateHand, openInspect, closeInspect, the pointer block and draw. Swap the meshes for your own object; nothing else names the joint except FROM, RR and Q_BENCH.

What is and is not in the source

Kibori (kibori.html, section "8b. Inspect — pick an item up") has click-to-lift with drag-to-turn, an idle wheel, a "Look closer" cue and a zoom-in cursor on hover, and close by click-away, Esc, scroll or its Close button. It has no hover orbit: the camera only leans a few degrees as scroll dollies it between stations. It also has no free zoom, because wheel is the page scroll and closes the piece. So:

  • Ported unchanged: the item is lifted, the camera is not flown; the layer-1 second pass; the 1-(1-k)^4.4 lift ease; 1.0 s in, 0.60 s home, 0.28 s fast close; the way home straightens before it descends (sstep(0,.36) then sstep(.24,1)); drag 0.0085 / 0.0065 rad per px; spin damping exp(-2.4 dt); idle turn 0.26 rad/s after 2.2 s; click = under 6 px and 600 ms; the band between the top bar and the caption; the cue pill. Not ported: Kibori's hand-written tone-mapping pass and blur of the room behind the piece; the demo uses ACESFilmic and a flat dim.
  • Built here in the Kibori manner: the hover lean, the wheel/pinch dolly, arrow-key and +/- control, focus return, the live region, the touch sway.

Rules, each with the failure it prevents

  1. Freeze the lean on pointerdown and while inspecting. Otherwise the camera keeps easing during a 100 ms click and the raycast at pointerup misses a piece that moved 6 px. Raycast again at pointerup, on the same layer the piece is on.
  2. Gain from a raycast, target from the pointer relative to the piece. g rises to 1 only while the ray hits (tau 0.30 s); the target is clamp((pointer - pieceCentre)/(0.18 W), -1, 1). Screen-centre pointer maps make the piece drift whenever the mouse is anywhere on the page.
  3. Bound and damp. Azimuth ±0.30 rad, elevation ±0.09 rad, 1 - exp(-4.5 dt). Unbounded lean turns the object edge-on; an undamped one jitters with a 1 px mouse.
  4. Hover is the last pointer, not a media query. (hover:hover) lies on hybrid laptops and headless Chrome. Read pointerType on pointermove; touch gets a slow sway (±0.6 of the bound at 0.55 rad/s) so the depth is still shown without hover.
  5. Click and drag are one pointerdown. Under 6 px and under 600 ms opens; anything else is a drag and never opens. In the hand the same 6 px test separates "turn" from "click away". Set touch-action:none on the canvas only while inspecting, so the page can still scroll when it is not.
  6. Lift by layer, not by clone or .visible. The piece moves to layer 1, the room renders on layer 0, a fullscreen quad at opacity 0.80 e goes between, clearDepth(), then layer 1. Lights need layers.enable(1) or the lifted piece renders black. The scene needs background = null and manual clear(); a scene background re-clears pass two.
  7. Perspective-true hand-off, no pop. Interpolate in camera space: x/z and y/z linearly, then multiply by the lerped depth. A world-space lerp slides in depth and the piece hardly moves sideways for the first 100 ms. The world "from" pose is stored, never re-read from the moving object, so a cancelled lift lands on the exact bench pose (verified equal to four decimals after Esc at 150 ms).
  8. Hold the camera in the hand; never chase it. The hand pose is defined in camera space; if the lean keeps easing while inspecting, the piece swims. Freeze it, resume on close.
  9. Sit at the band's middle, sized by it. fit = min(0.92 bandH, 0.88 W) / (2 R ppu) with ppu = H / (2 D tan(fov/2)), D = 1.0 m. Measure the caption's settled top (top - 14 - 40) before it fades in; a piece sized for the whole viewport runs under the caption.
  10. Close is its own clock. On close, e0 = r0 = insOut(k) and k restarts from 1; a piece turned further takes up to 1.5x longer (span). Closing from the middle of the lift must not restart from zero.
  11. Zoom is eased and bounded. 0.70 to 1.45 of fit, wheel exp(-dy 0.0012), ease 1 - exp(-8 dt), pinch by finger distance ratio. Past 1.5 the piece overlaps the caption; below 0.7 it is smaller than it sat on the bench. Reset to 1 on close.
  12. Pitch is bounded and relaxes. 0.25 to 1.30 rad, home 0.72, pull home at exp(-0.35 dt). A flat timber turned past ~1.3 is a hairline.
  13. Keyboard and focus. A real button opens (Inspect the joint); Enter and Space work. On open, focus the Close button and set the caption inert = false; Tab is trapped to it; arrows turn 0.14 rad, + - zoom 15%, Esc closes. Return focus to the opener after the page copy is visible again: the copy is visibility:hidden while inspecting and focus() on a hidden element silently does nothing.
  14. A live region, not aria-live on the dialog. Announce "Inspecting ..." on open and "Back on the bench." on close through a separate role=status node, cleared then set 30 ms later so a repeat is announced.

Numbers

thingvalue
lens, rest cameraFOV 34, elevation 0.40 rad, 1.50 m (portrait: 0.92 / aspect)
hover leanaz ±0.30, el ±0.09, damping 4.5/s, gain tau 0.30 s
lift1.0 s, 1 - (1-k)^4.4, arc sin(pi e) * 0.06 m, hand at D = 1.0 m
home0.60 s (0.28 s fast), rotate sstep(0,.36), descend sstep(.24,1)
dim behindopacity 0.80 e, same colour as the room
lamp on the piecewarm point, 3.0 e, up 0.5 / left 0.6 / toward eye 0.35 m, layer 1 only
turn0.0085 / 0.0065 rad/px, damp 2.4/s, idle 0.26 rad/s after 2.2 s
budgetdt clamp 1/30, DPR cap 2, shadow map static (2048), render only when something moved

Room

Dark room #0a0806, FogExp2 0.42 (materials that must not go murky set fog:false), key 0xfff2df 2.0 from upper left with the only shadow, fill 0xaec7df 0.76, rim 0xffdeb4 0.72, hemisphere 0.19, a low ember point 0xffb46a. Environment: a PMREM of three emissive cards in a black box at environmentIntensity 0.55. Ember #dbad64 is the only accent (cue ring, kicker, focus).

Reduced motion, hidden, cost

  • prefers-reduced-motion: no lean (the piece rests at the three-quarter bench pose), the lift is instant (rate = 1e3), no idle spin, no cue slide. Drag, keys and zoom still work: they are direct manipulation.
  • visibilitychange cancels the loop; resume resets last so the first frame does not integrate the pause.
  • Cost, measured on an Apple GPU in headless Chrome (Metal): about 0.3 ms of CPU per rendered frame, 3 draw calls, about 250 triangles; 8.6 ms per synchronous frame in the two-pass lifted state at 1440x900 and 14.2 ms at 2880x1800 (DPR 2). Idle at rest renders zero frames. The lever is DPR; the shadow map is static and free, and triangle count does not matter.

Expensive gotchas

  • The lifted piece is black. Lights are layer-filtered by the camera; light.layers.enable(1).
  • The room disappears when the piece appears. scene.background re-clears in the second render; use setClearColor and autoClear = false.
  • A shader compile hitch on the first click. The lit-piece pass has a different light set (the lamp). Draw both states once at load.
  • Shadow of a piece that has left. shadowMap.autoUpdate = false, then needsUpdate when the piece changes layer, or a ghost shadow stays on the bench.
  • Cue and cursor stay after a click. Clear both on open; zoom-in cursor only on hit, grab / grabbing in the hand.
  • Focus lost on close. See rule 13.
  • Object3D.clone serialises userData. If you clone instead of switching layers, empty userData first (render targets and polygons in there stall the clone).

Verify

Load at 1440x900 and 390x844; hover at three pointer positions (lean left, right, none); press and release on the piece (opens), press-drag-release across it (does not open); drag, wheel, pinch in the hand; click-away, Esc and Close each return the piece to the exact bench pose, including when cancelled 150 ms into the lift; Tab reaches the button, Enter opens, focus lands on Close and returns; reduced motion shows the rest pose and an instant lift; console is clean.

Extracted from KIBORI 木彫, a dark WebGL workshop page where every bench piece can be picked up and turned in the hand (section 8b). The hover lean and dolly are new, built to the same easing and HUD.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Build woven fabric in Three.js with visible threads, realistic sheen, seams, and thick hems. The skill covers folds that preserve texture proportions, light passing through thin cloth, and variations in dye and wear.

中文简介正在准备中。当前显示英语简介。

  • Creating backlit curtains
  • Differentiating linen, silk, and wool
  • Adding hems and stitches for close-up views
MengTo/Skills69492026年10月18日 更新

Build falling leaves in Three.js and similar 3D scenes, with visible front, edge, and back views as they tumble and drift. Coordinate wind, scene depth, camera-aware recycling, and performance checks for a convincing autumn effect.

中文简介正在准备中。当前显示英语简介。

  • Adding falling autumn leaves to a 3D scene
  • Coordinating leaf tumble, sideways drift, and wind
  • Managing leaves around buildings and camera movement
MengTo/Skills69492026年10月18日 更新

Add coordinated seasonal transitions to an existing 3D scene. Blend foliage, lighting, sky, ground, snow, and particles from one shared state while preserving camera controls and interactions, so the same place remains recognizable throughout spring, summer, fall, and winter.

中文简介正在准备中。当前显示英语简介。

  • Add a four-season selector to an outdoor 3D scene
  • Show seasonal changes in an architectural walkthrough
  • Coordinate foliage, sunlight, sky, and ground appearance
MengTo/Skills69492026年10月18日 更新

Create or integrate detailed 3D models with smooth outlines and believable edges for close-up views. Prepare lower-detail versions for distant viewing, prioritize loading, and measure rendering costs in the actual scene.

中文简介正在准备中。当前显示英语简介。

  • Creating a detailed 3D model for close-up views
  • Replacing visibly faceted curves with smoother geometry
  • Preparing detailed architectural assets for rendering
MengTo/Skills69492026年10月18日 更新

Prepare detailed 3D materials that remain clear in close-up views. The skill guides texture resolution, physically based material maps, surface mapping, filtering, and progressive loading while checking download size and GPU memory use.

中文简介正在准备中。当前显示英语简介。

  • Prepare wood, stone, and fabric materials for close-up views
  • Fix texture seams, grain direction, and normal map orientation
  • Preserve floor and foliage detail at oblique or distant views
MengTo/Skills69492026年10月18日 更新

Create realistic metal surfaces in Three.js by combining structured reflections, studio lighting, scratches, fingerprints, and beveled edges. Includes polished, brushed, and aged finishes for objects such as tools, rings, and metal product parts.

中文简介正在准备中。当前显示英语简介。

  • Fixing metal that looks like gray plastic
  • Creating polished or brushed metal finishes
  • Adding patina and worn edges
MengTo/Skills69492026年10月18日 更新

查看 MengTo 的全部技能

报告此技能的问题