滚动擦洗的帧序列

钉住一屏，用滚动位置挑帧，而不是让时间轴自己走。没有任何东西自己在播 —— 读者就是那个播放进度条，手一停画面就停。

怎么做到的

Scroll position becomes a number in 0…1, and that number picks a frame: i = floor(p × N). That is the entire effect. Everything below is about making it not stutter.
The progress comes from where the section actually is, never from accumulating wheel deltas: p = clamp(-section.getBoundingClientRect().top / (section.scrollHeight − innerHeight), 0, 1). Accumulated deltas drift out of sync with the page the first time someone drags the scrollbar, uses Home/End, or lands on a deep link — and the drift is silent and permanent.
The section is a tall track with a position: sticky stage inside it. It never calls preventDefault on the wheel: the page scrolls normally the whole time, which is why the scrollbar, keyboard paging and momentum all still behave.
One paint per animation frame, not one per scroll event. A trackpad fires scroll far faster than the display refreshes; painting inline spends the frame budget drawing pictures nobody ever sees. Scroll only records the target index; a requestAnimationFrame callback draws it.
Repainting is skipped entirely when the index has not changed. With 40 frames across a 460% track, most scroll events resolve to the frame already on screen.
This sample synthesises each frame on a <canvas> so the file stays self-contained. With a real asset only the last line changes: ctx.drawImage(images[i], …) for a JPEG sequence, or video.currentTime = p * video.duration for a video. The scroll → progress math above it is identical.
For the <video> variant: read duration only after loadedmetadata, throttle the seeks to about 24–30 per second, and set muted playsinline preload="auto" with a poster. Seeking faster than that queues seeks the decoder cannot retire, and the picture lags the scroll by a growing margin.

参数

Scroll distance: 460% of the viewport
Frames: 40
Progress: −rect.top / (scrollHeight − innerHeight)
Paint budget: 1 per rAF, skipped if index unchanged
Video seek rate: 24–30 /s (video variant)
Canvas backing: min(devicePixelRatio, 2)

提示词

把这段丢给任意一个写代码的模型，就能从零长出这个效果。它把每个数值都写死了 ——「丝滑」「现代感」这种词每次生成出来的都不一样。

Build a scroll-scrubbed frame sequence in vanilla HTML/CSS/JS. Markup: a scroll container holding a track 460% of the viewport height, with a sticky stage inside it — the stage MUST be height: 100vh, not height: 100%, because 100% resolves against the 460%-tall track. Never call preventDefault on wheel and never accumulate wheel deltaY; derive progress from layout every time: p = clamp(-section.getBoundingClientRect().top / (section.scrollHeight - window.innerHeight), 0, 1). Map it to a frame with i = Math.min(N - 1, Math.floor(p * N)) for N = 40. Register the scroll listener passive; in the handler only store the target index and schedule one requestAnimationFrame; paint inside that callback and return early if the index equals the one already drawn. Back the canvas at min(devicePixelRatio, 2) and rebuild it on resize. If you use a <video> instead of a canvas, set muted, playsinline, preload="auto" and a poster, read duration only after the loadedmetadata event, set currentTime = p * duration, and throttle seeks to 24-30 per second. Show a small monospace readout of progress and frame number so the mechanism is visible. Everything must work when scrolling backwards, when scrolling very fast, and after a window resize.

源码

整个文件。上面那个样板跑的就是它 —— 存成 .html 打开就能用，不需要别的任何东西。

scroll-scrubbed-sequence.html

<!doctype html>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Scroll-scrubbed frame sequence</title>
<style>
  * { margin: 0; padding: 0; box-sizing: border-box; }
  html, body { height: 100%; }
  body { font-family: -apple-system, BlinkMacSystemFont, "SF Pro Display", sans-serif; }

  .scroller { height: 100%; overflow-y: auto; background: #04040a; scrollbar-width: thin; }
  .track { height: 460%; }
  /* 🩸 100vh，不是 100%：`.track` 高 460%，`height: 100%` 会把舞台撑成
     整条轨道那么高，画面居中到看不见的地方去。 */
  .stage {
    position: sticky; top: 0; height: 100vh;
    overflow: hidden; background: #04040a;
  }
  canvas { position: absolute; inset: 0; width: 100%; height: 100%; display: block; }

  .hud { position: absolute; inset: 0; pointer-events: none; color: #fff; }
  .eyebrow {
    position: absolute; left: 5%; top: 8%;
    font-size: 10px; letter-spacing: 0.22em; text-transform: uppercase; color: #ffffff66;
  }
  /* 把机制摆到台面上：读数就是 scroll → 帧号这条链路本身 */
  .readout {
    position: absolute; left: 5%; bottom: 9%;
    font: 500 11px/1.7 ui-monospace, SFMono-Regular, Menlo, monospace;
    letter-spacing: 0.06em; color: #ffffff8c;
  }
  .readout b { color: #ffd8a8; font-weight: 600; }
  .rail {
    position: absolute; right: 5%; top: 22%; bottom: 22%; width: 2px;
    background: #ffffff1f; border-radius: 2px;
  }
  .rail i { position: absolute; left: 0; top: 0; width: 100%; background: #ffb066; border-radius: 2px; }
  .cue {
    position: absolute; right: 5%; bottom: 9%;
    font-size: 11px; letter-spacing: 0.18em; text-transform: uppercase; color: #ffffff55;
  }
</style>

<div class="scroller" id="scroller">
  <div class="track">
    <div class="stage">
      <canvas id="cv"></canvas>
      <div class="hud">
        <div class="eyebrow">01 — Sequence</div>
        <div class="readout">progress <b id="rp">0.000</b><br>frame <b id="rf">1</b> / 40</div>
        <div class="rail"><i id="ri"></i></div>
        <div class="cue">scroll ↓</div>
      </div>
    </div>
  </div>
</div>

<script>
  const N = 40;                    // frames in the sequence
  const scroller = document.getElementById("scroller");
  const cv = document.getElementById("cv");
  const ctx = cv.getContext("2d");
  const rp = document.getElementById("rp"), rf = document.getElementById("rf"), ri = document.getElementById("ri");

  /* ---------- the sequence ----------
     A real build points this at an asset: 40 JPEGs blitted with drawImage, or
     one <video> driven by `video.currentTime = progress * video.duration`.
     Here each frame is synthesised so the sample stays a single file — the
     scroll → index math below is identical either way. */

  const rot = (p, ax, ay) => {
    let { x, y, z } = p;
    let c = Math.cos(ay), s = Math.sin(ay);
    [x, z] = [x * c + z * s, -x * s + z * c];
    c = Math.cos(ax); s = Math.sin(ax);
    [y, z] = [y * c - z * s, y * s + z * c];
    return { x, y, z };
  };
  const F = 620;                                   // focal length
  const proj = (p, cx, cy, k) => {
    const d = F / (F + p.z);
    return { x: cx + p.x * k * d, y: cy + p.y * k * d, d };
  };

  // A blocky figure: two boxes and four limbs, defined once in body space.
  const box = (cx, cy, cz, hx, hy, hz) => ({
    v: [[-1,-1,-1],[1,-1,-1],[1,1,-1],[-1,1,-1],[-1,-1,1],[1,-1,1],[1,1,1],[-1,1,1]]
      .map(([sx, sy, sz]) => ({ x: cx + sx * hx, y: cy + sy * hy, z: cz + sz * hz })),
    f: [[0,1,2,3,0,0,-1],[5,4,7,6,0,0,1],[4,0,3,7,-1,0,0],[1,5,6,2,1,0,0],[4,5,1,0,0,-1,0],[3,2,6,7,0,1,0]],
  });
  const HEAD = box(0, -63, 0, 16, 16, 16);
  const TORSO = box(0, -10, 0, 19, 30, 11);
  const LIMBS = [
    [{ x: 0, y: -46, z: 0 }, { x: 0, y: -38, z: 0 }, 6],       // neck
    [{ x: -19, y: -32, z: 0 }, { x: -40, y: -6, z: 10 }, 7],   // upper arm L
    [{ x: -40, y: -6, z: 10 }, { x: -52, y: 24, z: 24 }, 5.5], // forearm L
    [{ x: 19, y: -32, z: 0 }, { x: 44, y: -30, z: -6 }, 7],    // upper arm R
    [{ x: 44, y: -30, z: -6 }, { x: 62, y: -56, z: 4 }, 5.5],  // forearm R
    [{ x: -11, y: 17, z: 0 }, { x: -20, y: 54, z: 8 }, 8],     // thigh L
    [{ x: -20, y: 54, z: 8 }, { x: -14, y: 90, z: 24 }, 6 ],   // shin L
    [{ x: 11, y: 17, z: 0 }, { x: 24, y: 52, z: -4 }, 8],      // thigh R
    [{ x: 24, y: 52, z: -4 }, { x: 40, y: 84, z: 8 }, 6],      // shin R
  ];
  // Points FROM the surface TOWARDS the key light: up, left, and towards the
  // camera (screen y grows downwards, and the camera sits at negative z).
  const LIGHT = (() => { const l = { x: -0.48, y: -0.72, z: -0.50 }; const m = Math.hypot(l.x, l.y, l.z); return { x: l.x/m, y: l.y/m, z: l.z/m }; })();

  function renderFrame(i, w, h) {
    const t = N > 1 ? i / (N - 1) : 0;
    ctx.clearRect(0, 0, w, h);

    // stage: deep blue room, one warm key light falling from the upper left
    const g = ctx.createRadialGradient(w * 0.34, -h * 0.2, 0, w * 0.34, -h * 0.2, h * 1.7);
    g.addColorStop(0, "#1b3d8f"); g.addColorStop(0.42, "#0d1636"); g.addColorStop(1, "#04060f");
    ctx.fillStyle = g; ctx.fillRect(0, 0, w, h);

    // The beam gets a gradient ACROSS its width, not along its length. A flat
    // polygon leaves two hard diagonal edges that read as a crease in the wall.
    const ax = w * 0.16, bx = w * 0.62;
    const dx = bx - ax, dy = h, dl = Math.hypot(dx, dy);
    const px = dy / dl, py = -dx / dl;                    // unit perpendicular
    const half = w * 0.19;
    const mx = (ax + bx) / 2, my = h / 2;
    const beam = ctx.createLinearGradient(mx - px * half, my - py * half, mx + px * half, my + py * half);
    beam.addColorStop(0, "rgba(255,206,150,0)");
    beam.addColorStop(0.5, "rgba(255,206,150,0.17)");
    beam.addColorStop(1, "rgba(255,206,150,0)");
    ctx.fillStyle = beam;
    ctx.beginPath();
    ctx.moveTo(mx - px * half - dx / 2, my - py * half - dy / 2);
    ctx.lineTo(mx + px * half - dx / 2, my + py * half - dy / 2);
    ctx.lineTo(mx + px * half + dx / 2, my + py * half + dy / 2);
    ctx.lineTo(mx - px * half + dx / 2, my - py * half + dy / 2);
    ctx.fill();

    const yaw = -1.05 + t * 3.6;
    const pitch = Math.sin(t * Math.PI * 1.35) * 0.6 - 0.18;
    const k = (h / 300) * (1.08 + t * 0.5);
    const cx = w * 0.5 + Math.sin(t * Math.PI * 1.1) * w * 0.07;
    const cy = h * 0.58 - t * h * 0.17;

    const draw = [];
    for (const [b, tint] of [[TORSO, 0], [HEAD, 1]]) {
      const rv = b.v.map((p) => rot(p, pitch, yaw));
      for (const f of b.f) {
        const q = [rv[f[0]], rv[f[1]], rv[f[2]], rv[f[3]]];
        const n = rot({ x: f[4], y: f[5], z: f[6] }, pitch, yaw);
        if (n.z > 0.02) continue;                       // back face, skip
        const lit = Math.max(0, n.x * LIGHT.x + n.y * LIGHT.y + n.z * LIGHT.z);
        draw.push({ z: (q[0].z + q[1].z + q[2].z + q[3].z) / 4, kind: "f", q, lit, tint });
      }
    }
    for (const [a, b, r] of LIMBS) {
      const p0 = rot(a, pitch, yaw), p1 = rot(b, pitch, yaw);
      draw.push({ z: (p0.z + p1.z) / 2, kind: "l", p0, p1, r });
    }
    draw.sort((m, n) => n.z - m.z);                     // painter's algorithm

    for (const d of draw) {
      if (d.kind === "f") {
        const pts = d.q.map((p) => proj(p, cx, cy, k));
        const v = (d.tint ? 40 : 24) + d.lit * (d.tint ? 205 : 150);
        ctx.fillStyle = d.tint
          ? `rgb(${Math.round(v)},${Math.round(v * 0.88)},${Math.round(v * 0.68)})`
          : `rgb(${Math.round(v * 0.86)},${Math.round(v * 0.90)},${Math.round(v * 1.0)})`;
        ctx.beginPath(); ctx.moveTo(pts[0].x, pts[0].y);
        for (let j = 1; j < 4; j++) ctx.lineTo(pts[j].x, pts[j].y);
        ctx.closePath(); ctx.fill();
      } else {
        const a = proj(d.p0, cx, cy, k), b = proj(d.p1, cx, cy, k);
        const wd = d.r * k * ((a.d + b.d) / 2) * 2;
        ctx.lineCap = "round";
        // A cylinder is a barrel plus a highlight, not a flat stroke. Drawing the
        // core once more, thinner and shifted towards the light, is enough to
        // stop the limbs reading as noodles.
        ctx.strokeStyle = "#4b5470"; ctx.lineWidth = wd;
        ctx.beginPath(); ctx.moveTo(a.x, a.y); ctx.lineTo(b.x, b.y); ctx.stroke();
        const ox = LIGHT.x * wd * 0.22, oy = LIGHT.y * wd * 0.22;
        ctx.strokeStyle = "#e8dcc6"; ctx.lineWidth = wd * 0.55;
        ctx.beginPath(); ctx.moveTo(a.x + ox, a.y + oy); ctx.lineTo(b.x + ox, b.y + oy); ctx.stroke();
      }
    }
  }

  /* ---------- scroll → frame ---------- */

  let w = 0, h = 0, shown = -1, pending = 0;

  const paint = (i) => {
    if (i === shown) return;                            // the throttle that matters:
    shown = i;                                          // never redraw the same frame
    renderFrame(i, w, h);
    rf.textContent = String(i + 1);
  };

  const update = () => {
    const max = scroller.scrollHeight - scroller.clientHeight;
    const p = max > 0 ? Math.min(1, Math.max(0, scroller.scrollTop / max)) : 0;
    rp.textContent = p.toFixed(3);
    ri.style.height = (p * 100).toFixed(2) + "%";
    pending = Math.min(N - 1, Math.floor(p * N));
    // One paint per animation frame, not one per scroll event. A trackpad fires
    // scroll far faster than the display refreshes; painting inline burns the
    // frame budget redrawing pictures nobody ever sees.
    if (!raf) raf = requestAnimationFrame(tick);
  };
  let raf = 0;
  const tick = () => { raf = 0; paint(pending); };

  const resize = () => {
    const dpr = Math.min(2, devicePixelRatio || 1);
    w = cv.clientWidth; h = cv.clientHeight;
    cv.width = Math.round(w * dpr); cv.height = Math.round(h * dpr);
    ctx.setTransform(dpr, 0, 0, dpr, 0, 0);
    shown = -1;                                         // size changed, cache is stale
    paint(pending);
  };

  scroller.addEventListener("scroll", update, { passive: true });
  addEventListener("resize", () => { resize(); update(); });
  resize(); update();
</script>

会翻车的地方

A real video only scrubs smoothly if its keyframe interval is short. Encode at roughly one keyframe every 5–10 frames; a normally-encoded clip has one every few seconds, and every seek in between decodes from the last keyframe forward — it feels like the video is stuck, and the code looks fine.
iOS Safari will not seek a <video> that has never been played, and it ignores preload on a metered connection. A poster image plus a canvas sequence is the reliable path on phones; the video path needs a real fallback there, not a promise.
Scroll-scrubbed media has no state a reduced-motion user can opt out of — there is no animation to disable, only content that will not appear. Under prefers-reduced-motion: reduce, pin it to a representative frame and let the section scroll past normally.
460% of viewport height buys about four thumb-flicks on a phone with nothing else on screen. On small viewports cut the track, not the frame count — a short sequence over a long track just makes each frame linger.

出处

@派大鑫 (Douyin) · “这是我 Vibe Coding 的个人站” · 0:00 & 1:09