低多边形日式车站立体场景对比

Build a single self-contained HTML file: an interactive low-poly 3D diorama of a rural Japanese railway station in spring, using Three.js (ES modules from a CDN). No external models or images — generate all geometry from primitives in code and paint textures (grass, signs) on <canvas>.

Style: a floating diorama block (grass top, layered earth and stone sides) on a soft pastel sky gradient. MeshToonMaterial with a 3-step gradient map, dark ink outlines (OutlineEffect), soft directional shadows, bright saturated palette.

Scene:
- A single railway track across the block, running into green tunnel hills with concrete portals at both ends.
- Concrete platform with a yellow tactile strip, a shelter with a corrugated roof, benches, a hanging clock, a small station building with a teal gable roof and a "春風駅" sign, a station name board (はるかぜ / Harukaze, with neighbouring stations), vending machines, bins, lamps, fence.
- A road crossing the track with a level crossing: striped poles, crossbuck, red flashing lamps, barrier arms; road markings and "止まれ".
- Utility poles with sagging wires, overhead catenary, small village houses, a vegetable patch, bushes, and cherry trees with puffy pink blob crowns and fallen petals painted on the grass.

Animation:
- A two-car commuter train (white with a pink stripe, pantograph) comes out of one tunnel, stops at the platform, then departs into the other; clipping planes hide it inside the tunnels. Loop every ~32 s.
- Crossing lamps blink alternately and the arms lower before the train passes, then rise.
- ~460 instanced sakura petals drifting in the wind.
- Slow auto-orbit camera; OrbitControls for drag and zoom.

UI (English): bottom-left title "春風駅 / Harukaze Station" with a small caption; bottom-right glass toolbar with Day / Night, Orbit on/off, and Train (call the next train). Night mode: dark sky with stars, lit windows, lamp point lights with additive glow sprites, glowing crossing lamps and train headlights.

Technical: merge static meshes by material into a few dozen draw calls; use a seeded random so the scene is identical on every load; responsive down to phone width; respect prefers-reduced-motion.

出处：作者在原帖下的回复。提示词保留原文语言，未做改写。

作者提到的工具

更多模型对比案例

• 0:42

  AI自画像挑战：Sonnet5.5对比GPT-6 Astra

  725.7万 播放

• 0:28

  Opus 5.5 对比 GPT-6 Astra 武士游戏演示

  111.8万 播放

• 0:13

  软糖熊软体物理场景

  62.6万 播放

• 0:30

  GPT-6 Astra对比Claude Opus5.5

  60.4万 播放

• 0:22

  Grok与Opus摸草模拟器成本对比

  55.3万 播放

• 0:44

  Opus5.5对比GPT-6 Sol虚幻引擎游戏开发

  54.3万 播放