3d-high-poly-models

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

3D High-Poly Models

3D高多边形模型

Build the requested close-up detail into the model, then make it practical to render. Judge realism through silhouette, proportions, material response, and lighting as well as polygon count.
在模型中构建所需的特写级细节,再确保其可高效渲染。真实感不能只看多边形数量,还要结合轮廓、比例、材质响应和光照效果综合判断。

Establish what the camera needs

明确相机需求

Inspect the reference, intended hero angle, closest approach, orbit range, model scale, current mesh, and target hardware. Record the current triangle count and draw calls rather than labeling an asset high-poly from appearance alone. Keep a high-detail source/master when building new geometry so runtime optimization remains reversible.
Separate detail into three scales:
  • Silhouette: outer contours, root flares, carved edges, folded fabric, roof curvature. Use geometry.
  • Surface form: bevels, dents, joints, raised seams, deep cracks. Use geometry or justified displacement when the view reveals parallax and shadow shape.
  • Microdetail: grain, pores, shallow scratches, fiber structure. Use material maps unless the extreme close-up actually resolves geometry.
Use the project's modeling tools or asset pipeline. Match real dimensions and proportions first, then add adaptive subdivisions, support topology, sculpting, or appropriately licensed scans. A dense mesh with incorrect proportions remains incorrect.
检查参考资料、预期主视角、最近观察距离、环绕观察范围、模型比例、当前网格和目标硬件。记录当前的三角形数量和Draw Call数,不要仅凭外观就将资产归类为高模。构建新几何结构时需保留高细节源文件/主文件,确保运行时优化可回退。
将细节分为三个层级:
  • 轮廓: 外轮廓、根部扩口、雕刻边缘、褶皱布料、屋顶曲率。使用几何结构实现。
  • 表面形态: 倒角、凹痕、接缝、凸起缝、深裂缝。当视角下能观察到视差和阴影形态时,采用几何结构或合理的位移技术实现。
  • 微观细节: 纹理颗粒、孔隙、浅划痕、纤维结构。除非极端特写下必须用几何结构才能呈现,否则均使用材质贴图实现。
使用项目的建模工具或资产管线。首先匹配真实尺寸和比例,再添加自适应细分、支撑拓扑、雕刻,或使用授权合规的扫描资产。比例错误的网格,即便再密集也是不合格的。

Model for believable close-ups

打造逼真的特写效果模型

Give manufactured edges a plausible bevel so they catch highlights. Preserve designed hard edges while smoothing curved surfaces; do not smooth every normal or indiscriminately weld UV seams. Recompute or preserve normals/tangents according to the modeling/export pipeline.
Spend curve segments where projected curvature is visible. Increase them around hero rims, roof silhouettes, and tight bends; use fewer on tiny distant parts. Check subdivision boundaries for pinching and displacement for cracks or detached surfaces.
For procedural branches, sweep rings along a smooth curve, transport the local frame to avoid sudden twist, taper continuously, and bury child starts inside the parent surface. Add denser rings at a flared root or strong curvature. Seijaku's
treeMesh
is a useful example of adaptive radial/longitudinal detail rather than uniform subdivision.
For scanned stones, preserve the scan's characteristic proportions. Excessive nonuniform scaling stretches both shape and surface features. Reuse the mesh with limited variation in rotation, placement, and scale; keep the most recognizable hero forms distinct.
给工业化生产的边缘添加合理的倒角,使其能够捕捉高光。平滑曲面的同时保留设计指定的硬边;不要平滑所有法线,也不要随意焊接UV接缝。根据建模/导出管线的要求,重新计算或保留法线/切线。
在投影曲率可见的位置分配更多曲线段。在主视角边缘、屋顶轮廓和急弯处增加曲线段;在远处的细小部件上减少用量。检查细分边界是否有挤压变形,位移效果是否存在裂缝或表面分离问题。
对于程序化生成的树枝,沿平滑曲线扫掠环形线,传递局部坐标系以避免突然扭转,实现连续锥度,并将子枝的起点埋入父级表面内部。在扩口根部或曲率大的位置添加更密集的环形线。Seijaku的
treeMesh
是很好的参考案例,它采用自适应径向/纵向细节处理,而非统一细分。
对于扫描生成的石头模型,要保留扫描资产的特征比例。过度的非均匀缩放会同时拉伸形状和表面特征。可复用网格,仅在旋转、位置和比例上做有限调整;让最具辨识度的主形态保持独特性。

Prepare runtime versions

制备运行时版本

  1. Keep the authored master and export a validated runtime asset, usually glTF/GLB when the stack supports it. Check units, transforms, materials, UV seams, normals, and bounds in the actual viewer.
  2. Create lower-detail versions for distant use, preserving silhouette and baked shading. Derive thresholds from projected size and compare transitions from the real camera. Use hysteresis or an appropriate transition to avoid rapid LOD switching. See Three.js LOD.
  3. Instance repeated geometry/material pairs such as roof tiles, stones, or foliage where suitable. Partition large repeated fields spatially so one visible object does not force the whole site to draw.
  4. Keep shadow geometry and shadow distance proportional to visible benefit. A detailed object rendered in several shadow maps can cost far more than its beauty pass suggests.
  5. Compress geometry transport where supported, then validate decoded geometry and appearance. Compression reduces transfer size; it does not inherently reduce runtime triangle count. Treat quantization and simplification as distinct operations with separate visual checks.
There is no universal correct triangle budget. Profile the representative scene with its actual materials, lights, shadows, transparency, and target resolution. Report unique asset triangles separately from the triangles drawn across instances and passes. Geometry compression, instancing, and LOD solve different costs.
  1. 保留制作完成的主文件,导出经过验证的运行时资产;技术栈支持时优先使用glTF/GLB格式。在实际查看器中检查单位、变换、材质、UV接缝、法线和边界框。
  2. 创建供远景使用的低细节版本,保留轮廓和烘焙着色效果。根据投影尺寸推导阈值,并从真实相机视角对比过渡效果。采用迟滞机制或合适的过渡方式,避免LOD频繁切换。参考Three.js LOD。
  3. 在合适的场景下,对屋顶瓦片、石头、植物等重复的几何/材质组合使用实例化技术。对大型重复区域进行空间分区,避免单个可见物体导致整个场景被绘制。
  4. 阴影几何结构和阴影距离应与可见收益成正比。细节丰富的物体如果在多张阴影贴图中渲染,其开销可能远高于主渲染通道的开销。
  5. 在支持的情况下压缩几何数据传输,随后验证解码后的几何结构和外观效果。压缩能减小传输体积,但本身并不会减少运行时的三角形数量。将量化和简化视为两项独立操作,分别进行视觉效果验证。
不存在通用的三角形数量预算标准。需使用实际的材质、光照、阴影、透明度和目标分辨率,对代表性场景进行性能分析。将单资产的三角形数量与实例化和多通道绘制的总三角形数量分开统计。几何压缩、实例化和LOD解决的是不同类型的性能开销。

Avoid a startup stall

避免启动卡顿

Load visible hero geometry first and defer optional interiors or distant detail. Generate procedural meshes offline or in bounded work, use workers where useful, and reuse buffers instead of regenerating identical parts. Keep a useful coarse representation until the detailed asset is ready. Prewarm shader variants with supported asynchronous compilation when it materially reduces first-use hitches.
For selection and collision, use suitable proxies or a spatial acceleration structure. Repeated full-resolution triangle scans during pointer movement can make a visually smooth scene feel unresponsive.
优先加载可见的主几何结构,延迟加载可选的内部结构或远景细节。离线或在有限工作量内生成程序化网格,合适的场景下使用Worker,复用缓冲区而非重新生成相同部件。在细节资产加载完成前,保留一个可用的粗略表示模型。如果异步编译能显著减少首次使用的卡顿,可利用支持的异步编译功能预热Shader变体。
对于选中和碰撞检测,使用合适的代理体或空间加速结构。指针移动时反复进行全分辨率三角形扫描,会让视觉上流畅的场景显得响应迟钝。

Verify

验证

  • Compare the master and runtime mesh at the closest camera distance under neutral and final lighting.
  • Inspect silhouette, bevel highlights, seams, contacts, shadow shape, and wireframe density.
  • Walk across LOD thresholds and confirm there are no visible holes or distracting jumps.
  • Measure startup, first interaction, full-frame draw calls/triangles, and frame time on desktop and mobile.
  • Claim ultra-realistic appearance only when the rendered result supports it; record the measured geometry count separately.
Read REFERENCES.md for source and APIs. Seijaku combines detailed swept trunks and scanned rock geometry with inexpensive instanced leaf cards. Do not interpret its appearance as evidence that every object needs a dense mesh.
  • 在最近相机距离下,分别在中性光照和最终光照环境中对比主文件和运行时网格。
  • 检查轮廓、倒角高光、接缝、衔接处、阴影形态和线框密度。
  • 遍历LOD阈值区间,确认没有可见孔洞或令人分心的跳变。
  • 在桌面端和移动端分别测量启动时间、首次交互时间、全帧Draw Call/三角形数量和帧耗时。
  • 仅在渲染结果足以支撑时,才能宣称超写实效果;单独记录实测的几何结构数量。
查阅REFERENCES.md获取源文件和API相关信息。Seijaku将精细的扫掠树干和扫描岩石几何结构,与低成本的实例化叶片面片相结合。不要因为它的外观效果,就认为每个物体都需要密集网格。