layer-3d

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Layer 3D Generation

Layer 3D 生成

Overview

概述

3D on Layer is a chain of separate use cases, not one call. Generate the mesh, then remesh, retexture, rig, or animate it as the deliverable requires. Each stage is its own

filter.use_case

and its own run, and each takes the previous stage's output

file_id

as input.

Stage	`filter.use_case`	Input
Mesh from a description	`text_to_3d`	prompt
Mesh from concept art	`image_to_3d`	`init_image` or `reference_image`
Cleaner topology, lower count	`remesh`	`init_mesh`
New surfaces on an existing mesh	`retexture`	`init_mesh` , often `texture_image`
Skeleton and weights	`rig`	`init_mesh`
Motion on a rigged mesh	`animate`	rigged `init_mesh`

The loop and the spend gate are in the

layer

skill. If a sibling skill named here is missing from your available skills, ask the user to install it (

npx skills add layerai/skills --skill <name>

); unattended, proceed from tool schemas and flag the gap.

Layer 上的 3D 并不是一次调用，而是由一系列独立 use case 组成的链式流程。先根据交付要求生成网格，然后按需 remesh、retexture、 rig 或 animate。每个阶段都有自己的

filter.use_case

和独立的运行任务，并以上一阶段输出的

file_id

作为输入。

阶段	`filter.use_case`	输入
根据描述生成网格	`text_to_3d`	prompt
根据概念图生成网格	`image_to_3d`	`init_image` 或 `reference_image`
更干净的拓扑、更低的面数	`remesh`	`init_mesh`
为现有网格添加新的表面	`retexture`	`init_mesh` ，通常还需 `texture_image`
骨架和权重	`rig`	`init_mesh`
为已绑定骨骼的网格添加运动	`animate`	已绑定骨骼的 `init_mesh`

循环机制和费用门槛定义在

layer

skill 中。如果这里提到的某个同级 skill 未包含在你的可用 skills 中，请让用户安装它（

npx skills add layerai/skills --skill <name>

）；在无人值守模式下，则根据工具 schema 继续执行，并标记这一缺失项。

Start from an image

从图像开始

image_to_3d

is the stronger path whenever the design is settled or settleable. A generated mesh inherits its silhouette and surface detail from the reference, so approving a still first with

layer-image

costs one cheap image run and removes most of the iteration from an expensive 3D one.

The reference image should show the subject clean: isolated on a plain background, lit evenly, in a three-quarter view that reveals depth. A dramatic, heavily shadowed key art frame makes a worse reference than a plain orthographic-ish view, because the model reads shadow as geometry.

text_to_3d

is right when nothing exists yet and the shape is simple or generic enough to describe.

只要设计已经确定或可以确定，

image_to_3d

就是更优的路径。生成的网格会继承参考图的轮廓和表面细节，因此先使用

layer-image

确认一张静帧只需一次低成本的图像运行，并且可以消除昂贵的 3D 运行中的大部分迭代。

参考图应清晰地呈现主体：置于纯色背景上、照明均匀，并采用能够展现深度的四分之三视角。戏剧化且阴影浓重的主视觉图并不是理想参考，因为模型会将阴影解读为几何结构。

当尚未存在任何素材，且形状足够简单或通用、可以用文字描述时，适合使用

text_to_3d

。

Say what the mesh is for

说明网格用途

The engine target changes the request, so state it rather than accepting defaults:

Topology:
```
quad_mesh
```
where the asset will be sculpted, deformed, or subdivided downstream. Triangles are fine for a static prop that ships as-is.
Budget: a
```
face_limit
```
reflecting the platform. A hero prop on desktop and a background prop on mobile are different requests, and remeshing later costs another run.
Surfaces:
```
pbr_materials
```
and
```
include_textures
```
when the asset goes into a lit scene. Without them you get geometry and a flat surface, which reads as broken in an engine.

The per-model inference schema is the authority on which of these apply, so read

get_base_model

before composing the run.

目标引擎会改变生成需求，因此应明确说明，而不是直接接受默认设置：

拓扑：如果资产后续还要用于雕刻、变形或细分，应选择
```
quad_mesh
```
。对于按原样发布的静态道具，三角形网格也没有问题。
预算：使用符合目标平台的
```
face_limit
```
。桌面端的核心道具和移动端的背景道具属于不同需求，之后再进行 remesh 还需要额外运行一次。
材质表面：当资产要放入有光照的场景时，使用
```
pbr_materials
```
和
```
include_textures
```
。如果不提供这些选项，得到的只是几何体和单一平面，在引擎中会显得像出现了故障。

每个模型的推理 schema 都是判断上述选项适用范围的依据，因此请先读取

get_base_model

，再配置本次运行。

Describing geometry

描述几何结构

3D prompts reward structural language and ignore atmosphere. Lighting, mood, and camera belong to an image prompt and do nothing here, because the mesh carries no lighting.

Name the form: overall silhouette, proportions, how parts join, surface material, and the level of detail. "Squat clay cooking pot, wide rounded belly, short flared neck, two small looped handles at the shoulder, hand-thrown surface with visible ridges" describes a mesh. "Beautiful ancient pot, dramatic lighting" describes a photograph and will produce a generic one.

Symmetry and closed forms generate reliably. Thin protrusions, chains, cloth, and hair are where generated meshes fail, so plan to model those separately or accept them as fused geometry.

3D 提示词适合使用结构性语言，氛围描述则会被忽略。灯光、情绪和相机参数属于图像提示词，在这里不起作用，因为网格本身不包含光照信息。

应描述形态：整体轮廓、比例、部件连接方式、表面材质以及细节程度。“矮胖的陶制炊锅，宽而圆润的腹部，短而外扩的颈部，肩部有两个小环形把手，手工拉坯表面带可见纹脊”描述的是一个网格。“美丽的古董锅，戏剧性灯光”描述的是一张照片，最终只会得到一个泛泛的结果。

对称且封闭的形体更容易稳定生成。细长凸起、链条、布料和头发是生成网格最容易失败的部分，因此应计划单独建模，或接受将它们作为融合后的几何体处理。

Rigging and animation

骨骼绑定与动画

rig

produces a skeleton and weights, and it expects a mesh in a neutral, limbs-apart rest pose. A mesh generated in a dynamic action pose rigs badly, so generate the rest pose first and animate into the action.

animate

needs a rigged mesh, so the two run in order. Where a model offers named motion presets, those are more reliable than describing the motion in prose.

rig

会生成骨架和权重，并要求网格处于中立、四肢分开的静止姿势（rest pose）。以动态动作姿势生成的网格很难正确绑定，因此应先生成静止姿势，再通过

animate

制作目标动作。

animate

需要已绑定骨骼的网格，因此这两个阶段必须按顺序运行。如果模型提供具名动作预设，其可靠性高于使用自然语言描述动作。

Worked example

实操示例

"Turn this approved prop concept into a game-ready mesh for mobile."

The concept exists from an earlier image run and already carries a
```
file_id
```
.
```
list_base_models
```
with
```
filter.use_case: "image_to_3d"
```
and
```
filter.capabilities: {textures: true}
```
(an object of booleans, not a list of names). Take the first result, then
```
get_base_model
```
for the fields it accepts.
```
estimate_forge_price
```
with the concept as
```
image_to_3d
```
input,
```
pbr_materials
```
and
```
include_textures
```
on,
```
quad_mesh
```
off for a static prop, and a
```
face_limit
```
matched to mobile. 3D runs are expensive, so expect this to clear 20 CUs and ask before proceeding.
Execute, poll at
```
poll_interval_seconds
```
.
Inspect the result before chaining. If the silhouette is right but the count is too high, the next run is
```
remesh
```
on the output
```
file_id
```
, not a regenerate.

“将这个已确认的道具概念设计转换为适合移动端的游戏可用网格。”

该概念设计来自之前的图像运行，并且已经带有
```
file_id
```
。
调用
```
list_base_models
```
，参数为
```
filter.use_case: "image_to_3d"
```
和
```
filter.capabilities: {textures: true}
```
（这是一个布尔值对象，而不是名称列表）。选择第一个结果，然后调用
```
get_base_model
```
获取它接受的字段。
调用
```
estimate_forge_price
```
，以该概念设计作为
```
image_to_3d
```
输入，开启
```
pbr_materials
```
和
```
include_textures
```
，对于静态道具关闭
```
quad_mesh
```
，并将
```
face_limit
```
设为适合移动端的数值。3D 运行成本较高，因此预计消耗将超过 20 CU，继续之前应先征求用户同意。
执行任务，然后按
```
poll_interval_seconds
```
轮询。
在串联后续阶段前检查结果。如果轮廓正确但面数过高，下一次运行应使用输出的
```
file_id
```
执行
```
remesh
```
，而不是重新生成。

Common mistakes

常见错误

Using
```
text_to_3d
```
when concept art exists or could be generated cheaply first.
Feeding a dramatic, shadow-heavy key art frame as the 3D reference.
Writing an image prompt (lighting, mood, camera) for a mesh.
Leaving topology, budget, and materials unspecified, then remeshing in a second paid run.
Rigging a mesh generated in an action pose.
Regenerating from scratch when
```
remesh
```
or
```
retexture
```
would fix the actual defect.
Expecting clean thin geometry such as chains, straps, or hair from a generated mesh.

在概念图已经存在或本可先低成本生成的情况下，仍然使用
```
text_to_3d
```
。
将戏剧化、阴影浓重的主视觉图作为 3D 参考图。
为网格编写图像提示词，例如描述灯光、氛围和相机。
不指定拓扑、预算和材质，随后再通过第二次付费运行进行 remesh。
对以动作姿势生成的网格进行骨骼绑定。
当
```
remesh
```
或
```
retexture
```
可以修复实际缺陷时，仍然从头重新生成。
指望生成网格自动产出链条、绑带或头发等干净的细薄几何结构。