layer-image

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Layer Image Generation

Layer 图像生成

Overview

概述

The loop is the one the

layer

skill teaches:

list_base_models

with

filter.use_case: "text_to_image"

get_base_model

, estimate, execute, poll. What separates a usable image from a near miss is the prompt and the per-model contract, not the model choice, which the curated ranking already handles.

Editing an image that already exists is a different use case with different rules: see

layer-image-editing

. Sprites, icons, and game UI have their own constraints: see

layer-game-assets

. Holding one look across a set: see

layer-reference-sets

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

skill 所教授的流程是：使用

list_base_models

并设置

filter.use_case: "text_to_image"

，然后依次执行

get_base_model

、估算、执行和轮询。决定图像可用与否的关键是提示词和每个模型的契约，而不是模型选择；模型选择已由精选排序处理妥当。

编辑已有图像属于另一种用例，需要遵循不同的规则：请参阅

layer-image-editing

。Sprites、图标和游戏 UI 有各自的约束：请参阅

layer-game-assets

。在一组图片中保持统一风格：请参阅

layer-reference-sets

。

如果此处提到的同级 skill 不在可用 skill 列表中，请让用户安装（

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

）；在无人值守场景下，则根据工具 schema 继续操作，并标记该缺失项。

What a prompt must name

提示词必须明确的内容

A model fills every gap you leave, and it fills it with the most average answer in its training data. Each of these is a gap worth closing, in roughly this order of impact:

Subject, concretely. "A dwarf blacksmith" beats "a fantasy character".
Style register, as a production term rather than an artist's name: hand-painted, cel-shaded, flat vector, painterly semi-realism, stylised PBR render, gouache, 90s airbrush.
Shot and framing: full body, three-quarter portrait, top-down, isometric, extreme close-up. This is the single most common omission and the most common cause of a reroll.
Lighting: key direction, hardness, time of day, practical sources. "Rim-lit from behind, warm forge glow below" is a decision; "good lighting" is not.
Palette, named or constrained: "muted teal and rust, two accent hues only".
Background treatment: flat colour, isolated on transparent, environment implied, full scene. Assets that will be composited need this stated, not assumed.

Leave out what does not matter. A prompt that specifies everything specifies nothing, because the model weights the whole string and dilutes the parts you cared about.

模型会填补你留下的每一个空白，而且通常会用其训练数据中最平庸的答案来填补。以下这些空白都值得补齐，其影响程度大致按此顺序排列：

主体，要具体。“一位矮人铁匠”胜过“一个奇幻角色”。
风格类别，应使用制作术语，而不是艺术家姓名：手绘、赛璐璐上色、扁平矢量、绘画式半写实、风格化 PBR 渲染、水粉、90 年代喷绘。
镜头与取景：全身、四分之三视角肖像、俯视、等距视角、特写。这是最常被省略的信息，也是导致重新生成的最常见原因。
光照：主光方向、硬度、时间以及实际光源。“从后方形成轮廓光，下方带暖色熔炉辉光”是一项明确决策，而“良好的光照”不是。
配色，应明确命名或加以限制：“低饱和青绿色与铁锈色，仅使用两种辅助色”。
背景处理：纯色、隔离在透明背景上、暗示环境存在或完整场景。后续需要合成的素材必须明确说明这些要求，不能依赖默认行为。

不要描述无关紧要的内容。如果提示词指定了一切，也就等于什么都没指定，因为模型会权衡整个字符串，从而稀释你真正关注的部分。

Sizing, and the fields that differ

尺寸，以及各模型不同的字段

Aspect ratio and resolution are per-model contracts. Two models that both do text-to-image disagree about whether they take an aspect ratio string, explicit width and height, or a fixed set of named sizes, so read

get_base_model

rather than reusing the shape that worked last time.

When the deliverable has a fixed ratio, filter for it.

filter.capabilities

is an object of booleans, not a list of names, so it is

{portrait_9_16: true}

square

portrait_9_16

landscape_16_9

, or

multiple_aspect_ratios

when the user will want several crops from one setup. Set only what the task needs, since

false

filters rather than defaults.

宽高比和分辨率由每个模型的契约决定。两个都支持 text-to-image 的模型，可能分别要求宽高比字符串、明确的宽度和高度，或一组固定的命名尺寸，因此应读取

get_base_model

，而不是沿用上次奏效的参数结构。

当交付物需要固定比例时，应按该比例筛选。

filter.capabilities

是一个布尔值对象，而不是名称列表，因此应写成

{portrait_9_16: true}

：可使用

square

、

portrait_9_16

、

landscape_16_9

，或在使用者需要从同一设置中生成多种裁切比例时使用

multiple_aspect_ratios

。只设置任务真正需要的字段，因为

false

表示过滤，而不是设置默认值。

Negatives, references, and seeds

负面提示、参考与种子

negative_prompt

is a capability, not a universal parameter. Filter for it when the run needs one, and do not send it blindly: on a model without it, it is either ignored or an error, and on a model with it, a long negative list costs more than it saves. Reach for it to suppress one specific, recurring defect.

Three reference types steer a text-to-image run, all passed through

guidance_files

after an upload:

```
reference_image
```
with the
```
image_editing
```
capability, for "reproduce this subject" and for new art that has to match a reference.
```
style_reference
```
is the older style-only transfer: it moves a look across, it does not carry a subject, so reaching for it to hold a character is the usual reason a cast drifts.
```
pose
```
with
```
character_pose
```
, for "put the character in this position".

depth

canny

lineart

, or

scribble

with

structure

outline

, for "keep this layout".

A reference is worth more than three adjectives. When a user has an image and is describing it in words, upload the image.

Seeds make a run repeatable, which matters when iterating: hold the seed and change one clause to see what that clause actually does. They do not make two different prompts consistent with each other. For a look held across a set, train it: see

layer-reference-sets

negative_prompt

是一项能力，而不是通用参数。仅当本次生成需要它时才进行筛选，不要盲目传入：对于不支持它的模型，该参数可能被忽略，也可能引发错误；对于支持它的模型，冗长的负面提示列表通常成本大于收益。它适合用于抑制某一个具体且反复出现的缺陷。

有三种参考类型可以引导 text-to-image 任务，上传后均通过

guidance_files

传入：

```
reference_image
```
配合
```
image_editing
```
能力，用于“复现这个主体”，以及生成必须与某张参考图匹配的新艺术内容。
```
style_reference
```
是较早期的纯风格迁移方式：它可以迁移视觉风格，但不会携带主体，因此试图用它来保持角色一致，正是角色群像经常发生风格漂移的原因。
```
pose
```
配合
```
character_pose
```
，用于“将角色置于这个姿势”。
```
depth
```
、
```
canny
```
、
```
lineart
```
或
```
scribble
```
配合
```
structure
```
或
```
outline
```
，用于“保留这个布局”。

一张参考图胜过三个形容词。当用户手中已有图像并试图用文字描述它时，应直接上传该图像。

种子可以让生成结果具备可重复性，这在迭代时很重要：保持种子不变，每次只修改一个子句，以观察该子句实际产生的影响。种子无法让两个不同的提示词彼此保持一致。如果需要在一组图片中保持统一风格，应训练该风格：请参阅

layer-reference-sets

。

Text in the image

图像中的文字

Most image models garble text. When the deliverable needs readable words, filter for a model whose description claims typography, keep the string short, and put it in quotes in the prompt. If it still fails after two attempts, stop rerolling: generate the art clean and composite the text, which is also what makes it editable and localisable later.

大多数图像模型无法准确生成文字。当交付物需要包含可读的词语时，应筛选描述中明确支持 typography 的模型，保持文本字符串简短，并在提示词中用引号将其标出。如果尝试两次后仍然失败，不要继续重新生成：先生成不含文字的干净画面，再合成文字；这样也能让文字在之后保持可编辑和可本地化。

Worked example

实际示例

"Key art for our roguelike, 16:9, for the Steam page."

list_base_models

with

filter.use_case: "text_to_image"

and

filter.capabilities: {landscape_16_9: true}

. Take the first result.

```
get_base_model
```
, which reports the sizing fields it accepts and whether it needs a trigger word.
Compose against the six points above: "Hooded rogue mid-leap over a collapsing stone bridge, seen three-quarter from below, hand-painted semi-realism, hard moonlight from the upper left with cold blue rim light, warm torchlight pooling below, muted slate and amber, storm-lit ruins receding into fog behind."
```
estimate_forge_price
```
with
```
batch_size: 4
```
, to see four compositions of one idea.
Under 20 CUs, so execute, then poll at
```
poll_interval_seconds
```
.
Present all four. Pick one, hold its seed, and iterate one clause at a time.

“用于我们 Roguelike 的主视觉，16:9，用于 Steam 页面。”

使用

list_base_models

，设置

filter.use_case: "text_to_image"

和

filter.capabilities: {landscape_16_9: true}

。选择第一个结果。

使用
```
get_base_model
```
，查看它接受的尺寸字段，以及是否需要触发词。
根据上述六点编写提示词：“披着兜帽的盗贼跃过正在坍塌的石桥，从下方以四分之三视角观察，手绘半写实风格，左上方硬质月光配冷蓝色轮廓光，下方暖色火光汇聚，低饱和石板灰与琥珀色，背后被风暴照亮的废墟渐隐入雾中。”
使用
```
estimate_forge_price
```
并设置
```
batch_size: 4
```
，查看同一构想的四种方案。
如果低于 20 CUs，则执行生成，然后按照
```
poll_interval_seconds
```
进行轮询。
展示全部四张结果。选择其中一张，保持其种子不变，每次只迭代修改一个子句。

Common mistakes

常见错误

Omitting shot and framing, then rerolling the same prompt hoping for a different crop.
Describing a reference image in words instead of uploading it.
Sending
```
negative_prompt
```
to a model that does not declare the capability.
Reusing the sizing fields from a different model rather than reading
```
get_base_model
```
.
Using
```
batch_size
```
for N different assets. It makes variations of one prompt.
Chasing readable text through a sixth reroll instead of compositing it.
Stacking five style adjectives, which averages them into none of them.

省略镜头与取景，然后反复使用同一提示词，希望得到不同的裁切结果。
用文字描述参考图，而不是直接上传参考图。
向未声明支持
```
negative_prompt
```
能力的模型传入该参数。
沿用其他模型的尺寸字段，而不是读取
```
get_base_model
```
。
使用
```
batch_size
```
一次生成 N 个不同素材；它只会生成同一提示词的变体。
为了追求可读文字而进行第六次重新生成，而不是将文字合成到图像中。
堆叠五个风格形容词，结果将它们平均化，最终失去所有风格特征。