# 36 张 AI 卡牌插画，怎么保持一致性？
**作者**: kshift
**日期**: 2026-07-24T08:22:27.000Z
**来源**: [https://x.com/kshift/status/2080569286184009993](https://x.com/kshift/status/2080569286184009993)
---
给自己留个档，顺便回答一个被问了几次的问题（让 fable 老师写的文章）。
《谜宫》图鉴里的 36 张卡面插画，全部是 Codex image-gen 生成的。问得最多的是两件事：这么多张图，怎么保证看起来是同一个画师的手笔？同一个角色反复生成，脸怎么不跑？
我的经验是：别指望找到一条万能 prompt，一致性是流程给的。生成之前，用设定和参考图把模型夹住；生成之后，再用脚本统一收口。拆开讲就三步。
36 张卡 · 同一套 image-gen 画风
## 一、动笔之前，先把脸写死
模型没有记忆，「同一个角色」这件事，只能你替它记着。更麻烦的是，你不管它，它就把所有人往一张「平均脸」上画——年轻女性角色尤其严重。
所以每个露脸的角色，我都先锁六条轴：脸型、眼型、鼻型、眉形、年龄、常态表情，再各配一个独占标志。还定了条死规矩：任意两个同性别角色，至少四条轴不一样。
（完整版还有鼻型、眉形两条轴，这里省了两列。）
这些设定最后落成一段英文描述，原样塞进每次生成的 prompt，一个词都不许改。比如吸血鬼这段：
```
sharp pale aristocratic face with a pointed chin, narrow piercing
dark-red eyes, pointed ears, high narrow nose, exposed fangs,
slicked-back black hair with a widow's peak, apparent age mid-30s
```
看着死板？脸这个东西，就得死板。效果对着看：
同一批生成的六张脸，没有一张重样
## 二、给每个角色定一张立绘，当身份证
角色的全身立绘定稿之后，这个角色往后所有的图，都挂这张立绘当参考图。
但光挂图不够，得跟模型把话说清楚：这张图只管「你是谁」，不管「你在干什么」。prompt 里的原话是这样的：
```
The attached image defines the character's identity only: keep the
same face, species anatomy, hairstyle, clothing, colors and signature
equipment. It does NOT define the acting — do not copy the reference's
pose, expression, camera angle or framing; the scene description below
overrides all of them.
```
脸、发型、配色、标志性装备，照参考图来；姿势、表情、镜头，全听场景描述的。效果是这样：
参考图只锁「这是谁」
同一条龙。左边是定稿立绘，右边是卡面——趴在金山上，瞪着刚闯进来的你。身份锁死了，戏随便演。
## 三、prompt 是拼出来的，不是写出来的
每张卡的 prompt 由五段拼成：画风、身份、场景、构图、约束。其中四段 36 张卡全部共用，逐字复制；每张卡真正要动笔的只有「场景」一段——写清楚这个角色在地下城的哪个瞬间。
画风段长这样（36 张共用）：
```
A full-bleed 3:4 portrait fantasy card illustration. Western painted
cartoon style matching the attached character reference exactly: bold
thick dark outlines, solid dimensional color blocks, flat colors with
simple cel shading, consistent detail density. Dramatic cinematic
lighting from the single in-scene light source specified in the scene
description; the background is painted in the same style with lower
detail density than the subject so the subject reads first.
```
场景段贴三个，怪物和冒险者都有，感受一下写法。
上一节那条龙（怪物 · 没有脸的物种，重点全在镜头上）：
```
Scene: the deepest treasure vault of the dungeon, seen from the
intruder's eye level. Low-angle close-up: the red dragon's huge head,
neck and one clawed forefoot dominate the frame, rising over a slope
of gold coins and goblets. Its half-open yellow eye with a narrowed
pupil stares straight at the viewer; a thin wisp of smoke rises from
one nostril. A few coins slide down the pile beneath its claw.
Light source: warm torch light from the lower front, as if from the
intruder's own torch.
```
死灵法师（有脸的角色——第一节那段脸部描述，拼的就是这一张。她的人设是「对活人不耐烦，对死者彬彬有礼」：翻着名册，礼貌地给列队走进黑暗的骷髅们引路）：
```
Scene: a green-lit crypt corridor. The necromancer stands at the side of the
corridor with courtly poise, one hand holding an open roster book with wisps
of green soul-fire floating above its pages like candle flames, the other
hand gesturing politely down the corridor as if ushering guests through. Her
skull-topped staff rests in the crook of her arm. Deeper in the corridor
behind her, a few vague skeletal silhouettes walk away in an orderly line
into the darkness. Light source: the green soul-fire, lighting her gaunt face
from the lower front.
```
牧师（冒险者。别人往黑暗里举火把，她自己就是那点光）：
```
Scene: the threshold where a lit dungeon camp gives way to a pitch-dark
corridor, chest-up framing in three-quarter view. The priestess walks calmly
forward into the darkness, ahead of everyone, her thumb mid-count on a string
of worn prayer beads held at her chest. Her cleric veil frames her oval face;
her soft narrow eyes look ahead into the dark with a faint serene smile,
utterly unafraid. The golden sun emblem on her chest gives off a gentle glow
that lights the first step of the darkness before her. Behind her at the
frame edge, warm torchlight of the camp she is leaving recedes. Light source:
the soft golden glow of her sun emblem from her chest, with warm torchlight
falling off behind her.
```
剩下两段固定块也贴全，想试的可以直接拿去拼：
构图段（36 张共用；最后那句安全区，是因为卡框会在代码里叠上去）：
```
Portrait 3:4 composition. The background scene fills the entire canvas edge
to edge, no borders, no vignette, no empty margins. Use the cinematic framing
specified in the scene description: like a movie still, the subject's body
may be naturally cropped by the frame, but the face and every signature
feature named in the scene description must stay fully inside the central
safe area, away from the outer 8 percent of every edge, because a card frame
will be overlaid in code.
```
约束段（36 张共用）：
```
No text, no letters, no numbers, no watermark, no signature, no card frame,
no UI elements, no photorealistic rendering, no depth-of-field blur, no
plastic highlights, no malformed anatomy, no extra fingers, no named or
recognizable third-party intellectual property.
```
写场景这件事，我攒下来三条心得：
- 光源要有来历。不写 "dramatic lighting" 这种空话。龙那张的光是「闯入者自己手里那支火把」，牧师那张的光是她胸口的圣徽——光有了出处，画面就有了戏。
- 姿势、机位都写进故事里。「她像给来宾引路一样，朝走廊深处礼貌地摆手」比 "waist-up shot, 45 degrees" 这种镜头术语好用得多，模型对「情境」的理解远强于对「参数」的理解。
- 约束要正着说，别说「不要」。想避免多余的手，就写「双手清晰可数」，别写「不要画多手」。模型对否定词的执行力，用过的都懂。
整个结构一张图说完——四段不动，换一段场景，就是一张新卡：
每张卡的 prompt = 五段逐字拼接
## 出图之后的事
生成只是前一半。图出来之后还有一道脚本后处理：统一裁切到固定尺寸，校验脸有没有出安全区；图标类资产还要过一遍描边和色彩归一。
说到底就一句：手气交给模型，一致性留给流程。
这 36 张卡都收在游戏图鉴里。
《谜宫》，每日地下城猜谜——六次机会，猜出深处的布阵：meikyu.app
## 相关链接
- [AI 托我给您带个话 reposted](https://x.com/AIWhispers4U)
- [kshift](https://x.com/kshift)
- [@kshift](https://x.com/kshift)
- [2.2K](https://x.com/kshift/status/2080569286184009993/analytics)
- [游戏图鉴](https://meikyu.app/world/)
- [meikyu.app](https://www.meikyu.app/)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [4:22 PM · Jul 24, 2026](https://x.com/kshift/status/2080569286184009993)
- [2,273 Views](https://x.com/kshift/status/2080569286184009993/analytics)
- [View quotes](https://x.com/kshift/status/2080569286184009993/quotes)
---
*导出时间: 2026/7/24 20:37:13*