学习路径Learning path · 第 4 阶段Stage 4
美术风格的一致性Keeping the art style consistent
Keeping the art style consistent
一句话：风格不一致不是提示词写得不够多：恰恰相反，官方最佳实践是提示词里少写风格词，把风格交给一个固定的「锚」。
（「风格一致性是生成模型弱项」是本站判断，见 Vibe Gaming 的拆解；下面的机制与做法是工具官方文档原话，当日核对。）
— the official best practice is the opposite: keep style words out of the prompt and hand style to one fixed "anchor".
("Style consistency is a generative-model weakness" is this site's judgement, see the Vibe Gaming breakdown; the mechanisms below are tool-doc wording, checked that day.)
官方机制：把风格固定成一个锚
| 工具 | 官方机制（当日核对） | 官方原话要点 |
|---|---|---|
| Midjourney | （--sw，0–1000，默认 100） | 官方文档：sref 捕捉参考图的「visual vibe」：颜色、媒介、纹理、光照："helping you achieve a consistent visual theme"（docs.midjourney.com，2026-10-29） |
| Midjourney | 风格码（--sref <code>） | 同一个码 reroll/变体保持同一风格；风格码可从 Style Explorer 选 |
⚠ 这张表核对的是「官方文档怎么说」，不是「效果一定稳」：本站不做实测。
The official mechanism: pin style to one anchor
| Tool | Official mechanism (checked that day) | Official wording |
|---|---|---|
| Midjourney | (--sw, 0–1000, default 100) | Docs: sref captures the reference's "visual vibe" — colors, medium, textures, lighting — "helping you achieve a consistent visual theme" (docs.midjourney.com, 2026-10-29) |
| Midjourney | Style codes (--sref <code>) | The same code keeps the same style across rerolls/variations; codes are browsable in the Style Explorer |
⚠ The table checks what the official docs say, not "the result is guaranteed stable" — this site runs no measurements.
官方最佳实践：提示词写内容，不写风格
Midjourney 官方文档对 sref 的最佳实践原话（当日核对）：
- Keep text prompts simple：别加会和参考图冲突的风格词
- Focus on content, not instructions：提示词描述「看到什么」，不是「模型该怎么改参考图」
- 官方给的坏例子：
the look of this image but a dog；好例子：detailed portrait of a dog
对游戏工作流（本站推论）：先花一次时间定锚（生成或挑一张你满意风格的图，记下它的 sref/风格码），之后整包素材：角色、地形、UI 图标：都挂同一个锚，提示词只写「这是什么」。
Official best practice: prompts describe content, not style
Midjourney's own best practices for sref (checked that day):
- Keep text prompts simple — avoid style words that conflict with the reference
- Focus on content, not instructions — describe what to see, not how the model should modify the reference
- Official bad example:
the look of this image but a dog; good:detailed portrait of a dog
For game workflows (this site's inference): spend the time once to set the anchor (generate or pick one image whose style you like, keep its sref/code), then hang the whole pack — characters, terrain, UI icons — on that same anchor, and let prompts say only "what this is".
更稳的路：成套素材包
生成模型再稳也有漂移；：同一个作者、同一套规格（见 音频素材的来源与授权 同款思路在美术侧的对应：Kenney 这类整包 CC0 素材，像素风/卡通风各有成套的）。
⚠ 原型阶段最省事的组合：成套开源素材做占位 + 风格锚生成做特色件。两者混用时，先保证「一套里一致」，再考虑「两套之间搭不搭」。
The steadier road: matching asset packs
However stable the generator, it drifts; open-source packs are consistent by construction — one author, one spec (the art-side twin of the audio sources logic: Kenney-style whole packs, pixel or cartoon, each internally matching).
⚠ The laziest workable combo in prototyping: a matching open pack for placeholders + style-anchored generations for signature pieces. When mixing the two, get "consistent within each set" first, then worry about whether the two sets sit well together.
有序路径 · 第 13 / 26 篇Ordered path · item 13 of 26