AlterU AlterU Part 6 · gen-image 实战技巧 Part 6 · gen-image Field Notes
大纲Outline
美术 · gen-image 4 条实战

出图不翻车
4 个钩子。

1 必带

测 plate prompt 一定带 ref_url

只测 txt2img 看似好,加 ref_url 后玩家原照被拼进画面 — Fit Check 真踩过

同款坑:Build-a-Boyfriend tier 卡烧字、Hour Capsule MFG 时间戳。任何 plate 都 ref + 真照测一次。

2 前置

写风格 prompt 前先 Read 库

第一版风格 prompt 脑补不看库 → ship 立刻被骂"风格不一"——Kiss Wall 9 commit 中 5 个都是因此。

已有 asset 是真相,prompt 是描述。先 Read 后写,否则脑补必出错。同样适用于角色锚定(The Bidding · The Couturier · The Locksmith)。

3 身份钩

烧字进画面 = 身份钩 ②

Hour Capsule @user #00042 / Pulp Hour 红色 COMING NEXT chip / Build-a-Boyfriend tier 标签 / Daily Arcana 牌面 username。

SD 文字精度差 — 别依赖 prompt 精确出字,烧字用 HTML overlay 或 Canvas 后处理。

4 玩法

5-8s 等待 = 玩法

不是 spinner 糊过去:Kiss Wall 化学浴显影 · Hour Capsule sealing 工序耳语 · Fit Check developing 药丸 · Field Guide 翻档案动画。

范式 ⑥ 通用解:第一个输入 lock prompt → 进入显影 / 显形 / 拆封 / 翻档案的仪式。

Craft · gen-image · 4 field notes

How to ship pictures
that don't blow up.

1 mandatory

Test plates with ref_url

txt2img-only tests pass; add ref_url and the player's photo stitches into the frame. Fit Check learned this in production.

Same trap: Build-a-Boyfriend tier card burn-in, Hour Capsule MFG stamps. Every plate test = ref + real photo, no exceptions.

2 precondition

Read the library before writing a style prompt

v1 prompts written from imagination, never reading the library → shipped, immediately roasted as "style drifted." 5 of Kiss Wall's 9 commits were this.

Existing assets are truth; prompts describe them. Read first, prompt second. Same for character anchor sheets (The Bidding · The Couturier · The Locksmith).

3 identity

Baked text = identity hook ②

Hour Capsule @user #00042 / Pulp Hour red COMING NEXT chip / Build-a-Boyfriend tier tags / Daily Arcana card-face username.

SD text rendering is imprecise — never rely on the prompt to spell anything. Bake type via HTML overlay or Canvas post-processing.

4 play

5-8s wait = play

Not a spinner: Kiss Wall develop / Hour Capsule sealing whisper / Fit Check developing pill / Field Guide archive flip.

Generalizable: first input → lock prompt → enter the develop / form / unseal / archive-flip ritual.

备忘 · 按 S 关闭Notes · S to close

这一张是给已经听完 Part 3 / 4 的同学的"上手实战"。4 个钩子全部是真踩过坑总结的。

关键深度点:SD 文字渲染精度差 = 烧字一律走 HTML / Canvas overlay 层。Hour Capsule 的 MFG 时间戳 / Pulp Hour 的 chip 文案,所有"看似烧进图的字"实际都是 HTML 叠加在 gen-image 出的底图上。

This slide is the "now actually do it" sheet after Parts 3 and 4. All four notes are real production scars.

Deeper point: SD text rendering is fragile = bake type via HTML / Canvas overlay always. Hour Capsule's MFG timestamp and Pulp Hour's chip are HTML layered on top of the gen-image base, not from the prompt.

34 / 41