Encoding Design Judgment — A Visual Standard Assistant

01 | 问题背景 Background

一家多品牌、多渠道的大健康电商公司,视觉设计团队为旗下多个品牌产出电商主图。我把范围聚焦在「商品主图」——它最标准化、杠杆也最高:主图直接决定搜索与列表页的点击和转化。

A multi-brand, multi-channel health & wellness e-commerce company, whose design team produces marketplace main-images across many brands. I scoped the project to product main-images — the most standardized, highest-leverage surface: the main-image directly drives click-through and conversion on search and listing pages.

团队真正的瓶颈不在「出图快不快」——产能和速度公司已有工具覆盖。瓶颈在新人判断不准方向:公司多品牌、多渠道,没有统一的「长相」,真正在起作用的,是一套功效驱动型主图语法。它内化在资深设计师脑中、从未被写下来——教不了新人,也无法检查。于是新人要么误判跑偏、反复返工,要么事事问资深、上手慢。

The real bottleneck wasn't production speed — throughput was already tool-covered. It was that junior designers couldn't judge direction. Across many brands there is no single "look"; what actually governs quality is a tacit efficacy-driven main-image grammar that lives in senior designers' heads, never written down — so it can't be taught and can't be checked. Juniors either misjudge and rework repeatedly, or ask seniors for everything and ramp slowly.

瓶颈不在「画得快不快」,而在「判断得准不准」。

The bottleneck isn't how fast you draw — it's how accurately you judge.

所以我把问题钉在生成之前的那一环——「标准对齐与判断」,刻意不碰已有工具覆盖的出图产能。这既不与现有工具重叠,也正对准老板真正要的目标:加速新人上手。

So I anchored the problem at the step before generation — standard alignment and judgment — deliberately not touching the image-generation throughput that existing tools already cover. This neither overlaps with current tooling nor misses the actual goal: ramping juniors faster.

02 | 方案:两层标准 + 对齐助手 The Solution: A Two-Layer Standard + an Alignment Assistant

核心思路:把资深设计师脑中那套隐性语法显性化、可检查化,做成一个新人上传稿子即得判断的对齐助手。它不生成图片(不碰出图工具)、不替代设计师——只把「判断力」做成随时可用的工具。

The core idea: make that tacit grammar explicit and checkable, as an assistant where a junior uploads a draft and gets a judgment back. It does not generate images (it stays away from the production tools) and does not replace designers — it turns judgment itself into an always-available tool.

两层标准是方案的战略核心:

The two-layer standard is the strategic core:

• 第一层 · 不变的功效语法:功效大标题、问题/利益点、产品主角、信任背书、背景克制、底部转化模块。

• 第二层 · 随品类的视觉皮肤:颜色、人物、气质、场景。

• Layer 1 · the invariant efficacy grammar: a benefit headline, problem/benefit points, the product as hero, trust endorsements, a restrained background, a bottom conversion module.

• Layer 2 · a per-category visual skin: color, models, mood, scene.

这意味着标准锁的是「转化语法」、放开的是「视觉皮肤」。所以它能正确处理多品牌的现实——婴童护理的粉白、不同品类的不同气质,都不会被误判成跑偏;真正的跑偏,是语法塌掉(变成种草氛围、功效不前置、产品被弱化)。

So the standard locks the conversion grammar and frees the visual skin. That lets it handle a multi-brand reality correctly: baby-care pink-white, different moods per category — none flagged as off-standard. The real failure is when the grammar collapses (it turns into lifestyle mood, efficacy isn't foregrounded, the product gets weakened).

我还引入了一条「视觉说服范式」判别轴(功效驱动 / 品牌驱动 / 混合),让系统判的是「这张图靠什么说服用户」,而不只是数「有没有标题、有没有产品」——避免把高端品牌感误当成功效型主图的加分项。

I also added a visual-persuasion paradigm axis (efficacy-driven / brand-driven / hybrid), so the system judges what an image persuades with, not just whether it has a headline and a product — preventing it from mistaking high-end brand polish for an efficacy-image merit.

03 | 能跑的 MVP A Working MVP

我在 Dify 上搭了一个双模式 agent,以「有没有上传图片」分流,覆盖新人两类真实场景:

I built a dual-mode agent on Dify, branching on whether an image was uploaded, to cover a junior's two real situations:

• 判图(主功能):有图 → 视觉分析 → 知识检索 → 图片判断 → 输出「总体判断 / 符合点 / 跑偏点 / 修改建议 / 可参考案例」

• 问答(辅助):无图 → 知识检索 → 文字回答,日常查标准、问设计边界

• Judge-image (primary): image in → visual analysis → knowledge retrieval → image judgment → output: overall verdict / on-standard / off-standard / fixes / reference cases.

• Q&A (secondary): no image → knowledge retrieval → text answer, for everyday questions about the standard and design boundaries.

几个设计取舍:

A few design tradeoffs:

• 先观察、后判断分离:先让模型客观描述图里有什么,再对照标准下判断,避免「边看边下结论」的草率。

• 判断前必过知识库:判定锚在显性标准 + 案例库(RAG)上,不靠模型凭空感觉,结果可追溯、可解释。

• 全链路用最小模型(gpt-4o-mini):运行成本近乎为零——而且下一节会看到,能力瓶颈根本不在模型。

• Separate observe-then-judge: first have the model objectively describe the image, then judge against the standard — avoiding snap conclusions.

• Judgment must pass the knowledge base: verdicts are anchored to the explicit standard + a case library (RAG), not the model's free-floating intuition — so results are traceable and explainable.

• Smallest model end-to-end (gpt-4o-mini): running cost is near zero — and as the next section shows, the capability bottleneck isn't the model at all.

04 | 关键迭代:v1.0 → v1.1 Key Iteration

从 v1.0 到 v1.1,系统暴露出两个判断错误。两处都靠系统设计修正,而不是换更贵的模型。

From v1.0 to v1.1, the system surfaced two judgment errors. Both were fixed through system design, not a pricier model.

问题一:把「粉色」误判为跑偏。 v1.0 把一张婴童护理品类的主图(粉白调)判成偏离标准。诊断:系统把「某个对标样本的配色」当成了标准本身。修复:引入两层模型——语法不变、皮肤随品类。验证:同一张图被判为「高度接近」,粉色不再被冤判。

Error 1: pink misjudged as off-standard. v1.0 flagged a baby-care main-image (pink-white) as deviating. Diagnosis: the system had mistaken one reference sample's palette for the standard itself. Fix: the two-layer model — grammar fixed, skin per category. Result: the same image was re-judged "highly aligned"; pink was no longer wrongly penalized.

问题二:品牌驱动型主图判得太松。 v1.0 把一张高端护肤品牌的品牌驱动型主图判成「基本接近」,甚至把「高端感」当加分项。诊断:系统在数表面要素(有没有标题 / 产品 / 背书),没判「靠什么说服用户」。修复:把视觉分析结构化,新增「视觉说服范式」判别轴。验证:同一张图从「基本接近」降为「部分接近」,范式被正确区分。

Error 2: a brand-driven image judged too leniently. v1.0 rated a high-end skincare brand's brand-driven main-image as "broadly aligned," even crediting its "premium feel." Diagnosis: the system was counting surface elements (headline / product / endorsement) instead of judging what it persuades with. Fix: structure the visual analysis and add the persuasion-paradigm axis. Result: the same image dropped from "broadly aligned" to "partially aligned"; the paradigm was correctly distinguished.

以下案例图为原创 mock 图，品牌与商品均为虚构，仅用于展示系统的判断逻辑。
Mock images below are original; brands and products are fictional, shown only to illustrate the system's judgment logic.

迭代验证 · 从「要素识别」升级为「范式判断」

测试现象

初版系统在判断高端护肤图时，容易把「产品清晰、标题存在、品牌感强」误认为符合功效驱动型标准，因此对品牌驱动型主图判得偏松。

A/B 测试结果

在相同任务下，将基础模型替换为更强模型后，判断结论并没有明显改善。说明问题的主要瓶颈不在模型能力，而在任务定义：系统只是在数「有没有标题、有没有产品、有没有背书」，没有判断这张图到底靠什么说服用户。

修正方式

v1.1 中新增「视觉说服范式」判别轴，将图像先区分为：功效驱动型 / 品牌驱动型 / 混合型 / 其他。同时加入两层判断模型：① 稳定语法——功效标题、问题 / 利益点、产品主角、信任背书、底部转化模块；② 品类皮肤——颜色、人物、气质、场景随品类变化。

迭代结果

系统不再单纯判断「要素齐不齐」，而是进一步判断「这张图的核心说服逻辑是什么」。因此，粉色婴童护理图可以被判为高度接近，而高端品牌海报型护肤图会被识别为部分接近。

项目价值

这次迭代证明：AI 判断类工具的关键不只是换更强模型，而是把任务定义设计得更准确。通过结构化观察、知识库检索和范式判断，系统才能更接近真实设计主管的判断逻辑。

同一任务换更大的模型,结果几乎不变——瓶颈不在模型能力,在任务定义。

Swap in a bigger model on the same task and the result barely moves — the bottleneck isn't model capability, it's task definition.

05 | 诚实的价值与局限 Honest Value & Limits

我刻意只把最硬、最可辩护的一块做量化:新人因「判断不准」产生的返工工时。按一组保守的、可替换的示意假设算下来,直接可省的工时约 ¥0.7–2 万 / 年,随团队扩招线性放大。这个数字不大——它是 ROI 的保底线,不是卖点。

I deliberately quantified only the hardest, most defensible piece: rework hours from juniors' misjudgment. On conservative, replaceable illustrative assumptions, the directly-saveable hours come to roughly ¥7k–20k/year, scaling with hiring. It's a small number — the floor of the ROI, not the pitch.

真正值钱的在战略层,我如实标为方向性上行、需真实数据验证,不编成数字塞进总账:

The real value is strategic, which I marked honestly as directional upside pending real data rather than forcing into the total:

• 新人上手提速(老板真正的目标)——一个 6 周就靠谱的新人,远胜 12 周才靠谱的。

• 更多主图落在被验证过的转化语法上——主图是点击命门,转化的微小改善对应的 GMV 增量,量级上通常压过全部工时节省。此项须用真实转化数据验证,不作承诺。

• 可复制资产——同一套「标准 + 判断助手」可平移到其他品类与物料,边际成本低。

• Faster ramp for juniors (the actual goal) — a junior reliable in 6 weeks beats one who takes 12.

• More main-images on a validated conversion grammar — the main-image is the click pressure-point; a small conversion lift usually dwarfs all hour-savings in GMV terms. This must be verified with real conversion data; no promises made.

• A replicable asset — the same "standard + judgment assistant" method ports to other categories and materials at low marginal cost.

诚实的局限:① 小模型偶尔仍会把「品牌型背书」误算为加分项(表面匹配的残留),下一版用更细的背书分类解决;② 目前只覆盖「商品主图」一类,其他物料需另立标准;③ 价值验证仍基于假设,需用真实业务数据(出图量、返工轮次、转化率)校准。

Honest limits: (1) the small model still occasionally counts brand-type endorsement as a merit (residual surface-matching); a finer endorsement taxonomy fixes it next version. (2) It covers only product main-images for now; other materials need their own standard. (3) The value case still rests on assumptions and needs calibration against real business data (image volume, rework rounds, conversion).