marketer-ad-variant-factory

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Ad variant factory

广告变体生成工具

Take one approved concept and explode it into 10-30 shippable ad variants for A/B testing on Meta, Google, TikTok, and Pinterest. Built for speed (parallel batch) and for direct upload to ad accounts (deterministic naming).

将一个已通过审核的创意概念扩展为10-30个可直接上线的广告变体，用于Meta、Google、TikTok和Pinterest平台的A/B测试。该工具专为高效批量处理（并行批量操作）和直接上传广告账户设计（采用确定性命名规则）。

When to Use

使用场景

User has a hero / concept and needs many variants across headline × visual × CTA × background for A/B tests.
"Fan out 30 variants of this ad for Meta" / "generate a test matrix" / "multiply this creative".
Prepping a new ad-set launch — needs 9:16, 1:1, 16:9 with 3-5 visual variants each.
Do NOT use for a single hero (use
```
gen-ai generate
```
) or for cross-channel creative (use
```
marketer-campaign-kit
```
). This skill is for depth on one concept, not breadth across channels.

用户拥有主图/创意概念，需要针对标题×视觉元素×行动号召（CTA）×背景生成大量变体用于A/B测试。
需求如“为Meta生成30版该广告的变体”、“生成测试矩阵”、“拓展该创意内容”。
准备新广告组上线——需要9:16、1:1、16:9三种比例，每种比例搭配3-5个视觉变体。
请勿在仅需生成单张主图时使用（请使用
```
gen-ai generate
```
），也勿用于跨渠道创意生成（请使用
```
marketer-campaign-kit
```
）。该工具专注于深化单个创意概念，而非跨渠道拓展。

Prerequisites

前置条件

Ask up front if the brief doesn't cover it (combine into one message):

Hero asset — path or URL to the approved concept image.
Axes to vary — visual direction (1-5), background/scene (1-5), focal composition (close-up vs wide), optional: color treatment.
Platforms / aspect ratios — Meta needs 9:16 + 1:1, TikTok is 9:16, Display wants 16:9. Confirm which.
Variant count — how many total? 10-15 is typical for a first test, 30+ for broad exploration.
Naming convention — what does the ad platform require (e.g.
```
{campaign}_{axis}_{variant}_{size}.webp
```
)?
Brand guardrails — colors (hex), forbidden elements, existing brand.md?

If the user just says "a lot", default to 5 visuals × 3 ratios = 15 variants.

若需求说明未涵盖以下内容，请一次性询问用户：

主图素材——已通过审核的概念图路径或URL。
可变维度——视觉方向（1-5种）、背景/场景（1-5种）、焦点构图（特写vs全景），可选：色彩处理。
平台/宽高比——Meta需要9:16 + 1:1，TikTok为9:16，展示广告需要16:9。请确认具体需求。
变体数量——总共需要多少？首次测试通常10-15个，广泛探索则需要30+个。
命名规则——广告平台要求的命名格式（例如
```
{campaign}_{axis}_{variant}_{size}.webp
```
）？
品牌规范——品牌色（十六进制代码）、禁用元素、是否有现成的brand.md文档？

若用户仅说“很多”，默认按5种视觉方向×3种比例=15个变体生成。

How to Run

操作步骤

Anchor on the hero. The hero is the reference image — every variant should feel like a sibling, not a cousin. Upload to Drive first if it's local so downstream jobs can reference a URL.
Define the variant matrix. Keep axes explicit. 5 visual directions × 3 aspect ratios = 15 jobs. Don't mix 8 axes — the test becomes unreadable.
Write the manifest. One job per variant, unique
```
id
```
that maps to your ad-platform naming convention. The
```
image
```
field references the hero.

Estimate + dry-run.

bash

gen-ai batch run variants.json --dry-run

Run at concurrency 6-8. Image variants are fast and independent — push concurrency higher than the default 3. Watch for 429s; back off to 4 if you see them.
bash
```
gen-ai batch run variants.json -c 8 -o ./ad-variants
```
Audit and resume. Filter
```
results.json
```
for non-completed jobs, retry.
bash
```
gen-ai batch resume ./ad-variants
```
Hand off. Ads platform uploaders expect a flat folder with standard naming —
```
results.json
```
has every path + URL for direct CSV import to Meta Ads Manager / TikTok Ads / Google Ads.

以主图为基准。主图是参考图像——所有变体都应保持风格统一，而非差异过大。如果是本地图片，请先上传至Drive，以便后续任务可以通过URL引用。
定义变体矩阵。明确可变维度。例如5种视觉方向×3种宽高比=15个任务。不要设置8个维度，否则测试结果将无法解读。
编写清单文件。每个变体对应一个任务，使用唯一的
```
id
```
，与广告平台的命名规则对应。
```
image
```
字段需引用主图。

预估成本+试运行。

bash

gen-ai batch run variants.json --dry-run

设置并发数6-8。图像变体生成速度快且相互独立——将并发数设置为高于默认的3。若遇到429错误，请降至4。
bash
```
gen-ai batch run variants.json -c 8 -o ./ad-variants
```
审核并续跑。筛选
```
results.json
```
中的未完成任务，重新尝试。
bash
```
gen-ai batch resume ./ad-variants
```
交付成果。广告平台上传工具需要一个包含标准命名文件的扁平文件夹——
```
results.json
```
中包含所有文件的路径和URL，可直接导入Meta Ads Manager / TikTok Ads / Google Ads的CSV模板。

Quick Reference

快速参考

Naming convention

{campaign}_{visual}_{composition}_{size}

keeps downstream uploads clean.

json

{
  "defaults": {
    "model": "recraftv4",
    "negativePrompt": "low quality, watermark, busy background",
    "imageUrls": ["https://cdn-pipeline-output.picsart.com/.../hero.webp"]
  },
  "jobs": [
    { "id": "launch_bright_closeup_9x16", "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "9:16" },
    { "id": "launch_bright_closeup_1x1",  "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "1:1"  },
    { "id": "launch_bright_closeup_16x9", "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "16:9" },
    { "id": "launch_studio_wide_9x16",    "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "9:16" },
    { "id": "launch_studio_wide_1x1",     "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "1:1"  },
    { "id": "launch_studio_wide_16x9",    "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "16:9" },
    { "id": "launch_urban_medium_9x16",   "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "9:16" },
    { "id": "launch_urban_medium_1x1",    "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "1:1"  },
    { "id": "launch_urban_medium_16x9",   "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "16:9" }
  ]
}

9 variants from 3 visual × 3 ratios. Scale to 15 or 30 by adding visual rows. Remember: no

count

field — emit one job per variant.

采用

{campaign}_{visual}_{composition}_{size}

的命名规则可简化后续上传流程。

json

{
  "defaults": {
    "model": "recraftv4",
    "negativePrompt": "low quality, watermark, busy background",
    "imageUrls": ["https://cdn-pipeline-output.picsart.com/.../hero.webp"]
  },
  "jobs": [
    { "id": "launch_bright_closeup_9x16", "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "9:16" },
    { "id": "launch_bright_closeup_1x1",  "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "1:1"  },
    { "id": "launch_bright_closeup_16x9", "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "16:9" },
    { "id": "launch_studio_wide_9x16",    "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "9:16" },
    { "id": "launch_studio_wide_1x1",     "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "1:1"  },
    { "id": "launch_studio_wide_16x9",    "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "16:9" },
    { "id": "launch_urban_medium_9x16",   "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "9:16" },
    { "id": "launch_urban_medium_1x1",    "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "1:1"  },
    { "id": "launch_urban_medium_16x9",   "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "16:9" }
  ]
}

通过3种视觉方向×3种比例生成9个变体。可通过添加视觉方向行扩展至15或30个变体。注意：不要使用

count

字段——每个变体对应一个任务。

Quick Reference

模型选择参考

Sub-task	Model	Why
Brand-consistent variants from a hero (default)	`recraftv4`	Strongest at keeping design language consistent across many renders
Photoreal product / lifestyle variants	`flux-2-pro`	Best photoreal adherence, great for Meta/TikTok product ads
Variants with readable headline text baked in	`ideogram-v3`	Only model that reliably renders legible copy — use when you can't overlay
Face/character continuity across variants	`gemini-3-pro-image`	Nano Banana Pro locks subject identity best
Background swaps on a fixed subject	`gen-ai change-bg` (subcommand)	Keeps the subject pixel-identical, only swaps the scene
Ultra-cheap exploration before the flagship run	`gemini-3.1-flash-image`	~5x cheaper, fast — use to pick winning prompts, then regenerate with flux/recraft

Confirm IDs with

gen-ai models --mode image

子任务	模型	选择理由
基于主图生成符合品牌规范的变体（默认）	`recraftv4`	在多次渲染中保持设计语言一致性的表现最佳
写实风格产品/生活方式变体	`flux-2-pro`	写实还原度最高，非常适合Meta/TikTok产品广告
内置清晰标题文本的变体	`ideogram-v3`	唯一能可靠生成清晰可读文字的模型——无法叠加文字时使用
变体中保持人脸/角色一致性	`gemini-3-pro-image`	Nano Banana Pro在保持主体身份一致性方面表现最优
固定主体更换背景	`gen-ai change-bg` （子命令）	保持主体像素完全一致，仅更换场景
旗舰版生成前低成本探索	`gemini-3.1-flash-image`	成本约低5倍，速度快——用于筛选优质提示词，再用flux/recraft重新生成最终版本

可通过

gen-ai models --mode image

确认模型ID。

Procedure

操作规范

Explicit variant axes. Decide 3-4 axes up front (visual, composition, color, setting). Scattershot prompts make A/B results unreadable.
Hero as reference image on every job. Use the
```
image
```
field in
```
defaults
```
— every variant inherits the brand look.
Deterministic naming = direct ad-platform import.
```
{campaign}_{visual}_{composition}_{size}
```
parses cleanly in Meta/TikTok/Google ads CSV templates.
Draft cheap, upgrade winners. Run 30 variants through
```
gemini-3.1-flash-image
```
for ~$1. Pick top 8. Regenerate those 8 through
```
flux-2-pro
```
or
```
recraftv4
```
for the final upload.
Text-safe zones per platform. Meta Stories reserve 250px top + 310px bottom. TikTok reserves ~300px at bottom for UI. Prompt focal into the center 60% of the canvas.
Concurrency 6-8 for image variants. Images are fast — higher concurrency finishes a 30-variant run in under 2 minutes. Drop to 4 if you see 429s.
Never overwrite silently. Unique
```
id
```
per variant means unique output filename — resume is safe and collision-free.
Never claim results in the prompt. "Viral ad, 10M views" doesn't improve output; describe framing, subject, lighting, mood.

明确变体维度。提前确定3-4个维度（视觉、构图、色彩、场景）。随意的提示词会导致A/B测试结果无法解读。
每个任务都引用主图作为参考。在
```
defaults
```
中使用
```
image
```
字段——所有变体都将继承品牌风格。
确定性命名=直接导入广告平台。
```
{campaign}_{visual}_{composition}_{size}
```
格式可在Meta/TikTok/Google广告CSV模板中被清晰解析。
低成本草稿，升级优质变体。用
```
gemini-3.1-flash-image
```
生成30个变体成本约1美元。选出前8个优质变体，再用
```
flux-2-pro
```
或
```
recraftv4
```
重新生成用于最终上传。
各平台文本安全区域。Meta Stories顶部预留250px、底部预留310px。TikTok底部预留约300px用于UI元素。提示词中应指定焦点位于画布中间60%区域。
图像变体生成设置并发数6-8。图像生成速度快——更高的并发数可在2分钟内完成30个变体的生成。若遇到429错误，降至4。
禁止静默覆盖。每个变体使用唯一的
```
id
```
意味着输出文件名唯一——续跑操作安全且不会产生冲突。
请勿在提示词中加入结果预期。“爆款广告，1000万浏览量”不会提升输出质量；应描述构图、主体、光线、氛围。

Pitfalls

常见误区

Too many axes → unreadable A/B. Vary 3-4 at most. If you change visual + composition + color + setting + headline in one variant, you can't isolate the winner.
Missing the hero reference. Without
```
image
```
in defaults, each variant drifts visually — the bundle doesn't feel like one campaign.
Wrong aspect ratio for the platform. TikTok is 9:16 full-bleed, Meta Reels is 9:16, Meta feed is 1:1 or 4:5 (not 1.91:1 anymore), Google Display is 300×250 / 728×90 / 160×600 — check the ad set requirements before fanning out.
Text in the image without localization plan. Baked-in copy blocks localization — keep ad copy in the ads-manager overlay unless it's a one-market run.
Running 100 variants in one go with no draft phase. 100 full-price flagship renders = wasted credits. Draft → pick → upgrade.
Overwriting results on re-runs. Keep output dirs per run (
```
./variants-$(date +%F-%H%M)
```
) so resume + audit work cleanly.

维度过多→A/B测试结果无法解读。最多变化3-4个维度。如果在一个变体中同时更改视觉、构图、色彩、场景、标题，将无法确定哪个因素是获胜关键。
未引用主图。若
```
defaults
```
中没有
```
image
```
字段，每个变体的视觉风格会逐渐偏离——整个系列不会像同一个广告活动的内容。
平台宽高比错误。TikTok是9:16全屏，Meta Reels是9:16，Meta信息流是1:1或4:5（不再是1.91:1），Google展示广告是300×250 / 728×90 / 160×600——生成变体前请确认广告组要求。
图像中包含文本但无本地化计划。内置文本会阻碍本地化——除非仅针对单一市场投放，否则请将广告文案放在广告管理平台的叠加层中。
一次性生成100个变体且无草稿阶段。100次全价旗舰版渲染会浪费额度。应遵循草稿→筛选→升级的流程。
重新运行时覆盖结果。每次运行使用单独的输出目录（例如
```
./variants-$(date +%F-%H%M)
```
），以便续跑和审核操作顺利进行。

Verification

验证方法

Run

gen-ai whoami

to confirm authentication, then re-run the failed command with

--debug

运行

gen-ai whoami

确认身份验证状态，然后添加

--debug

参数重新运行失败的命令。

Cost & time

成本与时间

Variant count	Model	Credits each	Total	Wall time @ concurrency 8
9 variants	`recraftv4`	~2	~18	~45s
15 variants	`recraftv4`	~2	~30	~90s
30 variants (exploration)	`gemini-3.1-flash-image`	~0.5	~15	~2 min
30 variants (flagship)	`flux-2-pro`	~2	~60	~3 min
30 drafts + 8 upgraded	mixed		~30	~4 min total

Always confirm with

gen-ai batch run variants.json --dry-run

and pause if the estimate exceeds the user's cap.

变体数量	模型	单个变体消耗额度	总消耗	并发数8时的耗时
9个变体	`recraftv4`	~2	~18	~45秒
15个变体	`recraftv4`	~2	~30	~90秒
30个变体（探索版）	`gemini-3.1-flash-image`	~0.5	~15	~2分钟
30个变体（旗舰版）	`flux-2-pro`	~2	~60	~3分钟
30个草稿+8个升级变体	混合模型		~30	总计~4分钟

请始终运行

gen-ai batch run variants.json --dry-run

进行预估，若预估额度超出用户上限，请暂停操作。

marketer-ad-variant-factory

Original

Translation

Ad variant factory

广告变体生成工具

When to Use

使用场景

Prerequisites

前置条件

How to Run

操作步骤

Quick Reference

快速参考

Quick Reference

模型选择参考

Procedure

操作规范

Pitfalls

常见误区

Verification

验证方法

Cost & time

成本与时间

See also

相关链接