What other skills deliver isimages, what this skill delivers isa retrospective-ready experiment。
The difference is: generating a few extra images at random is not A/B. A/B requires that between each grouponly one variable differs, and that you write down in advance which group you expect to win and why. Otherwise, after it runs, you still won't know what is actually having an effect.
1. Capability boundaries
| Can do | Description |
|---|---|
| Single-variable control group | Change only one dimension at a time, keep everything else consistent |
| Hypothesis record | Each group comes with "expected effect + reason", so it can be reviewed after running |
| Compliance check | Each image is automatically checked against platform-compliance, to avoid polluting the data due to rejection |
| Retrospective template | Produces a comparison table you can fill in with data |
| Cannot do | Description |
|---|---|
| Connect to platform data | Not connected to the backend; click-through rate and conversion rate must be filled back in by you |
| Judge statistical significance | Don't draw conclusions when the sample size is insufficient; see "FAQ" |
2. Testable variables
Change only one per experiment:
| Variable | A (control) | B (experiment) | Hypothesis |
|---|---|---|---|
| Background | Pure white | Lifestyle scene | Scene images have stronger immersion and a higher click-through rate |
| Shot type | Full body | Half-body close-up | Close-ups are clearer on small screens |
| With or without model | Flat lay image | Model wearing it | For apparel, worn images are more persuasive |
| Copy | No text | Selling point badge in the upper left corner | Text increases information density (pay attention to whether the platform allows it) |
| Model | Asian face | Western face | Match the target market |
| Color tone | Cool white | Warm tone | Warm tones are more appealing in the home category |
Don't change three things at once. If three variables move together, even if it wins, you won't know which one won.
3. Tool invocation
Generate control groups — two commands differ in only one prompt:
# A 组：纯白背景（对照）
node scripts/gen.mjs --task listing-optimizer --brand examples/brand.yaml \
--prompt 'E-commerce main image of the olive cable-knit sweater on a model. Pure seamless white background, RGB 255,255,255. Full body, product fills 88% of the frame. No text, no watermark.' \
--images docs/flat-lay/garment-flatlay.jpg docs/flat-lay/pose-reference.jpg \
--batch 3 --save out/ab/A-white.jpg
# B 组：生活场景（只改这一处）
node scripts/gen.mjs --task listing-optimizer --brand examples/brand.yaml \
--prompt 'E-commerce main image of the olive cable-knit sweater on a model. Warm minimal apartment interior, soft window light. Full body, product fills 88% of the frame. No text, no watermark.' \
--images docs/flat-lay/garment-flatlay.jpg docs/flat-lay/pose-reference.jpg \
--batch 3 --save out/ab/B-lifestyle.jpg
Generate for each group --batch 3 Pick the best one, to avoid treating the luck of a single generation as a conclusion.
Compliance check (rejected images will pollute the experimental data):
python3 scripts/check_listing.py out/ab/*.jpg --platform amazon
4. Retrospective template
Build this table when generating, and fill in the data after launch:
## 实验：主图背景 白底 vs 生活场景
- SKU：A001 军绿麻花针织衫
- 平台 / 坑位：Amazon 主图
- 起止：2026-09-01 ~ 2026-09-14
- 变量：背景（其余保持一致：同模特、同景别、同色调、无文字）
| 组 | 图 | 假设 | 曝光 | 点击 | 点击率 | 加购 | 转化率 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| A 对照 | A-white.jpg | 白底信息干净，符合平台预期 | | | | | |
| B 实验 | B-lifestyle.jpg | 场景图代入感强，点击率更高 | | | | | |
**结论**：
**下一步**：
"Hypothesis" must be written before running. Making up reasons afterward is the most common form of self-deception in retrospectives.
5. Execution process
- Clarify what metric you want to optimizeThe optimization directions for click-through rate and conversion rate are often opposite— A sensational main image can raise click-through rate, but it will lower conversion rate and drive up returns.
- Choose one variable. When the user wants to change three things, tell them to queue up three rounds.
- Write a hypothesis, and have the user confirm it.
- Generate + compliance check。
- Provide a review template, and clearly state that the user needs to fill the data back in themselves.
- Don't draw conclusions when the sample size is insufficient. See below.
Six, common questions
| Symptom | Cause | What to do |
|---|---|---|
| The difference between the two groups is very small | The variable chosen is too weak (for example, only a slight color temperature adjustment) | Switch to a variable with more differentiation |
| The data looks like B is winning but it's uncertain | The impression volume is too small | Look again only after at least several thousand impressions per group; be even more cautious when the difference is less than 10% |
| Click-through rate went up and conversion rate went down | The main image overpromises | The main image must be accurate; don't sacrifice consistency for click-through rate |
| More than one variable differs between groups | Prompt changes were not controlled | Diff the two prompts once to confirm there is only one difference |
Tips
- One variable at a time, one round for two weeks. Being in a hurry to reach a conclusion in one week is usually self-deception.
- Test the big variables first. Whether there is a model, a white background or a scene, differences like these are far greater than slight color temperature adjustments.
- Save the winning version into brand-kit, make it the default from then on, and only then will the experimental results accumulate.