One product image → a full setof commercial photos in different angles / scenes / compositions.
E-commerce main image slots usually need 5 images, and the detail page needs a dozen or more. This skill solves the problem of "having only one image but needing to fill a whole screen":The same product, multiple camera positions and scenes, with consistent visual identity。
Generation result example
| Input: product image |
|---|
product-watch.jpg — black crocodile-pattern leather strap, steel case, silver sunburst dial watch |
The command actually executed (set image 2; the other two only change the third camera-position segment):
dlazy gpt-image-2 \
--prompt 'E-commerce product photography, set image 2 of 3 — in-use lifestyle shot. The subject is the watch from the reference image: a polished stainless-steel watch with a silver sunburst dial, applied baton markers and a black crocodile-embossed leather strap. Keep the product 100% faithful: same case shape and polish, same dial colour and marker layout, same hand shapes, same crown, same strap embossing and stitching — it must be recognisably the identical watch as the reference. Show it worn on a man wrist resting on a wooden cafe table beside a white coffee cup, dark suit sleeve and white shirt cuff visible, warm window light, shallow depth of field with a blurred cafe background. Photorealistic, no text, no watermark.' \
--images docs/fission-pattern/product-watch.jpg \
--size 1024x1536 --quality medium --imageFormat jpeg \
--save docs/fission-pattern/example-output-2.jpg
Output: a set of three images
| 1 · Front main image | 2 · Scene in-use image | 3 · Detail macro image |
|---|---|---|
| Blueprint paper + brass straight ruler, cool-toned side light | Worn on wrist + coffee table, warm window light | Macro of dial/markers/crown, hard light rim |
The product fidelity segment of the three images is word-for-word identical; only the camera-position segment changes.
1. Capability boundaries
| Mode | Description |
|---|---|
| Product set images | Same product → front main image / 45-degree image / scene in-use image / detail macro image / size comparison image |
| Pose set images | Same model, same outfit → front / side / back / walking / sitting |
| Input | Description |
|---|---|
| Product image | 1 image, best from the main viewpoint |
| Product name | Example:撞色长款风衣 |
| Product selling points | Example:100% 纯棉，轻盈舒适透气，法式复古撞色元素(used to determine scene and atmosphere) |
Do not do: do not change the product's shape, color, material, or structure; do not invent functional selling points the product does not have; do not generate false promotional information.
2. Input asset rules
Before generating, self-check these hard constraints:
- Size:20KB ~ 15MB
- Resolution:greater than 400×400
- Format:jpg / jpeg / png / webp
Input recommendations
| Practice | Description |
|---|---|
| ✅ Main viewpoint + clean background | The cleaner it is, the higher the product consistency across the whole set |
| ✅ Write selling points concretely | 防水防汗 will bring out a sports scene, and法式复古 will bring out a cafe scene |
| ✅ Product structure fully visible | The detail images in the set rely on this image to infer structure |
| ❌ Images that already contain marketing text | The text will be copied into every image in the set; first use remove-watermark to clean it |
| ❌ Product blocked by hands/props | The blocked parts will differ in every image in the set |
3. Set recipe: how to arrange the 5 main image slots
Break "a set of images" into a fixed camera-position checklist, one prompt per image;the product description segment is fully reused, only the camera-position segment changes：
| # | Camera position | Purpose | Shot Position Segment Example |
|---|---|---|---|
| 1 | Front Main Image | First image in search listings, must be the clearest | straight-on hero shot filling the frame, clean seamless background, even studio light |
| 2 | 45-degree 3D View | Convey volume and thickness | 45-degree three-quarter angle showing depth and side profile, soft gradient background |
| 3 | Scene Usage Image | Build usage associations | in-use lifestyle shot: [场景 + 人物动作], warm window light, shallow depth of field |
| 4 | Detail Macro Image | Prove material and craftsmanship | macro close-up of [关键工艺部位] filling the frame, hard rim light, extreme detail |
| 5 | Comparison / Interior View | Convey dimensions or interior | [尺寸对比物 / 内部结构] shown alongside the product, top-down layout, neutral background |
The Key to Consistency: In the three-segment structure, the first segment (faithful product description) is in all 5 promptsword-for-word identical, only the third segment (shot position) changes.
[段1 商品保真：不变] + [段2 卖点氛围：不变] + [段3 镜位：每张不同]
4. Mapping Selling Points → Scenes
Selling points determine what the 3rd scene image looks like; don't let the model improvise:
| Selling Point Type | Scene Writing |
|---|---|
| Warmth / Thickened | snowy outdoor street, breath visible, cold blue ambient with warm rim light |
| Breathable / Quick-dry | gym or running track, dynamic mid-motion, bright daylight |
| Waterproof / Three-proof | rainy pavement with water droplets beading on the surface, overcast light |
| Commute / Business | office lobby or subway station, dark suit context, cool neutral light |
| French / Vintage | French cafe interior, marble table, warm window light, film colour grading |
| Skin-friendly / Baby | soft nursery bedding, pastel palette, very soft diffused light |
| Precision / Craftsmanship | dark navy blueprint paper with a brass ruler, cool directional side light |
Pose Setuses another set: front standing pose / side turn head / back look over shoulder / natural walking / sitting pose—each one only changes the pose segment, while the model and outfit segments remain word-for-word unchanged (the writing can directly reuse creative-scene 's pose-changing sentence pattern).
5. Tool Invocation
This skill uses dLazy's gpt-image-2(the core metric of a set isthe product in the entire set is the same item, requiring the strongest reference image fidelity; in practice, low-cost models drift to a different product when switching between multiple scenes).
Invocation Method
Two equivalent ways to write it, choose one. The unified entry point automatically selects the backend, retries on failure, creates directories and writes to disk, and estimates cost:
# A. 统一入口（推荐）：可切任意后端，加 --dry-run 不计费空跑
node scripts/gen.mjs --task fission-pattern \
--prompt '<见下方 Prompt 模板>' \
--images <按下表顺序> \
--save output/fission-pattern-<sku>.jpg
# B. 直接用 dLazy CLI（不想引入 Node 依赖时，效果等价）
dlazy gpt-image-2 --prompt '...' --images ... --save output/fission-pattern.jpg
Parameter Conventions (Fixed Usage for This Skill)
| Parameter | Value | Reason |
|---|---|---|
--images |
[商品图](all N commands in the set pass the same one) |
Ensure the product in the entire set comes from the same source |
--size |
1024x1536 vertical set;1024x1024 square main image slot |
Follow platform main image specifications |
--quality |
medium scene image;high detail macro image |
Macro images rely on texture to speak |
--imageFormat |
jpeg |
Common format for e-commerce listings |
--batch |
1(a set relies on multiple prompts, not batch) |
batch only gives multiple versions of the same shot position |
--save |
docs/fission-pattern/output-<sku>-<序号>.jpg |
Archive by set number |
Cost Tip: To reduce cost, you can switch to
dlazy seedream-5.0(about 1/6 the price per image). The trade-off is lower product fidelity—in practice it drifts to a different product when switching scenes, only suitable for atmospheric images that are not sensitive to product consistency.
Command Examples
# basic call: 套图第 1 张（正面主图）
dlazy gpt-image-2 \
--prompt 'E-commerce product photography, set image 1 of 5 — hero front shot. The subject is the product in the reference image. Keep it 100% faithful: same shape, colour, material texture and structural details. Straight-on hero shot filling the frame, clean seamless background, even studio light. Photorealistic, no text, no watermark.' \
--images docs/fission-pattern/product-watch.jpg \
--size 1024x1536 --quality medium
# complex call: 用 shell 循环一次跑完整套 5 张
PRODUCT='a polished stainless-steel watch with a silver sunburst dial, applied baton markers and a black crocodile-embossed leather strap'
KEEP="Keep the product 100% faithful: same case shape and polish, same dial colour and marker layout, same hand shapes, same crown, same strap embossing and stitching."
i=0
for SHOT in \
'straight-on hero shot filling the frame, clean seamless background, even studio light' \
'45-degree three-quarter angle showing depth and side profile, soft gradient background' \
'in-use lifestyle shot: worn on a man wrist beside a coffee cup on a wooden cafe table, warm window light, shallow depth of field' \
'macro close-up of the dial edge, applied markers and knurled crown filling the frame, hard rim light, extreme detail' \
'top-down flat layout beside a brass ruler for scale, dark navy blueprint paper, cool side light'
do
i=$((i+1))
dlazy gpt-image-2 \
--prompt "E-commerce product photography, set image $i of 5. The subject is $PRODUCT from the reference image. $KEEP $SHOT. Photorealistic commercial product photography, no text, no watermark." \
--images docs/fission-pattern/product-watch.jpg \
--size 1024x1536 --quality medium --imageFormat jpeg \
--save docs/fission-pattern/output-sku001-$i.jpg
done
# 先估价不真跑（乘以套图张数就是总价）
dlazy gpt-image-2 --dry-run --prompt '...' --images a.jpg --size 1024x1536
Further Reading
| What to Look Up | Where to Go |
|---|---|
| Authentication, multi-backend configuration, output structure, error codes | references/provider-cli.md |
gpt-image-2 all available parameters of |
references/model-flags.md |
| All options of the unified entry point | node scripts/gen.mjs --help |
6. Prompt Template
Three-segment structure, the first two segments are reused across the whole set, the third segment differs for each image:
E-commerce product photography, set image [N] of [总数].
【段1 · 商品保真，整套逐字相同】
The subject is [商品名称 + 颜色 + 材质] from the reference image.
Keep the product 100% faithful: same [外形], same [颜色], same [材质纹理],
same [结构细节：五金/缝线/图案/logo 位置] — it must be recognisably the identical product
across the whole set.
【段2 · 卖点氛围，整套逐字相同】
[从第四节表格取对应的氛围与色调描述]
【段3 · 镜位，每张不同】
[从第三节表格取镜位描述]
Photorealistic commercial product photography, no text, no watermark.
Correction Sentences Appended by Issue
| Issue | Append to the end of the prompt |
|---|---|
| In the set, it looks like two different products | Cross-check against the reference image: [关键识别特征] must match exactly. Any deviation is a failure. |
| The scene steals the spotlight from the product | The product must occupy at least 40% of the frame and be the sharpest element; keep the environment subordinate and softly defocused. |
| The overall color tone is not unified | Grade the whole set consistently: [色温 + 对比度描述]. |
| The detail image is blurry | Change --quality high, and append resolve individual [纹理单位：yarn plies / gear teeth / leather pores] |
7. Execution process
- Clean the input: If the product image contains marketing text, first run remove-watermark。
- Write the product fidelity paragraph: List all the key identifying features of the product (shape / color / material / hardware / stitching / logo position) — this paragraph is reused across the whole set.
- Selling points → atmosphere paragraph: Check the table in Section 4, translate the selling points into scenes and color tones, and reuse them across the whole set.
- List the shot positions: Check Section 3, and take 5 entries according to the number of main image slots on the platform (or 8–12 entries for the detail page).
- Run the loop: Use the shell loop from Section 4, changing only the shot position paragraph each time.
- Quality-check the whole set: View the N images side by side — is it the same product? Is the color tone unified? Is any scene stealing the spotlight?
- Rerun only the individual images that fail, no need to redo the whole set.
8. Common issues
| Symptom | Cause | Handling |
|---|---|---|
| The products in the image set are clearly not the same item | The product fidelity paragraph is too vague, or a low-fidelity model was used | List all key identifying features; switch to gpt-image-2 |
| The color tone is different in every image | The atmosphere paragraph was not unified | Make the atmosphere paragraph identical word for word across the whole set, and append a unified color grading sentence |
| The scene is too full, and the product becomes smaller | The scene description is longer than the product description | Append a constraint sentence requiring 40% of the frame |
| The material cannot be seen in the detail image | --quality medium |
Change high and specify the texture unit to be resolved |
| Marketing text appears in the image set | The input image itself contains text | Clean the image first; keep it at the end of the prompt no text, no watermark |
| Want to create 12 images but the cost is too high | Number of images × unit price | Use gpt-image-2for the 5 main image slots, and use seedream-5.0 to supplement the volume for detail page atmosphere images |
Tips
Visit https://dlazy.com for more information.