batch-image

batch-image — Multi-product batch image generation

Turn single-image generation into apipeline: one product list → a whole set of commercial shots with unified style.

The hard part of batching is not "running it many times", but thatthe generated images must look like one set. The core of this skill isseparating the spec segment from the variable segment: the visual spec stays word-for-word unchanged, and only the product description varies with the SKU.

Generation result example

Input: product list (2 SKUs)
`SKU001` `军绿色麻花针织圆领毛衣，落肩宽松版型`
`SKU002` `黑色亮面皮革布洛克德比鞋，厚底系带`

The command actually executed (loop body, the two SKUs differ only in DESC and --images , while the spec segment is word-for-word identical):

for sku in a-sweater b-shoes; do
  case $sku in
    a-sweater) DESC='军绿色麻花针织圆领毛衣，落肩宽松版型' ;;
    b-shoes)   DESC='黑色亮面皮革布洛克德比鞋，厚底系带' ;;
  esac
  dlazy seedream-5.0 \
    --prompt "电商商拍图。图1 是商品：${DESC}。商品的颜色、材质纹理、款式细节必须与图1完全一致。放置在同一套统一视觉里：纯米白色摄影棚背景，柔和顶光加左侧补光，45 度视角，画面下方留出统一的商品投影，构图与留白在整组图中保持一致。真实商业产品摄影，无文字无水印。" \
    --images docs/batch-image/sku-${sku}.jpg --size 1:1 --resolution 2k \
    --save docs/batch-image/example-output-${sku}.jpg
done

Output: two SKUs under the same visual spec

SKU001	SKU002

1:1 / 2K，5 credits	1:1 / 2K，5 credits

The two images have consistent background color, lighting position, camera angle, and shadow direction—because the spec segment is word-for-word identical; each product remains faithful. 100 SKUs means running this loop 100 times, at a total cost of about 500 credits.

1. Capability boundaries

Capability	Description
Batch size	Up to 100 products per batch (corresponding to the enterprise feature's specifications)
Cost advantage	In batch scenarios, the compute cost per image can drop to about 70%
Collaboration	Multi-user online collaboration, real-time tracking of image production progress
Asset sharing	Shared models / reference images / works / compute credits
Implementation in this skill	List-driven loop + concurrency control + failure retries + archiving by SKU + summary report

Do not: do not mix different visual specs within the same batch (that is not one batch); do not skip sampling quality checks and list everything directly; do not relax product fidelity requirements just because it is a batch.

2. Input asset rules

Before generation, self-check these hard constraints:

Size:20KB ~ 15MB
Resolution:greater than 400×400
Format:jpg / jpeg / png / webp

Product list (manifest) is the input for batching; CSV is recommended:

sku,image,desc
SKU001,docs/batch-image/sku-a-sweater.jpg,军绿色麻花针织圆领毛衣，落肩宽松版型
SKU002,docs/batch-image/sku-b-shoes.jpg,黑色亮面皮革布洛克德比鞋，厚底系带

Each product image must still satisfy the single-image rules：20KB~15MB、>400×400、jpg/jpeg/png/webp。

Must-dos before batching

Step	Description
✅ First get 1 SKU working	If a single one is wrong, running 100 just means being wrong 100 times
✅ Then sample 5 SKUs	Cover edge cases across categories/colors/light and dark shades
✅ Confirm cost	`--dry-run` unit price × number of SKUs × number of batches
❌ Go straight to full volume	The easiest way to burn compute credits

3. Spec segment / variable segment separation

This is the only way to produce unified visuals in batch:

prompt = [规范段：整批逐字不变]  +  [变量段：每个 SKU 不同]

Segment	Content	Does it change
Spec segment	Background, lighting position, camera position, composition whitespace, shadow, color tone, output style,`no text, no watermark`	Word-for-word unchanged for the whole batch
Variable segment	Product category + color + material + key structural features	Different for each SKU

Write as shell:

SPEC='放置在同一套统一视觉里：纯米白色摄影棚背景，柔和顶光加左侧补光，45 度视角，画面下方留出统一的商品投影，构图与留白在整组图中保持一致。真实商业产品摄影，无文字无水印。'

while IFS=, read -r SKU IMG DESC; do
  dlazy seedream-5.0 \
    --prompt "电商商拍图。图1 是商品：${DESC}。商品的颜色、材质纹理、款式细节必须与图1完全一致。${SPEC}" \
    --images "$IMG" --size 1:1 --resolution 2k \
    --save "docs/batch-image/out/${SKU}.jpg"
done < manifest.csv

Once the specification section is changed, the entire batch must be rerun—otherwise two visual styles will appear within the same batch.

4. Concurrency, Retries, and Progress

Concurrency: dlazy handles one task per call; batching relies on shell concurrency. Keep it at 4–6 parallel jobs; too high easily triggers rate limiting.

run_one() { ... }               # 单个 SKU 的生成逻辑
export -f run_one
xargs -P 5 -I{} bash -c 'run_one {}' < sku-list.txt     # 5 路并发

Retries: Occasional network and server-side failures are normal; wrap it in a retry layer.

for attempt in 1 2 3; do
  dlazy ... && break
  echo "retry $attempt for $SKU" >&2; sleep $((attempt * 10))
done

Async mode: When there are many SKUs, use --no-wait to submit all tasks first, get the generateId and then poll uniformly, avoiding occupying the process for a long time.

GID=$(dlazy seedream-5.0 --no-wait ... | python3 -c 'import sys,json;print(json.load(sys.stdin)["result"]["task"]["generateId"])')
echo "$SKU,$GID" >> tasks.csv
# 稍后统一收结果
while IFS=, read -r SKU GID; do dlazy status "$GID" --wait; done < tasks.csv

Progress and summary: Record one status line per SKU, and output a report when finished.

echo "$SKU,$(test -f out/$SKU.jpg && echo ok || echo fail)" >> report.csv
awk -F, '{c[$2]++} END{for(k in c) print k, c[k]}' report.csv

5. Tool Invocation

This skill uses dLazy's seedream-5.0(the key metrics for batch scenarios arecost per imageandreference image support: this model costs about 5 credits per image, supports up to 10 reference images, and supports 2K/3K/4K; the cost for 100 SKUs drops from hundreds to tens. For batches with extremely high product fidelity requirements, switch to gpt-image-2）。

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 batch-image \
  --prompt '<见下方 Prompt 模板>' \
  --images <按下表顺序> \
  --save output/batch-image-<sku>.jpg

# B. 直接用 dLazy CLI（不想引入 Node 依赖时，效果等价）
dlazy seedream-5.0 --prompt '...' --images ... --save output/batch-image.jpg

Parameter conventions (fixed usage for this skill)

Parameter	Value	Reason
`--images`	`[该 SKU 的商品图]`(+ reference images / model images shared by the entire batch)	Shared assets are placed in a fixed location and referenced in the prompt by fixed image numbers
`--size`	Uniform across the entire batch (`1:1` main image / `3:4` vertical)	If not uniform, it is not a set
`--resolution`	`2k`	Sufficient for listing; 4K only for batches that need printing
`--batch`	`1`(batch scenarios rely on the number of SKUs, not on batch)	batch doubles the cost
`--no-wait`	Enable when there are many SKUs	Submit first, then poll, avoiding long blocking
`--save`	`docs/batch-image/out/<SKU>.jpg`	Name by SKU for easy backfilling into the product database

Model switching：seedream-5.0 is cheap and suitable for atmosphere and scene images; for batches with extremely high product fidelity requirements (pattern position, hardware details), switch to gpt-image-2, at about 6 times the cost. Do not mix the two models within the same batch.

Command Examples

# basic call: 单个 SKU（先跑通这个，再谈批量）
dlazy seedream-5.0 \
  --prompt '电商商拍图。图1 是商品：军绿色麻花针织圆领毛衣，落肩宽松版型。商品的颜色、材质纹理、款式细节必须与图1完全一致。放置在同一套统一视觉里：纯米白色摄影棚背景，柔和顶光加左侧补光，45 度视角，画面下方留出统一的商品投影。真实商业产品摄影，无文字无水印。' \
  --images docs/batch-image/sku-a-sweater.jpg --size 1:1 --resolution 2k

# complex call: 完整流水线（清单驱动 + 5 路并发 + 3 次重试 + 汇总报告）
cat > manifest.csv <<'CSV'
SKU001,docs/batch-image/sku-a-sweater.jpg,军绿色麻花针织圆领毛衣，落肩宽松版型
SKU002,docs/batch-image/sku-b-shoes.jpg,黑色亮面皮革布洛克德比鞋，厚底系带
CSV

SPEC='放置在同一套统一视觉里：纯米白色摄影棚背景，柔和顶光加左侧补光，45 度视角，画面下方留出统一的商品投影，构图与留白在整组图中保持一致。真实商业产品摄影，无文字无水印。'
mkdir -p docs/batch-image/out

run_one() {
  IFS=, read -r SKU IMG DESC <<< "$1"
  for attempt in 1 2 3; do
    dlazy seedream-5.0 \
      --prompt "电商商拍图。图1 是商品：${DESC}。商品的颜色、材质纹理、款式细节必须与图1完全一致。${SPEC}" \
      --images "$IMG" --size 1:1 --resolution 2k \
      --save "docs/batch-image/out/${SKU}.jpg" >/dev/null 2>&1 && break
    sleep $((attempt * 10))
  done
  if [ -f "docs/batch-image/out/${SKU}.jpg" ]; then echo "${SKU},ok"; else echo "${SKU},fail"; fi
}
export -f run_one; export SPEC

xargs -P 5 -I{} bash -c 'run_one "{}"' < manifest.csv | tee docs/batch-image/report.csv
awk -F, '{c[$2]++} END{printf "\n=== 汇总 ===\n"; for(k in c) printf "%s: %d\n", k, c[k]}' docs/batch-image/report.csv

# 先估价：单价 × SKU 数
dlazy seedream-5.0 --dry-run --prompt '...' --images a.jpg --size 1:1 | grep estimatedCost
echo "SKU 数：$(wc -l < manifest.csv)"

What to look up	Where to go
Authentication, multi-backend configuration, output structure, error codes	`references/provider-cli.md`
`seedream-5.0` all available parameters of	`references/model-flags.md`
All options of the unified entry point	`node scripts/gen.mjs --help`

6. Prompt Templates

【变量段 · 每个 SKU 不同】
电商商拍图。图1 是商品：[品类 + 颜色 + 材质 + 关键结构特征]。
商品的颜色、材质纹理、款式细节必须与图1完全一致。

【规范段 · 整批逐字不变】
放置在同一套统一视觉里：[背景]，[光位]，[机位/视角]，
画面下方留出统一的商品投影，构图与留白在整组图中保持一致。
真实商业产品摄影，无文字无水印。

Correction sentences appended by issue(if changed, the entire batch must be rerun)

Issue	Handling
Dark products look gray on a light background	Add to the specification section `深色商品需加右侧轮廓光勾边，避免与背景糊在一起`
Product size varies from SKU to SKU	Add to the specification section `商品在画面中的占比统一为约 [X]%，边距一致`
Inconsistent shadow direction	Add to the specification section `投影统一朝画面右下方，长度约为商品高度的 [X]`
Some SKUs lack sufficient fidelity	Rerun those SKUs separately with `gpt-image-2` and note it in the report

7. Execution Flow

Build the manifest：sku,image,desc Three-column CSV; each product image first goes through single-image rule validation.
Write the specification section: background / lighting position / camera position / proportion / shadow / color tone / output style—no changes allowed once finalized。
Run through 1 SKU: confirm both the specification section and the variable section are correct.
Sample 5 SKUs: cover the boundaries of dark / light / reflective / large items / small items.
--dry-run Estimate total price: unit price × number of SKUs.
Full run: 5-way concurrency + 3 retries; when there are many SKUs, use --no-wait and poll uniformly after submission.
Check the summary report: rerun failures individually.
Sampling quality check: randomly sample 10% to pass detect-task; for non-compliant SKUs, individually use gpt-image-2 to rerun.

8. Common issues

Symptom	Cause	Handling
Visual inconsistency within the same batch	The specification section was changed midway	No changes allowed once the specification section is finalized; if changed, rerun the entire batch
Dark products blend into the background	Missing rim light	Add a rim light sentence to the specification section and rerun the entire batch
Product proportion varies too much	Proportion not constrained	Add proportion and margin sentences to the specification section
Some SKUs lack sufficient fidelity	Limitations of low-cost models	For those SKUs, individually use `gpt-image-2` to rerun
A bunch of failures halfway through the run	Concurrency too high, triggering rate limiting	Reduce concurrency to 4–5, add retries and backoff
Cost exceeds budget	Used `--batch` or 4K	For batch scenarios `--batch 1` + `2k`
After the run, don't know which ones failed	No report	Use report.csv from Section 4 + summary

Tips

Visit https://dlazy.com for more information.

About this skill