brand-kit

brand-kit — Store visual consistency

The difference between a set of images and one store lies inwhether the 200th SKU still looks like the 1st。

Rewriting the prompt every time can't achieve this—people forget, change words, switch phrasing. Only a reusable spec file can.

1. What it solves

Without brand-kit	With brand-kit
Each skill writes its own model description	Defined in one place, all skills read the same one
Model faces differ across SKUs in the same store	Locked to the same face reference image
Color temperature drifts cold and warm, background drifts dark and light	Light position, color tone, and background hard-coded in the spec
Changing the tone once means digging through 26 prompts	Change one line of yaml

2. Quick start

# 1. 生成模板
node scripts/brand.mjs --init > brand.yaml   # 或直接改随技能分发的 examples/brand.yaml

# 2. 改成自己店铺的（见下方字段说明）

# 3. 看它会给某个技能追加什么
node scripts/brand.mjs --brand examples/brand.yaml --for flat-lay

# 4. 生成时带上，自动追加
node scripts/gen.mjs --task flat-lay --brand examples/brand.yaml \
  --prompt 'On-model photo of the olive cable-knit sweater.' \
  --images garment.jpg pose.jpg --save out/sku001.jpg

--brand Works for gen.mjs / run_loop.mjs / batch.mjs / video.mjs . Add it when generating a whole batch, and the visual tone of the entire batch is locked in.

3. Fields

brand:
  name: 示例品牌
  tone: quiet minimalist, warm and lived-in, never glossy or commercial

model:
  reference: assets/model/face-a.jpg          # 锁脸
  description: East Asian woman, late twenties, natural makeup
  body: slim, height around 168cm

photography:
  background: seamless off-white studio backdrop, RGB 248 248 246
  lighting: soft large softbox from camera left, gentle fill, no hard shadows
  camera: 85mm equivalent, eye level, shallow depth of field
  grade: neutral white balance around 5200K, low contrast
  crop: full body with headroom, product centered

layout:
  margin: at least 8% empty margin on all sides
  typeface: clean sans-serif, no decorative fonts

forbid:
  - no text or watermark
  - no oversaturated colors

compliance:
  platform: amazon

Only the relevant sections go into the prompt: the detail-page skill reads layout but not model, the watermark-removal skill reads neither. For the full mapping table see references/brand-schema.md。

4. Three lessons for writing a good spec

1. Write visual constraints in English, notes in Chinese. Image models respond more consistently to English visual terms, tone / background / lighting write these directly in English;name ones like this are just for people, so be casual.

2. Write physical quantities, not adjectives. warm lighting The model interprets it differently every time; neutral white balance around 5200K, soft large softbox from camera left only that is reproducible. Write color values directly as RGB.

3. forbid is more effective than tone . Positive descriptions are easily diluted; prohibitions are hard constraints. "What not to do" often stabilizes style better than "what to do."

5. Face lock

model.reference Give a clear front-facing face image, and it will be passed to the model as thelast reference image, working together with the Model must stay consistent in the prompt.

Requirement	Description
Front-facing, clear, evenly lit	Side profiles and backlighting won't lock
Face only, no complex background	Background will pollute the generation
Sufficient resolution	At least 512×512
One fixed image for the whole store	Changing the image equals changing the person

When the file isn't found, it only warns without erroring, and generation proceeds as usual—with one fewer reference image, the face is determined by the prompt description.

6. Execution flow

When the user says "unify the style," first ask three things:Whether the model should be locked、Main color tone、Whether there are elements that must be avoided。
--init Generate the template and fill it in based on the answers.
Run one SKU once with --brand and once without for comparison, and have the user confirm the tone is right.
After confirmation, run the whole batch—the spec is meant for the whole batch; a single image won't show its value.
Every subsequent image generation should include --brand. Without it, it's as if there's no specification.

VII. Common Issues

Symptom	Cause	What to Do
Adding brand makes it look less alike	Descriptions conflict with each other (for example, `tone` says minimalist,`photography` writes dramatic lighting)	Read through it once and delete contradictory items
The model's face still changes	The reference image is not a frontal face, or the resolution is too low	Replace it with a standard frontal face image
The prompt is too long and gets truncated	The specification is written too finely	Keep only the items that truly affect the visuals, and delete irrelevant paragraphs
The detail page image includes the model description	Used `--for` in a way other than manual stitching	Use `--brand`, it will automatically filter by skill

Tips

Put the specification into version control, together with the product data. It is an asset, not a temporary file.
One per store. Do not share one copy across multiple stores and keep changing it; store each separately.
Seasonal changes can be branched：brand-aw25.yaml、brand-ss26.yaml, switching is just changing the file name.