簡介
賣家有設計、標籤——現在需要產品清單：SEO友善的標題、有吸引力的描述和專業的模型圖像。本文使用LLM進行文案寫作，並使用透視變換來產生模型。
1.產品內容生成架構
Design Tags Seller Preferences
(style, theme, colors) (brand, tone, marketplace)
│ │
└──────────┬───────────────────┘
│
┌────────▼────────┐
│ LLM Content │
│ Generation │
│ (GPT-4 / Llama) │
└────────┬────────┘
│
┌──────────┼──────────┐
▼ ▼ ▼
Title Description SEO Tags
│
┌────────▼────────┐
│ Mockup Engine │
│ (CV Transform) │
└────────┬────────┘
│
┌──────────┼──────────┐
▼ ▼ ▼
Front Mock Lifestyle Detail Shot
2. LLM 產品文案
from openai import AsyncOpenAI
class ProductCopywriter:
"""Generate product title & description bằng LLM"""
def __init__(self):
self.client = AsyncOpenAI()
async def generate_listing(
self,
tags: DesignTags,
product_type: str = "t-shirt",
marketplace: str = "etsy",
language: str = "en",
) -> ProductListing:
# Build context from tags
tag_context = self._build_tag_context(tags)
# Marketplace-specific rules
rules = self._marketplace_rules(marketplace)
prompt = f"""Generate a product listing for a {product_type}.
Design characteristics:
{tag_context}
Marketplace: {marketplace}
{rules}
Return JSON:
{{
"title": "SEO-optimized title ({rules['title_max_chars']} chars max)",
"description": "Compelling product description",
"bullet_points": ["feature 1", "feature 2", ...],
"seo_tags": ["tag1", "tag2", ...]
}}"""
response = await self.client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{
"role": "system",
"content": "You are an expert e-commerce copywriter."
},
{"role": "user", "content": prompt},
],
response_format={"type": "json_object"},
temperature=0.7,
)
data = json.loads(response.choices[0].message.content)
return ProductListing(**data)
def _marketplace_rules(self, marketplace: str) -> dict:
rules = {
"etsy": {
"title_max_chars": 140,
"guidelines": (
"Title should be keyword-rich. "
"Use format: [Main keyword] - [Description] "
"- [Gift occasion]. "
"Include 13 SEO tags."
),
},
"amazon": {
"title_max_chars": 200,
"guidelines": (
"Title format: Brand + Product + "
"Key Feature + Material + Size. "
"Description: 5 bullet points + paragraph."
),
},
"shopify": {
"title_max_chars": 70,
"guidelines": (
"Short, clean title. "
"Rich HTML description with features."
),
},
}
return rules.get(marketplace, rules["shopify"])
def _build_tag_context(self, tags: DesignTags) -> str:
lines = []
if tags.style:
styles = [t.label for t in tags.style[:3]]
lines.append(f"Style: {', '.join(styles)}")
if tags.theme:
themes = [t.label for t in tags.theme[:3]]
lines.append(f"Theme: {', '.join(themes)}")
if tags.colors:
lines.append(f"Colors: {', '.join(tags.colors)}")
if tags.mood:
moods = [t.label for t in tags.mood[:2]]
lines.append(f"Mood: {', '.join(moods)}")
if tags.audience:
audiences = [t.label for t in tags.audience[:2]]
lines.append(f"Audience: {', '.join(audiences)}")
return "\n".join(lines)
3. 模型產生引擎
import cv2
import numpy as np
from PIL import Image
class MockupGenerator:
"""Generate product mockups bằng perspective transform"""
# Template library
MOCKUP_TEMPLATES = {
"flat_lay": {
"image": "mockups/tshirt_flat_lay.png",
"corners": [(320, 180), (680, 180), (700, 550), (300, 550)],
"mask": "mockups/tshirt_flat_lay_mask.png",
},
"model_front": {
"image": "mockups/model_male_front.png",
"corners": [(285, 210), (515, 210), (530, 480), (270, 480)],
"mask": "mockups/model_male_front_mask.png",
},
"hanger": {
"image": "mockups/tshirt_hanger.png",
"corners": [(310, 200), (690, 200), (710, 560), (290, 560)],
"mask": "mockups/tshirt_hanger_mask.png",
},
}
def generate_mockup(
self,
design: Image.Image,
template_name: str = "flat_lay",
shirt_color: str = "#FFFFFF",
) -> Image.Image:
template = self.MOCKUP_TEMPLATES[template_name]
# Load template image and mask
bg = cv2.imread(template["image"], cv2.IMREAD_UNCHANGED)
mask = cv2.imread(template["mask"], cv2.IMREAD_GRAYSCALE)
# Apply shirt color to template
bg = self._apply_shirt_color(bg, mask, shirt_color)
# Perspective transform design onto template
design_cv = cv2.cvtColor(np.array(design), cv2.COLOR_RGBA2BGRA)
# Source corners (design image)
h, w = design_cv.shape[:2]
src_corners = np.float32([
[0, 0], [w, 0], [w, h], [0, h]
])
# Destination corners (on template)
dst_corners = np.float32(template["corners"])
# Compute perspective transform
matrix = cv2.getPerspectiveTransform(
src_corners, dst_corners
)
warped = cv2.warpPerspective(
design_cv, matrix, (bg.shape[1], bg.shape[0]),
flags=cv2.INTER_LINEAR,
)
# Create mask for warped design
design_mask = np.zeros(
(bg.shape[0], bg.shape[1]), dtype=np.uint8
)
cv2.fillConvexPoly(
design_mask, dst_corners.astype(int), 255
)
# Blend design onto template
result = self._blend_design(bg, warped, design_mask, mask)
return Image.fromarray(
cv2.cvtColor(result, cv2.COLOR_BGRA2RGBA)
)
def _apply_shirt_color(
self, bg, mask, hex_color: str
) -> np.ndarray:
"""Apply shirt color to template using mask"""
r = int(hex_color[1:3], 16)
g = int(hex_color[3:5], 16)
b = int(hex_color[5:7], 16)
# Create color overlay
overlay = bg.copy()
overlay[mask > 128] = [b, g, r, 255]
# Blend with multiply mode
result = cv2.addWeighted(bg, 0.3, overlay, 0.7, 0)
return result
def generate_all_mockups(
self,
design: Image.Image,
shirt_color: str = "#FFFFFF",
) -> dict[str, Image.Image]:
"""Generate all mockup variants"""
return {
name: self.generate_mockup(design, name, shirt_color)
for name in self.MOCKUP_TEMPLATES
}
4. 批量產品創建
class ProductCreationPipeline:
"""End-to-end product creation pipeline"""
def __init__(self):
self.tagger = AutoTagPipeline()
self.copywriter = ProductCopywriter()
self.mockup_gen = MockupGenerator()
async def create_product(
self,
design: Image.Image,
shirt_colors: list[str],
marketplaces: list[str],
) -> Product:
# 1. Auto-tag
tags = await self.tagger.tag_design(design)
# 2. Generate listings per marketplace
listings = {}
for marketplace in marketplaces:
listing = await self.copywriter.generate_listing(
tags, marketplace=marketplace
)
listings[marketplace] = listing
# 3. Generate mockups per color
mockups = {}
for color in shirt_colors:
color_mockups = self.mockup_gen.generate_all_mockups(
design, shirt_color=color
)
mockups[color] = color_mockups
return Product(
design=design,
tags=tags,
listings=listings,
mockups=mockups,
)
總結
人工智慧產品生成：
- LLM文案 - GPT-4o-mini產生標題、描述、SEO標籤
- 多市場 — Etsy、Amazon、Shopify 的格式規則
- 模型引擎 — 將設計透視轉換到範本照片上
- 批次管道 — 一個工作流程中的標記 → 複製 → 模型
下一篇文章：趨勢檢測和內容審核。