Spaces:
Sleeping
Sleeping
Lcmind commited on
Commit ยท
b895841
1
Parent(s): f94d4dc
fix: S-tier prompt - hex to color name, blur UI text, remove watermarks
Browse files- app/core/config.py +5 -2
- app/services/qwen.py +94 -0
app/core/config.py
CHANGED
|
@@ -37,6 +37,9 @@ class Settings(BaseSettings):
|
|
| 37 |
gemini_model: str = "gemma-3-27b-it"
|
| 38 |
groq_model: str = "meta-llama/llama-4-maverick-17b-128e-instruct"
|
| 39 |
|
|
|
|
|
|
|
|
|
|
| 40 |
# Pyppeteer Settings
|
| 41 |
puppeteer_executable_path: str = "/usr/bin/chromium"
|
| 42 |
puppeteer_args: List[str] = [
|
|
@@ -46,8 +49,8 @@ class Settings(BaseSettings):
|
|
| 46 |
"--disable-gpu",
|
| 47 |
]
|
| 48 |
|
| 49 |
-
# ๋ถ์ ๋ชจ๋ธ ์ ํ: "gemini"
|
| 50 |
-
analysis_model: str = "
|
| 51 |
|
| 52 |
class Config:
|
| 53 |
env_file = ".env"
|
|
|
|
| 37 |
gemini_model: str = "gemma-3-27b-it"
|
| 38 |
groq_model: str = "meta-llama/llama-4-maverick-17b-128e-instruct"
|
| 39 |
|
| 40 |
+
# Qwen Model (๋ฉํฐ๋ชจ๋ฌ)
|
| 41 |
+
qwen_model: str = "Qwen/Qwen3-VL-32B-Instruct" # 32B ๋ชจ๋ธ๋ก ๊ต์ฒด
|
| 42 |
+
|
| 43 |
# Pyppeteer Settings
|
| 44 |
puppeteer_executable_path: str = "/usr/bin/chromium"
|
| 45 |
puppeteer_args: List[str] = [
|
|
|
|
| 49 |
"--disable-gpu",
|
| 50 |
]
|
| 51 |
|
| 52 |
+
# ๋ถ์ ๋ชจ๋ธ ์ ํ: "gemini", "groq", "qwen"
|
| 53 |
+
analysis_model: str = "qwen" # "gemini", "groq", "qwen" ์ค ์ ํ
|
| 54 |
|
| 55 |
class Config:
|
| 56 |
env_file = ".env"
|
app/services/qwen.py
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Qwen-VL ๋ฉํฐ๋ชจ๋ฌ ๋ถ์ ์๋น์ค."""
|
| 2 |
+
|
| 3 |
+
import httpx
|
| 4 |
+
import base64
|
| 5 |
+
import json
|
| 6 |
+
from app.core.config import settings
|
| 7 |
+
|
| 8 |
+
async def analyze_with_qwen(screenshot_path: str) -> dict:
|
| 9 |
+
"""
|
| 10 |
+
Analyze website screenshot using Qwen-VL API.
|
| 11 |
+
Args:
|
| 12 |
+
screenshot_path: Path to the screenshot file
|
| 13 |
+
Returns:
|
| 14 |
+
dict: Analysis results with title, atmosphere, colors, and keywords
|
| 15 |
+
Raises:
|
| 16 |
+
Exception: If analysis fails
|
| 17 |
+
"""
|
| 18 |
+
# ์ด๋ฏธ์ง ํ์ผ์ base64๋ก ์ธ์ฝ๋ฉ
|
| 19 |
+
with open(screenshot_path, "rb") as img_file:
|
| 20 |
+
img_b64 = base64.b64encode(img_file.read()).decode("utf-8")
|
| 21 |
+
|
| 22 |
+
prompt = """
|
| 23 |
+
You are a Senior Creative Director analyzing a website screenshot for a commercial poster design.
|
| 24 |
+
|
| 25 |
+
=== TASK ===
|
| 26 |
+
Extract key information to create a poster that VISUALLY REPRESENTS what this company does.
|
| 27 |
+
|
| 28 |
+
=== ANALYSIS STEPS ===
|
| 29 |
+
|
| 30 |
+
1. **WHAT IS THIS?** (Read the screen carefully)
|
| 31 |
+
- Company/Brand name (if Korean, romanize: ๋ฌด์ ์ฌโMUSINSA)
|
| 32 |
+
- What do they sell or provide? Be SPECIFIC.
|
| 33 |
+
- Who is the target user?
|
| 34 |
+
|
| 35 |
+
2. **VISUAL TRANSLATION** (Convert business to imagery)
|
| 36 |
+
The poster must show OBJECTS that represent the business:
|
| 37 |
+
| Business Type | What to Show |
|
| 38 |
+
|--------------|--------------|
|
| 39 |
+
| Productivity Tool | Organized workspace, floating UI panels, clean desk, glass screens with icons |
|
| 40 |
+
| Fashion Store | Clothes on racks, sneakers, fashion photography studio |
|
| 41 |
+
| Search/Tech | Holographic interfaces, data streams, futuristic screens |
|
| 42 |
+
| Delivery | Flying boxes, warehouse, conveyor belts |
|
| 43 |
+
| Food | The food items, kitchen, restaurant interior |
|
| 44 |
+
|
| 45 |
+
3. **COLOR EXTRACTION**
|
| 46 |
+
- What is the main brand color from the logo/design?
|
| 47 |
+
- Is it single color or multi-color brand?
|
| 48 |
+
|
| 49 |
+
=== OUTPUT (JSON) ===
|
| 50 |
+
{
|
| 51 |
+
"brand_name": "ENGLISH brand name",
|
| 52 |
+
"business_type": "Productivity/Fashion/Tech/Delivery/Food/Other",
|
| 53 |
+
"what_they_provide": "Specific description in 15 words",
|
| 54 |
+
"poster_objects": "List concrete objects: 'glass panels, folder icon, chat icon, checklist, modern desk, soft lighting'",
|
| 55 |
+
"background_style": "Clean gradient/Studio/Futuristic/Warehouse/Minimal",
|
| 56 |
+
"primary_color": "#hexcode",
|
| 57 |
+
"mood": "Clean/Premium/Energetic/Calm"
|
| 58 |
+
}
|
| 59 |
+
"""
|
| 60 |
+
|
| 61 |
+
url = "https://api-inference.huggingface.co/models/Qwen/Qwen-VL-Chat" # ๋๋ ์ง์ ๋์ด ์๋ฒ ์ฃผ์
|
| 62 |
+
headers = {
|
| 63 |
+
"Authorization": f"Bearer {settings.hf_token}",
|
| 64 |
+
"Content-Type": "application/json"
|
| 65 |
+
}
|
| 66 |
+
payload = {
|
| 67 |
+
"inputs": {
|
| 68 |
+
"image": f"data:image/png;base64,{img_b64}",
|
| 69 |
+
"question": prompt
|
| 70 |
+
}
|
| 71 |
+
}
|
| 72 |
+
async with httpx.AsyncClient(timeout=60.0) as client:
|
| 73 |
+
response = await client.post(url, headers=headers, json=payload)
|
| 74 |
+
response.raise_for_status()
|
| 75 |
+
result = response.json()
|
| 76 |
+
# Qwen-VL์ ๋ต๋ณ์ด result['answer']์ ๋ค์ด์์
|
| 77 |
+
text = result.get("answer", "").strip()
|
| 78 |
+
|
| 79 |
+
# ๊ธฐ์กด JSON ํ์ฑ ๋ก์ง ์ฌ์ฌ์ฉ
|
| 80 |
+
if "```json" in text:
|
| 81 |
+
text = text.split("```json")[1].split("```", 1)[0].strip()
|
| 82 |
+
elif "```" in text:
|
| 83 |
+
text = text.split("```", 1)[1].split("```", 1)[0].strip()
|
| 84 |
+
text = text[text.find('{'):text.rfind('}')+1]
|
| 85 |
+
try:
|
| 86 |
+
analysis = json.loads(text)
|
| 87 |
+
return analysis
|
| 88 |
+
except json.JSONDecodeError as e:
|
| 89 |
+
text = text.replace("'", '"').replace('\n', ' ')
|
| 90 |
+
try:
|
| 91 |
+
analysis = json.loads(text)
|
| 92 |
+
return analysis
|
| 93 |
+
except:
|
| 94 |
+
raise Exception(f"Failed to parse Qwen response as JSON: {text[:200]}")
|