Top Vision LLMs Compared: Qwen 2.5-VL vs LLaMA 3.2

# Download Llama 3.2 Vision model
ollama pull llama3.2-vision

# Download Qwen 2.5-VL model
ollama pull qwen2.5vl

import ollama

# Define the consistent prompts to be used for both models
prompts = {
    "document_understanding": "Extract all key-value pairs from this invoice image. List them in a table.",
    "image_captioning": "Generate a detailed caption for this image, including all visible objects and their spatial relationships.",
    "visual_qa": "How many people are in this image, and what are they doing?",
    "multilingual_tasks": "Extract all text in Chinese and Japanese from this image and translate it into English.",
    "medical_imaging": "Analyze this X-ray. Describe all visible anatomical structures and note any abnormalities.",
    "retail_analytics": "Extract all brand names and their corresponding product types from this shelf image.",
    "autonomous_systems": "Describe the spatial relationships between vehicles and traffic signs in this image.",
    "industrial_qc": "Detect any visible defects on this manufactured part."
}

# --- Inference Functions ---

def analyze_with_llama(prompt, image_path):
    """Sends a prompt and an image to the Llama 3.2 Vision model."""
    print(f"--- Llama 3.2 Vision Analyzing: {image_path} ---")
    response = ollama.chat(
        model='llama3.2-vision',
        messages=[{
            'role': 'user',
            'content': prompt,
            'images': [image_path]
        }]
    )
    print("Llama 3.2 Vision Response:\n", response['message']['content'])
    return response['message']['content']

def analyze_with_qwen(prompt, image_path):
    """Sends a prompt and an image to the Qwen 2.5-VL model."""
    print(f"--- Qwen 2.5-VL Analyzing: {image_path} ---")
    response = ollama.chat(
        model='qwen2.5vl',
        messages=[{
            'role': 'user',
            'content': prompt,
            'images': [image_path]
        }]
    )
    print("Qwen 2.5-VL Response:\n", response['message']['content'])
    return response['message']['content']

def compare_models(prompt_key, image_path):
    """Compare both models using the same prompt."""
    prompt = prompts[prompt_key]
    print(f"\n=== Testing: {prompt_key.upper()} ===")
    print(f"Prompt: {prompt}")
    print(f"Image: {image_path}\n")
    
    llama_response = analyze_with_llama(prompt, image_path)
    print("\n" + "="*50 + "\n")
    qwen_response = analyze_with_qwen(prompt, image_path)
    
    return llama_response, qwen_response

# Test document understanding with identical prompt
llama_doc, qwen_doc = compare_models("document_understanding", "invoice.png")

"""
=== Testing: DOCUMENT_UNDERSTANDING ===
Prompt: Extract all key-value pairs from this invoice image. List them in a table.
Image: invoice.png

--- Llama 3.2 Vision Analyzing: invoice.png ---
Llama 3.2 Vision Response:
Here are the key-value pairs from the invoice:
| Key              | Value                 |
|------------------|----------------------|
| Invoice Number   | INV-00123            |
| Date             | 2025-10-26           |
| Vendor           | Office Supplies Inc. |
| Total Amount     | $150.75              |

==================================================

--- Qwen 2.5-VL Analyzing: invoice.png ---
Qwen 2.5-VL Response:
| Key              | Value                    |
|------------------|--------------------------|
| Invoice Number   | INV-00123               |
| Date             | 2025-10-26              |
| Vendor           | Office Supplies Inc.     |
| Customer         | ABC Corporation          |
| Total Amount     | $150.75                  |
| Payment Terms    | Net 30                   |
| Items            | A4 Paper (5x), Pens (3x) |
"""

# Test image captioning with identical prompt
llama_caption, qwen_caption = compare_models("image_captioning", "living_room.jpg")

"""
=== Testing: IMAGE_CAPTIONING ===
Prompt: Generate a detailed caption for this image, including all visible objects and their spatial relationships.
Image: living_room.jpg

--- Llama 3.2 Vision Analyzing: living_room.jpg ---
Llama 3.2 Vision Response:
A cozy living room with a gray sofa positioned in the center. A wooden coffee table sits in front of the sofa, and a large window to the left lets in natural light. The room has warm lighting and appears well-organized.

==================================================

--- Qwen 2.5-VL Analyzing: living_room.jpg ---
Qwen 2.5-VL Response:
This is an indoor scene of a living room. It contains one gray fabric sofa positioned centrally, one rectangular oak coffee table in front of the sofa, two blue cushions on the left side of the sofa, and a potted plant in the corner. The coffee table has a stack of three books on it. A floor lamp stands to the right of the sofa.
"""

# Test visual question answering with identical prompt
llama_vqa, qwen_vqa = compare_models("visual_qa", "people_playing.jpg")

"""
=== Testing: VISUAL_QA ===
Prompt: How many people are in this image, and what are they doing?
Image: people_playing.jpg

--- Llama 3.2 Vision Analyzing: people_playing.jpg ---
Llama 3.2 Vision Response:
There are three people in this image. They are gathered around a table and appear to be playing a board game together. The atmosphere looks casual and friendly.

==================================================

--- Qwen 2.5-VL Analyzing: people_playing.jpg ---
Qwen 2.5-VL Response:
There are 3 people in the image. Person 1 (left) is rolling dice, Person 2 (center) is moving a game piece on the board, and Person 3 (right) is observing the game while holding cards.
"""

# Test multilingual capabilities with identical prompt
llama_multi, qwen_multi = compare_models("multilingual_tasks", "signs.jpg")

"""
=== Testing: MULTILINGUAL_TASKS ===
Prompt: Extract all text in Chinese and Japanese from this image and translate it into English.
Image: signs.jpg

--- Llama 3.2 Vision Analyzing: signs.jpg ---
Llama 3.2 Vision Response:
I can see some text that appears to be in Asian languages in this image, but I cannot accurately distinguish between Chinese and Japanese characters or provide reliable translations. I would recommend using specialized OCR tools for accurate multilingual text extraction.

==================================================

--- Qwen 2.5-VL Analyzing: signs.jpg ---
Qwen 2.5-VL Response:
Chinese Text: 欢迎光临 (Huānyíng guānglín) - Translation: Welcome
Japanese Text: ありがとうございます (Arigatou gozaimasu) - Translation: Thank you very much
Korean Text: 안녕하세요 (Annyeonghaseyo) - Translation: Hello
"""

# Test medical imaging analysis with identical prompt
llama_medical, qwen_medical = compare_models("medical_imaging", "wrist_xray.png")

"""
=== Testing: MEDICAL_IMAGING ===
Prompt: Analyze this X-ray. Describe all visible anatomical structures and note any abnormalities.
Image: wrist_xray.png

--- Llama 3.2 Vision Analyzing: wrist_xray.png ---
Llama 3.2 Vision Response:
This is an anteroposterior view of a right wrist. The radius, ulna, carpal bones, and metacarpals appear intact with normal bone density. The joint spaces are preserved. There is a slight soft tissue swelling noted around the distal radius, but no clear fracture line is visible. The overall alignment appears normal.

==================================================

--- Qwen 2.5-VL Analyzing: wrist_xray.png ---
Qwen 2.5-VL Response:
Anatomical structures visible: radius, ulna, scaphoid, lunate, triquetrum, pisiform, trapezium, trapezoid, capitate, hamate, and metacarpals 1-5. 
Abnormalities detected: Possible hairline fracture at distal radius (coordinates: 145,230 to 152,245). Recommend clinical correlation and possible additional imaging.
"""

# Test retail analytics with identical prompt
llama_retail, qwen_retail = compare_models("retail_analytics", "shelf.jpg")

"""
=== Testing: RETAIL_ANALYTICS ===
Prompt: Extract all brand names and their corresponding product types from this shelf image.
Image: shelf.jpg

--- Llama 3.2 Vision Analyzing: shelf.jpg ---
Llama 3.2 Vision Response:
Visible brands and products:
- Coca-Cola: Soft drinks/sodas
- Pepsi: Soft drinks/sodas  
- Frito-Lay: Snack chips
- Oreo: Cookies/biscuits
- Pringles: Potato chips

==================================================

--- Qwen 2.5-VL Analyzing: shelf.jpg ---
Qwen 2.5-VL Response:
Brand-Product pairs identified:
- Coca-Cola: Carbonated soft drinks (12-pack cans, $7.99)
- Pepsi: Carbonated soft drinks (12-pack cans, $7.99)
- Frito-Lay: Potato chips (Family size, $4.29)
- Oreo: Sandwich cookies (Original, $3.49)
- Pringles: Stackable potato chips (Original flavor, $2.99)
"""