시각적 질문 답변

시각적 질문 응답(VQA)

기본 VQA 프롬프트 구조

import anthropic, base64

client = anthropic.Anthropic(api_key='YOUR_API_KEY')

def ask_about_image(image_path, question, answer_format='Direct answer. No extra explanation.'):
    with open(image_path, 'rb') as f:
        img_b64 = base64.standard_b64encode(f.read()).decode('utf-8')

    prompt = f'{question}\n\n{answer_format}'

    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=150,
        messages=[{'role': 'user', 'content': [
            {'type': 'image', 'source': {'type': 'base64', 'media_type': 'image/jpeg', 'data': img_b64}},
            {'type': 'text', 'text': prompt}
        ]}]
    )
    return r.content[0].text

# Example VQA calls (replace image.jpg with actual image)
print('VQA function defined. Ready for image questions.')

개수 세기 질문

# Vague (bad):
vague_prompt = 'How many people are there?'

# Precise (good): specifies what counts and what does not
counting_prompt = '''
How many people are visible in this image?
Count only: people whose faces OR bodies are at least 50% visible.
Do NOT count: people who are heavily cropped, cut off at the edge, or only partially visible.
Return a single number.
'''

# Even more precise: handles partial visibility explicitly
precise_count = '''
Count the number of distinct individuals visible in this image.
If a person is partially obscured, count them if more than half their body is visible.
Return JSON: {"count": integer, "partially_visible": integer, "notes": "string or null"}
'''

print('Counting prompts: vague vs precise.')
print('Precise prompts define edge cases explicitly.')

브랜드와 로고 식별

logo_prompt = '''
Identify all visible brand logos, company names, and product labels in this image.

For each, note:
- Brand/company name
- Where it appears in the image (top-left, center, on a product, etc.)
- Confidence: high (clearly legible) | medium (partially visible) | low (partially obscured)

Return JSON: {"brands": [{"name": str, "location": str, "confidence": str}]}
If no logos are visible, return: {"brands": []}
'''

import anthropic, base64, json
client = anthropic.Anthropic(api_key='YOUR_API_KEY')

def identify_brands(image_path):
    with open(image_path, 'rb') as f:
        img_b64 = base64.standard_b64encode(f.read()).decode('utf-8')
    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=200,
        messages=[{'role': 'user', 'content': [
            {'type': 'image', 'source': {'type': 'base64', 'media_type': 'image/jpeg', 'data': img_b64}},
            {'type': 'text', 'text': logo_prompt}
        ]}]
    )
    return json.loads(r.content[0].text)

print('Brand identification function defined.')

감정 및 표정 인식

emotion_prompt = '''
Describe the emotional expression of the person in this image.

Assess:
- Primary emotion: (happy, sad, angry, surprised, fearful, disgusted, neutral, or other)
- Intensity: (low, moderate, high)
- Confidence: (high if expression is clear, medium if subtle, low if face is obscured or turned away)
- Evidence: which specific facial features support your assessment

Return JSON:
{
  "primary_emotion": str,
  "intensity": str,
  "confidence": str,
  "evidence": str,
  "secondary_emotion": str or null
}

If no person or face is clearly visible, return: {"primary_emotion": null, "confidence": "none", "reason": str}
'''

print(emotion_prompt)

공간 관계 질문

spatial_prompt = '''
Answer questions about the spatial relationships of objects in this image.
Use these spatial terms consistently:
- Position in frame: top-left, top-center, top-right, middle-left, center, middle-right, bottom-left, bottom-center, bottom-right
- Relative position: in front of, behind, to the left of, to the right of, above, below, overlapping
- Distance: in the foreground, in the midground, in the background

Question: {question}

Answer in one or two sentences using the spatial vocabulary above.
'''

# Example questions:
questions = [
    'Where is the red cup relative to the laptop?',
    'Is the plant in the foreground or background?',
    'What object is to the left of the person?'
]

for q in questions:
    print(spatial_prompt.replace('{question}', q)[:200])
    print('---')

품질 및 상태 평가

condition_prompt = '''
Assess the condition of the main subject in this image.

Rate on these dimensions (1-5 scale, 5=excellent):
- Physical condition: (1=heavily damaged, 5=like new)
- Cleanliness: (1=very dirty, 5=spotless)
- Completeness: (1=major parts missing, 5=fully intact)

For each rating, provide one-sentence evidence.

Return JSON:
{
  "physical_condition": {"score": int, "evidence": str},
  "cleanliness": {"score": int, "evidence": str},
  "completeness": {"score": int, "evidence": str},
  "overall_grade": "excellent|good|fair|poor",
  "recommendation": str
}
'''

print('Condition assessment prompt defined.')
print('Useful for: product inspection, real estate, equipment maintenance.')

예/아니요 VQA 질문

def yes_no_question(image_path, question):
    with open(image_path, 'rb') as f:
        img_b64 = base64.standard_b64encode(f.read()).decode('utf-8')

    prompt = f'''
Answer this yes/no question about the image.
Return JSON: {{"answer": "yes|no", "confidence": "high|medium|low", "reason": str}}
Do NOT answer with maybe, possibly, or a hedged statement.
If you genuinely cannot determine the answer, return {{"answer": "unclear", "confidence": "low", "reason": str}}

Question: {question}
'''

    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=100,
        messages=[{'role': 'user', 'content': [
            {'type': 'image', 'source': {'type': 'base64', 'media_type': 'image/jpeg', 'data': img_b64}},
            {'type': 'text', 'text': prompt}
        ]}]
    )
    return json.loads(r.content[0].text)

# Example: 'Is there a safety helmet visible in the image?'
print('Yes/no VQA function defined.')

VQA 질문 연결

multi_question_prompt = '''
Answer all of the following questions about this image.
Return a JSON object where each key is the question ID.

Questions:
1. How many people are visible?
2. What is the approximate age range of the youngest person?
3. Is there any food visible in the image?
4. What is the dominant color in the image?
5. Is the setting indoors or outdoors?

Return JSON:
{
  "q1": {"answer": str},
  "q2": {"answer": str},
  "q3": {"answer": "yes|no", "details": str or null},
  "q4": {"answer": str},
  "q5": {"answer": "indoors|outdoors|unclear"}
}
'''

print('Multi-question VQA prompt — answers 5 questions in one API call.')

VQA 불확실성 처리

uncertainty_vqa_prompt = '''
Answer this question about the image as precisely as possible.
If the answer is not clearly visible or is ambiguous, say so explicitly.

Question: {question}

Return JSON:
{
  "answer": str,
  "confidence": "high|medium|low|cannot_determine",
  "limitation": str or null
}

For confidence levels:
- high: Answer is clearly visible and unambiguous
- medium: Visible but some uncertainty
- low: Partially visible or requires inference
- cannot_determine: Not enough visual information

Question: What brand is printed on the water bottle?
'''

print(uncertainty_vqa_prompt)

도메인별 VQA 프롬프트

# Manufacturing quality control VQA
qc_prompt = '''
Inspect this product image for quality defects.
Answer each question:
1. Are there any visible scratches or surface damage? (yes/no + location)
2. Is the product alignment within expected tolerance? (yes/no)
3. Are all required labels/markings present? (yes/no + list missing ones)
4. Overall QC result: PASS or FAIL?

Return JSON:
{"scratches": {"present": bool, "location": str or null},
 "alignment_ok": bool,
 "labels_complete": bool, "missing_labels": [str],
 "qc_result": "PASS|FAIL",
 "fail_reasons": [str]}
'''

# Food safety VQA  
food_prompt = '''
Inspect this food preparation image.
1. Are gloves being worn? 2. Is hair covered? 3. Any visible contamination risk?
Return JSON: {"gloves": bool, "hair_covered": bool, "contamination_risk": bool, "details": str}
'''

print("Domain-specific QC and food safety VQA prompts defined.")

간단 확인

VQA 프롬프트 — 핵심 요점