gistfile1.py

def compare_batch(self, proxy_image_b64: str, candidate_images_b64: List[str]) -> List[Optional[int]]:
    """
    Compare proxy to multiple candidates in one API call.
    Returns list of scores corresponding to candidate_images_b64.
    """
    num_candidates = len(candidate_images_b64)

    prompt = f"""You are comparing Magic: The Gathering card artwork. I will show you:
                1. First image: The PROXY card (reference)
                2. Next {num_candidates} images: CANDIDATE printings (numbered 1-{num_candidates})

                Your task: Rate how similar each CANDIDATE's artwork is to the PROXY artwork.

                Focus ONLY on the central artwork/illustration. Completely ignore borders, frames, card text, and other non-artwork elements.

                Scoring guide:
                - 100 = Identical artwork (same illustration, same artist, same composition)
                - 90-99 = Same artwork with very minor variations (slight color/cropping differences)
                - 70-89 = Similar artwork but noticeable differences (extended art, alternate version)
                - 40-69 = Different artwork but similar subject/theme
                - 0-39 = Completely different artwork

                Respond with ONLY the scores as a comma-separated list. Example: 100, 95, 30, 100, 20"""

    # Build content list: prompt + proxy + all candidates
    content = [{"type": "text", "text": prompt}]

    # Add proxy image
    content.append({
        "type": "image_url",
        "image_url": {"url": f"data:image/jpeg;base64,{proxy_image_b64}"}
    })

    # Add all candidate images
    for candidate_b64 in candidate_images_b64:
        content.append({
            "type": "image_url",
            "image_url": {"url": f"data:image/jpeg;base64,{candidate_b64}"}
        })

    payload = {
        "model": "nm-testing/Qwen3-VL-8B-Instruct-NVFP4",
        "messages": [{"role": "user", "content": content}],
        "max_tokens": 100,  # Need more tokens for multiple scores
        "temperature": 0.7,
        "top_p": 0.8,
        "top_k": 20,
        "repetition_penalty": 1.0,
        "presence_penalty": 1.5
    }

    try:
        response = requests.post(self.chat_endpoint, json=payload, timeout=60)

        if response.status_code != 200:
            print(f"    ERROR: API returned {response.status_code}")
            return [None] * num_candidates

        result = response.json()
        content = result['choices'][0]['message']['content'].strip()

        print(f"    Model response: {content[:200]}...")  # Debug output

        # Parse comma-separated scores
        numbers = re.findall(r'\d+', content)
        if len(numbers) < num_candidates:
            print(f"    WARNING: Expected {num_candidates} scores, got {len(numbers)}")

        scores = [int(n) for n in numbers[:num_candidates]]  # Take first N numbers
        scores = [min(100, max(0, s)) for s in scores]  # Clamp to 0-100

        # Pad with None if we got fewer scores than expected
        while len(scores) < num_candidates:
            scores.append(None)

        return scores[:num_candidates]

    except Exception as e:
        print(f"    ERROR: {e}")
        return [None] * num_candidates