src/methods/captionsmiths.py

"""M1 — CaptionSmiths: Fine-tuned LLaVA-1.5 with structural style conditioning.

Uses published numbers only (CAPTIONSMITHS_USE_PUBLISHED=true).
Valid comparison for Track D only.
"""

from typing import Optional

import numpy as np

from src.methods.base import SteeringMethod


class CaptionSmiths(SteeringMethod):
    """CaptionSmiths — Fine-tuned baseline (ICCV 2025, Track D only)."""

    def __init__(self, checkpoint_path: Optional[str] = None, use_published: bool = True, **kwargs):
        self.checkpoint_path = checkpoint_path
        self.use_published = use_published

    @property
    def name(self) -> str:
        return "CaptionSmiths"

    @property
    def method_id(self) -> str:
        return "M1"

    @property
    def is_training_free(self) -> bool:
        return False

    def extract_vector(
        self,
        h_pos: np.ndarray,
        h_neg: np.ndarray,
        **kwargs,
    ) -> Optional[np.ndarray]:
        """CaptionSmiths does not use activation steering — returns None."""
        return None

    def generate_caption(self, model, processor, image, style: str, **kwargs) -> str:
        """Generate a caption using the fine-tuned CaptionSmiths model.

        If use_published=True, returns a placeholder indicating published
        numbers should be used instead.
        """
        if self.use_published:
            return f"[PUBLISHED] CaptionSmiths published result for style={style}"

        if self.checkpoint_path is None:
            raise NotImplementedError(
                "CaptionSmiths requires a checkpoint. "
                "Set CAPTIONSMITHS_USE_PUBLISHED=true to use published numbers."
            )
        raise NotImplementedError("CaptionSmiths inference not yet implemented.")