novel-downloader 1.4.5__py3-none-any.whl → 2.0.0__py3-none-any.whl

novel_downloader/utils/fontocr/ocr_v2.py

@@ -1,752 +0,0 @@
-#!/usr/bin/env python3
-"""
-novel_downloader.utils.fontocr.ocr_v2
--------------------------------------
-
-This class provides utility methods for optical character recognition (OCR)
-and font mapping, primarily used for decrypting custom font encryption
-on web pages (e.g., the Qidian website).
-"""
-
-import json
-import logging
-import math
-import os
-from collections.abc import Generator
-from pathlib import Path
-from typing import Any, TypeVar
-
-import cv2
-import numpy as np
-import paddle
-from fontTools.ttLib import TTFont
-from paddle.inference import Config
-from paddle.inference import create_predictor as _create_predictor
-from PIL import Image, ImageDraw, ImageFont
-from PIL.Image import Transpose
-
-try:
-    # pip install cupy-cuda11x
-    import cupy as array_backend  # GPU acceleration
-except ImportError:
-    import numpy as array_backend  # CPU only
-
-from novel_downloader.utils.constants import (
-    REC_CHAR_MODEL_FILES,
-    REC_IMAGE_SHAPE_MAP,
-)
-from novel_downloader.utils.hash_store import img_hash_store
-
-from .model_loader import (
-    get_rec_char_vector_dir,
-    get_rec_chinese_char_model_dir,
-)
-
-T = TypeVar("T")
-logger = logging.getLogger(__name__)
-
-
-class CTCLabelDecode:
-    """
-    Convert between text-index and text-label for CTC-based models.
-
-    :param character_dict_path: Path to the file containing characters, one per line.
-    :param beg_str: Token representing the start of sequence.
-    :param end_str: Token representing the end of sequence.
-    """
-
-    __slots__ = ("idx_to_char", "char_to_idx", "blank_id", "beg_str", "end_str")
-
-    def __init__(
-        self,
-        character_dict_path: str | Path,
-        beg_str: str = "sos",
-        end_str: str = "eos",
-    ):
-        # Store special tokens
-        self.beg_str = beg_str
-        self.end_str = end_str
-
-        # Read and clean character list (skip empty lines)
-        path = Path(character_dict_path)
-        chars = [
-            line.strip()
-            for line in path.read_text(encoding="utf-8").splitlines()
-            if line.strip()
-        ]
-
-        # Reserve index 0 for the CTC blank token, then actual characters
-        self.idx_to_char: list[str] = ["blank"] + chars
-        self.blank_id: int = 0
-
-        # Build reverse mapping from character to index
-        self.char_to_idx = {ch: i for i, ch in enumerate(self.idx_to_char)}
-
-    def decode(
-        self,
-        text_indices: np.ndarray,
-        text_probs: np.ndarray | None = None,
-    ) -> list[tuple[str, float]]:
-        """
-        Decode index sequences to strings and compute average confidence.
-
-        :param text_indices: (batch_size, seq_len) class indices.
-        :param text_probs: Optional per-step probabilities, same shape.
-        :return: List of (string, avg_confidence) per sample.
-        """
-        results: list[tuple[str, float]] = []
-        batch_size = text_indices.shape[0]
-
-        for i in range(batch_size):
-            seq = text_indices[i]
-            # Collapse repeated tokens: keep first of any run
-            mask = np.concatenate(([True], seq[1:] != seq[:-1]))
-            # Remove blanks
-            mask &= seq != self.blank_id
-
-            # Map indices to characters
-            chars = [self.idx_to_char[idx] for idx in seq[mask]]
-
-            # Compute average confidence, or default to 1.0 if no probs provided
-            if text_probs is not None:
-                probs = text_probs[i][mask]
-                avg_conf = float(probs.mean()) if probs.size else 0.0
-            else:
-                avg_conf = 1.0
-
-            results.append(("".join(chars), avg_conf))
-
-        return results
-
-    def __call__(self, preds: Any) -> list[tuple[str, float]]:
-        """
-        Decode raw model outputs to final text labels and confidences.
-
-        :param preds: Model output array/tensor of shape (batch, seq_len, num_classes),
-                      or a tuple/list whose last element is that array.
-        :returns: A list of (decoded_string, average_confidence).
-        """
-        # If passed as (logits, ...), take the last element
-        if isinstance(preds, (tuple | list)):
-            preds = preds[-1]
-
-        # Convert framework tensor to numpy if needed
-        if hasattr(preds, "numpy"):
-            preds = preds.numpy()
-
-        # Get the most likely class index and its probability
-        text_idx = preds.argmax(axis=2)
-        text_prob = preds.max(axis=2)
-
-        return self.decode(text_idx, text_prob)
-
-
-class TextRecognizer:
-    def __init__(
-        self,
-        rec_model_dir: str,
-        rec_image_shape: str,
-        rec_batch_num: int,
-        rec_char_dict_path: str,
-        use_gpu: bool = False,
-        gpu_mem: int = 500,
-        gpu_id: int | None = None,
-    ):
-        self.rec_batch_num = int(rec_batch_num)
-        self.rec_image_shape = tuple(map(int, rec_image_shape.split(",")))  # (C, H, W)
-        self.postprocess_op = CTCLabelDecode(
-            character_dict_path=rec_char_dict_path,
-        )
-
-        self._create_predictor(
-            model_dir=rec_model_dir,
-            use_gpu=use_gpu,
-            gpu_mem=gpu_mem,
-            gpu_id=gpu_id,
-        )
-
-    def _get_infer_gpu_id(self) -> int:
-        """
-        Look at CUDA_VISIBLE_DEVICES or HIP_VISIBLE_DEVICES,
-        pick the first entry and return as integer. Fallback to 0.
-        """
-        if not paddle.device.is_compiled_with_rocm:
-            gpu_env = os.environ.get("CUDA_VISIBLE_DEVICES", "0")
-        else:
-            gpu_env = os.environ.get("HIP_VISIBLE_DEVICES", "0")
-
-        first = gpu_env.split(",")[0]
-        try:
-            return int(first)
-        except ValueError:
-            return 0
-
-    def _create_predictor(
-        self,
-        model_dir: str,
-        use_gpu: bool,
-        gpu_mem: int,
-        gpu_id: int | None = None,
-    ) -> None:
-        """
-        Internal helper to build the Paddle predictor + I/O handles
-        """
-        model_file = f"{model_dir}/inference.pdmodel"
-        params_file = f"{model_dir}/inference.pdiparams"
-
-        cfg = Config(model_file, params_file)
-        if use_gpu:
-            chosen = gpu_id if gpu_id is not None else self._get_infer_gpu_id()
-            cfg.enable_use_gpu(gpu_mem, chosen)
-        else:
-            cfg.disable_gpu()
-
-        # enable memory optim
-        cfg.enable_memory_optim()
-        cfg.disable_glog_info()
-        # Use zero-copy feed/fetch for speed
-        cfg.switch_use_feed_fetch_ops(False)
-        # Enable IR optimizations
-        cfg.switch_ir_optim(True)
-
-        self.config = cfg
-        self.predictor = _create_predictor(cfg)
-
-        in_name = self.predictor.get_input_names()[0]
-        self.input_tensor = self.predictor.get_input_handle(in_name)
-
-        out_names = self.predictor.get_output_names()
-        preferred = "softmax_0.tmp_0"
-        selected = [preferred] if preferred in out_names else out_names
-        self.output_tensors = [self.predictor.get_output_handle(n) for n in selected]
-
-    def __call__(self, img_list: list[np.ndarray]) -> list[tuple[str, float]]:
-        """
-        Perform batch OCR on a list of images and return (text, confidence) tuples.
-        """
-        img_num = len(img_list)
-        results: list[tuple[str, float]] = []
-
-        C, H, W0 = self.rec_image_shape
-
-        # Process images in batches
-        for start in range(0, img_num, self.rec_batch_num):
-            batch = img_list[start : start + self.rec_batch_num]
-            # Compute width-to-height ratios for all images in the batch
-            wh_ratios = [img.shape[1] / float(img.shape[0]) for img in batch]
-            max_wh = max(W0 / H, *wh_ratios)
-
-            B = len(batch)
-            # Pre-allocate a numpy array for the batch
-            batch_tensor = np.zeros(
-                (B, C, H, int(math.ceil(H * max_wh))), dtype=np.float32
-            )
-
-            # Normalize and pad each image into the batch tensor
-            for i, img in enumerate(batch):
-                norm = self.resize_norm_img(img, max_wh)
-                batch_tensor[i, :, :, : norm.shape[2]] = norm
-
-            # Run inference
-            self.input_tensor.copy_from_cpu(batch_tensor)
-            self.predictor.run()
-
-            # Retrieve and post-process outputs
-            outputs = [t.copy_to_cpu() for t in self.output_tensors]
-            preds = outputs[0] if len(outputs) == 1 else outputs
-
-            rec_batch = self.postprocess_op(preds)
-            results.extend(rec_batch)
-
-        return results
-
-    def resize_norm_img(self, img: np.ndarray, max_wh_ratio: float) -> np.ndarray:
-        C, H, W0 = self.rec_image_shape
-        if img.ndim == 2:
-            # Convert grayscale images to RGB
-            img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
-        assert (
-            img.ndim == 3 and img.shape[2] == C
-        ), f"Expect {C}-channel image, got {img.shape}"
-
-        h, w = img.shape[:2]
-        # Determine new width based on the height and max width-height ratio
-        new_w = min(int(math.ceil(H * (w / h))), int(H * max_wh_ratio))
-        resized = cv2.resize(img, (new_w, H)).astype("float32")
-        # Change to CHW format and scale to [0,1]
-        resized = resized.transpose(2, 0, 1) / 255.0
-        # Normalize to [-1, 1]
-        resized = (resized - 0.5) / 0.5
-
-        return resized
-
-
-class FontOCRV2:
-    """
-    Version 2 of the FontOCR utility.
-
-    :param use_freq: if True, weight scores by character frequency
-    :param cache_dir: base path to store font-map JSON data
-    :param threshold: minimum confidence threshold [0.0-1.0]
-    :param font_debug: if True, dump per-char debug images under cache_dir
-    """
-
-    # Default constants
-    CHAR_IMAGE_SIZE = 64
-    CHAR_FONT_SIZE = 52
-    _freq_weight = 0.05
-
-    # shared resources
-    _global_char_freq_db: dict[str, int] = {}
-    _global_ocr: TextRecognizer | None = None
-    _global_vec_db: np.ndarray | None = None
-    _global_vec_label: tuple[str, ...] = ()
-    _global_vec_shape: tuple[int, int] = (32, 32)
-
-    def __init__(
-        self,
-        cache_dir: str | Path,
-        use_freq: bool = False,
-        use_ocr: bool = True,
-        use_vec: bool = False,
-        batch_size: int = 32,
-        gpu_mem: int = 500,
-        gpu_id: int | None = None,
-        ocr_weight: float = 0.6,
-        vec_weight: float = 0.4,
-        ocr_version: str = "v1.0",
-        threshold: float = 0.0,
-        font_debug: bool = False,
-        **kwargs: Any,
-    ) -> None:
-        self.use_freq = use_freq
-        self.use_ocr = use_ocr
-        self.use_vec = use_vec
-        self.batch_size = batch_size
-        self.gpu_mem = gpu_mem
-        self.gpu_id = gpu_id
-        self.ocr_weight = ocr_weight
-        self.vec_weight = vec_weight
-        self.ocr_version = ocr_version
-        self.threshold = threshold
-        self.font_debug = font_debug
-        self._max_freq = 5
-
-        self._cache_dir = Path(cache_dir)
-        self._cache_dir.mkdir(parents=True, exist_ok=True)
-        self._fixed_map_dir = self._cache_dir / "fixed_font_map"
-        self._fixed_map_dir.mkdir(parents=True, exist_ok=True)
-
-        if font_debug:
-            self._debug_dir = self._cache_dir / "font_debug" / "badcase"
-            self._debug_dir.mkdir(parents=True, exist_ok=True)
-
-        # load shared OCR + frequency DB
-        if self.use_ocr:
-            self._load_ocr_model()
-        if self.use_freq:
-            self._load_char_freq_db()
-        if self.use_vec:
-            self._load_char_vec_db()
-
-    def _load_ocr_model(self) -> None:
-        """
-        Initialize the shared PaddleOCR model if not already loaded.
-        """
-        if FontOCRV2._global_ocr is not None:
-            return
-
-        gpu_available = paddle.device.is_compiled_with_cuda()
-        self._char_model_dir = get_rec_chinese_char_model_dir(self.ocr_version)
-
-        for fname in REC_CHAR_MODEL_FILES:
-            full_path = self._char_model_dir / fname
-            if not full_path.exists():
-                raise FileNotFoundError(f"[FontOCR] Required file missing: {full_path}")
-
-        char_dict_file = self._char_model_dir / "rec_custom_keys.txt"
-        FontOCRV2._global_ocr = TextRecognizer(
-            rec_model_dir=str(self._char_model_dir),
-            rec_char_dict_path=str(char_dict_file),
-            rec_image_shape=REC_IMAGE_SHAPE_MAP[self.ocr_version],
-            rec_batch_num=self.batch_size,
-            use_gpu=gpu_available,
-            gpu_mem=self.gpu_mem,
-            gpu_id=self.gpu_id,
-        )
-
-    def _load_char_freq_db(self) -> bool:
-        """
-        Loads character frequency data from a JSON file and
-        assigns it to the instance variable.
-
-        :return: True if successfully loaded, False otherwise.
-        """
-        if FontOCRV2._global_char_freq_db is not None:
-            return True
-
-        try:
-            char_freq_map_file = self._char_model_dir / "char_freq.json"
-            with char_freq_map_file.open("r", encoding="utf-8") as f:
-                FontOCRV2._global_char_freq_db = json.load(f)
-            self._max_freq = max(FontOCRV2._global_char_freq_db.values())
-            return True
-        except Exception as e:
-            logger.warning("[FontOCR] Failed to load char freq DB: %s", e)
-            return False
-
-    def _load_char_vec_db(self) -> None:
-        """
-        Initialize the shared Char Vector if not already loaded.
-        """
-        if FontOCRV2._global_vec_db is not None:
-            return
-
-        char_vec_dir = get_rec_char_vector_dir(self.ocr_version)
-        char_vec_npy_file = char_vec_dir / "char_vectors.npy"
-        char_vec_label_file = char_vec_dir / "char_vectors.txt"
-
-        # Load and normalize vector database
-        vec_db = array_backend.load(char_vec_npy_file)
-        _, dim = vec_db.shape
-        side = int(np.sqrt(dim))
-        FontOCRV2._global_vec_shape = (side, side)
-
-        norm = array_backend.linalg.norm(vec_db, axis=1, keepdims=True) + 1e-6
-        FontOCRV2._global_vec_db = vec_db / norm
-
-        # Load corresponding labels
-        with open(char_vec_label_file, encoding="utf-8") as f:
-            FontOCRV2._global_vec_label = tuple(line.strip() for line in f)
-
-    @staticmethod
-    def _generate_char_image(
-        char: str,
-        render_font: ImageFont.FreeTypeFont,
-        is_reflect: bool = False,
-    ) -> Image.Image | None:
-        """
-        Render a single character into a square image.
-        If is_reflect is True, flip horizontally.
-        """
-        size = FontOCRV2.CHAR_IMAGE_SIZE
-        img = Image.new("L", (size, size), color=255)
-        draw = ImageDraw.Draw(img)
-        bbox = draw.textbbox((0, 0), char, font=render_font)
-        w, h = bbox[2] - bbox[0], bbox[3] - bbox[1]
-        x = (size - w) // 2 - bbox[0]
-        y = (size - h) // 2 - bbox[1]
-        draw.text((x, y), char, fill=0, font=render_font)
-        if is_reflect:
-            img = img.transpose(Transpose.FLIP_LEFT_RIGHT)
-
-        img_np = np.array(img)
-        if np.unique(img_np).size == 1:
-            return None
-
-        return img
-
-    def match_text_by_embedding(
-        self,
-        images: Image.Image | list[Image.Image],
-        top_k: int = 1,
-    ) -> list[tuple[str, float]] | list[list[tuple[str, float]]]:
-        """
-        Match input image to precomputed character embeddings using cosine similarity.
-
-        :param images: a PIL.Image or a list of PIL.Image to match
-        :param top_k: int, how many top matches to return
-
-        :return:
-        - If a single Image was passed in,
-                returns a list of (label, score) tuples sorted descending.
-
-        - If a list of Images was passed in, returns a list of such lists.
-        """
-        if self._global_vec_db is None:
-            return []
-        try:
-            imgs: list[Image.Image] = (
-                [images] if isinstance(images, Image.Image) else images
-            )
-
-            # Convert images to normalized 1D vectors
-            vecs = []
-            for img in imgs:
-                pil_gray = img.convert("L").resize(self._global_vec_shape)
-                arr = np.asarray(pil_gray, dtype=np.float32) / 255.0
-                v = array_backend.asarray(arr).ravel()
-                v /= array_backend.linalg.norm(v) + 1e-6
-                vecs.append(v)
-
-            batch = array_backend.stack(vecs, axis=0)  # (N, D)
-            # Compute all cosine similarities in one batch:
-            sims_batch = batch.dot(self._global_vec_db.T)  # (N, num_chars)
-
-            all_results: list[list[tuple[str, float]]] = []
-            for sims in sims_batch:
-                k = min(top_k, sims.shape[0])
-                top_unsorted = array_backend.argpartition(-sims, k - 1)[:k]
-                top_idx = top_unsorted[array_backend.argsort(-sims[top_unsorted])]
-                results = [
-                    (self._global_vec_label[int(i)], float(sims[int(i)]))
-                    for i in top_idx
-                ]
-                all_results.append(results)
-
-            # Unwrap single-image case
-            return all_results[0] if isinstance(images, Image.Image) else all_results
-        except Exception as e:
-            logger.warning("[FontOCR] Error: %s", e)
-            default = [("", 0.0)]
-            if isinstance(images, Image.Image):
-                return default
-            else:
-                return [default for _ in range(len(images))]
-
-    def run_ocr_on_images(
-        self,
-        images: Image.Image | list[Image.Image],
-    ) -> tuple[str, float] | list[tuple[str, float]]:
-        """
-        Run OCR on one or more PIL.Image(s) and return recognized text with confidence
-
-        :param images: A single PIL.Image or list of PIL.Images to recognize.
-        :return:
-        - If a single image is passed, returns Tuple[str, float].
-
-        - If a list is passed, returns List[Tuple[str, float]].
-        """
-        if self._global_ocr is None:
-            return []
-        try:
-            # Normalize input to a list of numpy arrays (RGB)
-            img_list = [images] if isinstance(images, Image.Image) else images
-            np_imgs: list[np.ndarray] = [
-                np.array(img.convert("RGB")) for img in img_list
-            ]
-
-            # Run OCR
-            ocr_results = self._global_ocr(np_imgs)
-
-            # Return result depending on input type
-            return ocr_results if isinstance(images, list) else ocr_results[0]
-
-        except Exception as e:
-            logger.warning("[FontOCR] OCR failed: %s", e)
-            fallback = ("", 0.0)
-            return (
-                fallback
-                if isinstance(images, Image.Image)
-                else [fallback for _ in images]
-            )
-
-    def query(
-        self,
-        images: Image.Image | list[Image.Image],
-        top_k: int = 3,
-    ) -> list[tuple[str, float]] | list[list[tuple[str, float]]]:
-        """
-        For each input image, run OCR + embedding match, fuse scores,
-        and return a sorted list of (char, score) above self.threshold.
-        """
-        # normalize to list
-        single = isinstance(images, Image.Image)
-        imgs: list[Image.Image] = [images] if single else images
-
-        # try the hash store
-        hash_batch = [img_hash_store.query(img, k=top_k) or [] for img in imgs]
-
-        fallback_indices = [i for i, h in enumerate(hash_batch) if not h]
-        fallback_imgs = [imgs[i] for i in fallback_indices]
-
-        # OCR scores
-        raw_ocr: tuple[str, float] | list[tuple[str, float]] = (
-            self.run_ocr_on_images(fallback_imgs)
-            if (self.use_ocr and fallback_imgs)
-            else []
-        )
-        if isinstance(raw_ocr, tuple):
-            ocr_fallback: list[tuple[str, float]] = [raw_ocr]
-        else:
-            ocr_fallback = raw_ocr
-
-        # Vec-embedding scores
-        raw_vec: list[tuple[str, float]] | list[list[tuple[str, float]]] = (
-            self.match_text_by_embedding(fallback_imgs, top_k=top_k)
-            if (self.use_vec and fallback_imgs)
-            else []
-        )
-        if raw_vec and isinstance(raw_vec[0], tuple):
-            vec_fallback: list[list[tuple[str, float]]] = [raw_vec]  # type: ignore
-        else:
-            vec_fallback = raw_vec  # type: ignore
-
-        # Fuse OCR+vector for the fallback set
-        fused_fallback: list[list[tuple[str, float]]] = []
-        for ocr_preds, vec_preds in zip(ocr_fallback, vec_fallback, strict=False):
-            scores: dict[str, float] = {}
-
-            # OCR weight
-            if ocr_preds:
-                ch, s = ocr_preds
-                scores[ch] = scores.get(ch, 0.0) + self.ocr_weight * s
-                logger.debug(
-                    "[FontOCR] OCR with weight: scores[%s] = %s", ch, scores[ch]
-                )
-            # Vec weight
-            for ch, s in vec_preds:
-                scores[ch] = scores.get(ch, 0.0) + self.vec_weight * s
-                logger.debug(
-                    "[FontOCR] Vec with weight: scores[%s] = %s", ch, scores[ch]
-                )
-            # Optional frequency
-            if self.use_freq:
-                for ch in list(scores):
-                    level = self._global_char_freq_db.get(ch, self._max_freq)
-                    freq_score = (self._max_freq - level) / max(1, self._max_freq)
-                    scores[ch] += self._freq_weight * freq_score
-                    logger.debug(
-                        "[FontOCR] After Freq weight: scores[%s] = %s", ch, scores[ch]
-                    )
-
-            # Threshold + sort + top_k
-            filtered = [(ch, sc) for ch, sc in scores.items() if sc >= self.threshold]
-            filtered.sort(key=lambda x: -x[1])
-
-            fused_fallback.append(filtered[:top_k])
-
-        # Recombine hash hits + fallback in original order
-        fused_batch: list[list[tuple[str, float]]] = []
-        fallback_iter = iter(fused_fallback)
-        for h_preds in hash_batch:
-            if h_preds:
-                fused_batch.append(h_preds)
-            else:
-                fused_batch.append(next(fallback_iter))
-
-        # Unwrap single-image case
-        return fused_batch[0] if single else fused_batch
-
-    def _chunked(self, seq: list[T], size: int) -> Generator[list[T], None, None]:
-        """Yield successive chunks of `seq` of length `size`."""
-        for i in range(0, len(seq), size):
-            yield seq[i : i + size]
-
-    def generate_font_map(
-        self,
-        fixed_font_path: str | Path,
-        random_font_path: str | Path,
-        char_set: set[str],
-        refl_set: set[str],
-        chapter_id: str | None = None,
-    ) -> dict[str, str]:
-        """
-        Generates a mapping from encrypted (randomized) font characters to
-        their real recognized characters by rendering and OCR-based matching.
-
-        :param fixed_font_path: Path to the reference (fixed) font.
-        :param random_font_path: Path to the obfuscated (random) font.
-        :param char_set: Characters to process normally.
-        :param refl_set: Characters to process as horizontally flipped.
-        :param chapter_id: Chapter ID
-
-        :returns mapping_result: { obf_char: real_char, ... }
-        """
-        mapping_result: dict[str, str] = {}
-        fixed_map_file = self._fixed_map_dir / f"{Path(fixed_font_path).stem}.json"
-
-        # load existing cache
-        try:
-            with open(fixed_map_file, encoding="utf-8") as f:
-                fixed_map = json.load(f)
-        except Exception:
-            fixed_map = {}
-
-        # prepare font renderers and cmap sets
-        try:
-            fixed_ttf = TTFont(fixed_font_path)
-            fixed_chars = {chr(c) for c in fixed_ttf.getBestCmap()}
-            fixed_font = ImageFont.truetype(str(fixed_font_path), self.CHAR_FONT_SIZE)
-
-            random_ttf = TTFont(random_font_path)
-            random_chars = {chr(c) for c in random_ttf.getBestCmap()}
-            random_font = ImageFont.truetype(str(random_font_path), self.CHAR_FONT_SIZE)
-        except Exception as e:
-            logger.error("[FontOCR] Failed to load TTF fonts: %s", e)
-            return mapping_result
-
-        def _render_batch(
-            chars: list[tuple[str, bool]]
-        ) -> list[tuple[str, Image.Image]]:
-            out = []
-            for ch, reflect in chars:
-                if ch in fixed_chars:
-                    font = fixed_font
-                elif ch in random_chars:
-                    font = random_font
-                else:
-                    continue
-                img = self._generate_char_image(ch, font, reflect)
-                if img is not None:
-                    out.append((ch, img))
-            return out
-
-        # process normal and reflected sets together
-        debug_idx = 1
-        for chars, reflect in [(list(char_set), False), (list(refl_set), True)]:
-            for batch_chars in self._chunked(chars, self.batch_size):
-                # render all images in this batch
-                to_render = [(ch, reflect) for ch in batch_chars]
-                rendered = _render_batch(to_render)
-                if not rendered:
-                    continue
-
-                # query OCR+vec simultaneously
-                imgs_to_query = [img for (ch, img) in rendered]
-                fused_raw = self.query(imgs_to_query, top_k=3)
-                if isinstance(fused_raw[0], tuple):
-                    fused: list[list[tuple[str, float]]] = [fused_raw]  # type: ignore
-                else:
-                    fused = fused_raw  # type: ignore
-
-                # pick best per char, apply threshold + cache
-                for (ch, img), preds in zip(rendered, fused, strict=False):
-                    if ch in fixed_map:
-                        mapping_result[ch] = fixed_map[ch]
-                        logger.debug(
-                            "[FontOCR] Using cached mapping: '%s' -> '%s'",
-                            ch,
-                            fixed_map[ch],
-                        )
-                        continue
-                    if not preds:
-                        if self.font_debug and chapter_id:
-                            dbg_path = (
-                                self._debug_dir / f"{chapter_id}_{debug_idx:04d}.png"
-                            )
-                            img.save(dbg_path)
-                            logger.debug(
-                                "[FontOCR] Saved debug image for '%s': %s", ch, dbg_path
-                            )
-                            debug_idx += 1
-                        continue
-                    real_char, _ = preds[0]
-                    mapping_result[ch] = real_char
-                    fixed_map[ch] = real_char
-                    if self.font_debug:
-                        logger.debug(
-                            "[FontOCR] Prediction for char '%s': top_pred='%s'",
-                            ch,
-                            real_char,
-                        )
-                        logger.debug("[FontOCR] All predictions: %s", preds)
-
-        # persist updated fixed_map
-        try:
-            with open(fixed_map_file, "w", encoding="utf-8") as f:
-                json.dump(fixed_map, f, ensure_ascii=False, indent=2)
-        except Exception as e:
-            logger.error("[FontOCR] Failed to save fixed map: %s", e)
-
-        return mapping_result