File size: 21,481 Bytes
844c9ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
"""Detect speech-bubble masks from an inpainted image + Lens text centers.

The Lens tree only tells us where the *text* sits (each item is a tight
rotated rectangle around a run of glyphs).  The actual speech *bubble*
extends well past those rectangles β€” that's the canvas the AI translation
should be rendered into, especially when the target language has the
opposite reading direction from the source (Japanese vertical β†’ Thai
horizontal, or English horizontal β†’ Chinese vertical).

This module recovers each bubble's polygon entirely from data we already
have, without bringing in any extra dependencies:

1. The pipeline already inpaints the source text (``erase_text_with_boxes``
   in :mod:`backend.render.erase`), leaving a clean speech-bubble interior
   on most pages.
2. We threshold the inpainted image to a binary "interior vs ink" mask.
3. ``cv2.connectedComponentsWithStats`` gives a label for every connected
   white region (bubble interiors, blank margins, panel gaps, …).
4. For each accepted label, ``cv2.findContours`` + ``cv2.fitEllipse``
   produce the bubble's ellipse fit, and the *largest axis-aligned
   rectangle inscribed in that ellipse* (``A/sqrt(2) Γ— B/sqrt(2)``) is
   what we return as the bubble bounds.  Unlike the raw component bbox
   β€” which includes the bubble's *corners*, often outside the visible
   outline β€” the inscribed rectangle is guaranteed to sit inside the
   bubble's outline, so text rendered into it never spills over the
   bubble edge.  This is the same approach scanlation typesetters use
   when picking text frames inside elliptical bubbles.
5. For each paragraph in the Lens tree, we seed-look-up the labels its
   item centers fall in, then return the inscribed rectangle of the
   matched label.

When a paragraph's seeds land on dark pixels (no clean interior β€” text
sat on a flat background, or the bubble has no outline) the function
returns ``None`` for that paragraph and the renderer falls back to the
existing text-only AABB.  Components that are larger than
``_MAX_BUBBLE_AREA_FRAC`` of the page are also rejected (they're almost
certainly the page background leaking through, not a real bubble).
"""

from __future__ import annotations

import math
import os
from typing import Any, Final

import cv2
import numpy as np
from PIL import Image

# Optional YOLO segmentation backend ---------------------------------------
#
# When ``ultralytics`` is installed AND the user opts in via the env vars
# below, the combined detector runs a YOLO segmentation pass first and uses
# its bubble polygons as the source of truth.  YOLO is much more accurate
# on hand-drawn bubbles than the threshold-based connected-components pass,
# but it adds 100-500ms on CPU, so we keep it strictly opt-in.
#
#   TP_BUBBLE_USE_YOLO=1           β€” enable the YOLO pass
#   TP_BUBBLE_YOLO_MODEL=path.pt   β€” path or HF id of a YOLO-seg checkpoint
#   TP_BUBBLE_YOLO_IMGSZ=640       β€” inference resolution (default 640)
#   TP_BUBBLE_YOLO_CONF=0.25       β€” detection confidence floor

try:
    from ultralytics import YOLO as _YOLO  # type: ignore
    _ULTRALYTICS_OK = True
except Exception:  # pragma: no cover - optional dependency
    _YOLO = None  # type: ignore[assignment]
    _ULTRALYTICS_OK = False

_yolo_model_cache: dict[str, Any] = {}

# Threshold above which a pixel counts as bubble interior.  Manga pages
# are typically inked with a hard black outline on white interiors, so a
# fixed threshold near pure white works without per-image tuning.
_BG_THRESHOLD: Final[int] = 220

# Components that swallow more than half the page are rejected β€” that's
# almost always the background "panel" component, not a real bubble.
_MAX_BUBBLE_AREA_FRAC: Final[float] = 0.50

# Skip tiny specks that survived inpainting (dust, JPEG noise).
_MIN_BUBBLE_AREA_PX: Final[int] = 64

# When a paragraph seed lands on a dark pixel (e.g. it's right on what
# used to be an ink stroke that the inpainter only partially cleaned),
# spiral outward this many pixels looking for a valid interior label.
_SEED_FALLBACK_RADIUS: Final[int] = 6

# 1 / sqrt(2) β€” the largest axis-aligned rectangle that fits inside an
# ellipse of diameters (A, B) has dimensions (A/sqrt(2), B/sqrt(2)).
# Cached as a constant because it shows up everywhere in the inscribed-
# rect maths and ``math.sqrt`` isn't free.
_INV_SQRT2: Final[float] = 1.0 / math.sqrt(2.0)


def _safe_float(x: Any, default: float = 0.0) -> float:
    """Coerce ``x`` to a finite float, falling back to ``default``."""
    try:
        n = float(x)
    except (TypeError, ValueError):
        return float(default)
    return n if math.isfinite(n) else float(default)


def _seed_label(
    labels: np.ndarray, seed_x: int, seed_y: int, radius: int = _SEED_FALLBACK_RADIUS
) -> int:
    """Return the connected-component label that owns ``(seed_x, seed_y)``.

    Label 0 is the binary-image background (= ink / outside the bubble);
    if the seed lands there we spiral outward up to ``radius`` pixels and
    return the first non-zero label we find.  Returns 0 if no interior
    label is reachable within the search radius (the renderer treats this
    as "no bubble" and falls back to the text AABB).
    """
    h, w = labels.shape
    seed_x = max(0, min(w - 1, int(seed_x)))
    seed_y = max(0, min(h - 1, int(seed_y)))

    lbl = int(labels[seed_y, seed_x])
    if lbl > 0:
        return lbl

    # Ring-by-ring outward search: ring r is the square perimeter at
    # Chebyshev distance r from the seed.
    for r in range(1, max(1, int(radius)) + 1):
        for dy in range(-r, r + 1):
            ny = seed_y + dy
            if ny < 0 or ny >= h:
                continue
            for dx in range(-r, r + 1):
                if abs(dx) != r and abs(dy) != r:
                    continue  # interior of the ring β€” already covered
                nx = seed_x + dx
                if nx < 0 or nx >= w:
                    continue
                lbl = int(labels[ny, nx])
                if lbl > 0:
                    return lbl
    return 0


def _inscribed_rect_from_contour(
    contour: np.ndarray, img_w: int, img_h: int
) -> tuple[float, float, float, float] | None:
    """Largest *axis-aligned* rectangle that fits inside ``contour``.

    Manga speech bubbles are dominantly elliptical, so we fit an ellipse
    to the contour (``cv2.fitEllipse``) and use the closed-form solution
    for the inscribed axis-aligned rectangle:

        width  = A / sqrt(2)
        height = B / sqrt(2)

    centered on the ellipse center, where ``(A, B)`` are the ellipse's
    *full* diameters along its major / minor axes.  For a rotated
    ellipse the formula is conservative β€” the returned rectangle still
    fits entirely inside the ellipse, just not always the absolute
    largest axis-aligned one possible.  That's fine for our use: we want
    text that never overflows the visible bubble outline, even if we
    leave a little unused room at the corners.

    Falls back to a simple bbox-scale (Γ—0.707) when the contour is too
    short for ``fitEllipse`` (it needs β‰₯ 5 points).

    Returns ``(left, top, width, height)`` in image pixels, clamped to
    the image, or ``None`` when the contour produces no usable rect.
    """
    if contour is None or len(contour) == 0:
        return None

    if len(contour) >= 5:
        try:
            (cx, cy), (a, b), _angle = cv2.fitEllipse(contour)
        except cv2.error:
            return None

        # Inscribed AABB inside the ellipse.  For a non-axis-aligned
        # ellipse this is still inside the ellipse; we're trading a few
        # percent of area for a closed-form solution.  When the major
        # axis is at ~45Β° the inscribed AABB shrinks toward the smaller
        # of the two β€” use the smaller axis on both sides to be safe.
        rect_w = float(a) * _INV_SQRT2
        rect_h = float(b) * _INV_SQRT2
        new_x = float(cx) - rect_w / 2.0
        new_y = float(cy) - rect_h / 2.0
    else:
        # fitEllipse needs β‰₯ 5 points β€” degenerate contour, just inset
        # the bbox by 1 - 1/sqrt(2) β‰ˆ 29.3% on each side.
        bx, by, bw, bh = cv2.boundingRect(contour)
        rect_w = float(bw) * _INV_SQRT2
        rect_h = float(bh) * _INV_SQRT2
        new_x = float(bx) + (float(bw) - rect_w) / 2.0
        new_y = float(by) + (float(bh) - rect_h) / 2.0

    # Clamp into the image and drop degenerate rectangles.
    new_x = max(0.0, min(float(img_w) - 1.0, new_x))
    new_y = max(0.0, min(float(img_h) - 1.0, new_y))
    if new_x + rect_w > img_w:
        rect_w = float(img_w) - new_x
    if new_y + rect_h > img_h:
        rect_h = float(img_h) - new_y
    if rect_w <= 1.0 or rect_h <= 1.0:
        return None
    return (new_x, new_y, rect_w, rect_h)


def _inscribed_rect_from_label(
    labels: np.ndarray,
    label_id: int,
    bbox: tuple[int, int, int, int],
    img_w: int,
    img_h: int,
) -> tuple[float, float, float, float] | None:
    """Inscribed rect for a single connected component.

    Extracts a tight sub-mask for the component (avoids running
    ``findContours`` over the whole page per label), grabs the outermost
    contour, and forwards it to :func:`_inscribed_rect_from_contour`.
    """
    bx, by, bw, bh = bbox
    if bw <= 0 or bh <= 0:
        return None
    # Crop to the component's bounding box, then build a binary submask
    # of just this label.  Contour points come back in local coordinates;
    # add the bbox offset to put them back into image space.
    submask = (labels[by : by + bh, bx : bx + bw] == label_id).astype(np.uint8)
    submask *= 255
    contours, _ = cv2.findContours(submask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    if not contours:
        return None
    contour = max(contours, key=cv2.contourArea)
    if contour.size == 0:
        return None
    contour = contour + np.array([[bx, by]])
    return _inscribed_rect_from_contour(contour, img_w, img_h)


def _item_seed_points(
    items: list[dict], img_w: int, img_h: int
) -> list[tuple[int, int]]:
    """Return one ``(x, y)`` seed in image pixels per text-bearing item.

    Uses ``box.center`` when available (the rotated rectangle's geometric
    centre, computed by :func:`backend.lens.tree.decode_tree`); falls
    back to ``left + width/2`` so older trees without ``center`` still
    seed correctly.
    """
    seeds: list[tuple[int, int]] = []
    for it in items or []:
        text = str(it.get("text") or "").strip()
        if not text:
            continue
        box = it.get("box") or {}
        center = box.get("center") or {}
        cx_n = _safe_float(
            center.get("x"),
            _safe_float(box.get("left")) + _safe_float(box.get("width")) / 2.0,
        )
        cy_n = _safe_float(
            center.get("y"),
            _safe_float(box.get("top")) + _safe_float(box.get("height")) / 2.0,
        )
        seeds.append((int(round(cx_n * img_w)), int(round(cy_n * img_h))))
    return seeds


def detect_bubble_bounds_by_paragraph(
    erased_image: Image.Image,
    paragraphs: list[dict],
    img_w: int,
    img_h: int,
) -> dict[int, tuple[float, float, float, float] | None]:
    """Compute a bubble bounding box (in image pixels) for every paragraph.

    Parameters
    ----------
    erased_image
        The inpainted image that the rest of the overlay sits on (i.e.
        ``result["imageDataUri"]`` decoded).  Connected components are
        labelled on this image so the bubble interior comes back as one
        big white region.
    paragraphs
        ``tree["paragraphs"]`` from any decoded Lens tree.  Item centers
        from each paragraph are used as seeds.
    img_w, img_h
        Image dimensions in pixels (must match ``erased_image``).

    Returns
    -------
    dict
        Mapping from ``para_index`` to ``(left, top, width, height)`` in
        pixels.  ``None`` is returned for paragraphs whose seeds didn't
        find a usable interior component (the renderer falls back to the
        text-only AABB for those).
    """
    if not paragraphs or img_w <= 0 or img_h <= 0:
        return {}

    try:
        gray = np.asarray(erased_image.convert("L"), dtype=np.uint8)
    except Exception:
        return {}

    # Make absolutely sure the array matches the declared dimensions β€”
    # PIL/Numpy occasionally disagrees on orientation when EXIF rotation
    # is in play.
    if gray.shape != (img_h, img_w):
        # Resize is cheap and avoids index-out-of-bounds when seeds are
        # computed from the declared dims.
        gray = cv2.resize(gray, (int(img_w), int(img_h)), interpolation=cv2.INTER_AREA)

    # Bright pixels are bubble interior; dark = ink, panel border, or art.
    _, binary = cv2.threshold(gray, _BG_THRESHOLD, 255, cv2.THRESH_BINARY)

    # 8-way connectivity so diagonally-touching pixels join the same
    # component (matters for jaggy hand-drawn bubble outlines).
    n_labels, labels, stats, _ = cv2.connectedComponentsWithStats(binary, connectivity=8)
    if n_labels <= 1:
        return {int(p.get("para_index", i)): None for i, p in enumerate(paragraphs)}

    page_area = max(1, int(img_w) * int(img_h))
    max_area = int(page_area * _MAX_BUBBLE_AREA_FRAC)

    # Pre-compute one inscribed rectangle per accepted label.  Doing it
    # here (once per label) instead of inside the paragraph loop avoids
    # re-running ``findContours`` for paragraphs that share a bubble.
    inscribed_by_label: dict[int, tuple[float, float, float, float] | None] = {}
    for lbl in range(1, n_labels):
        l, t, w, h, area = (int(v) for v in stats[lbl])
        if area > max_area or area < _MIN_BUBBLE_AREA_PX:
            continue
        inscribed_by_label[lbl] = _inscribed_rect_from_label(
            labels, lbl, (l, t, w, h), img_w, img_h
        )

    out: dict[int, tuple[float, float, float, float] | None] = {}
    for i, para in enumerate(paragraphs):
        pi = int(para.get("para_index", i))
        seeds = _item_seed_points(para.get("items") or [], img_w, img_h)
        if not seeds:
            out[pi] = None
            continue

        # Look up each seed's component label.  Labels not in the
        # accepted set (too big / too small) are dropped silently.
        seen_labels: list[int] = []
        for sx, sy in seeds:
            lbl = _seed_label(labels, sx, sy)
            if lbl > 0 and lbl in inscribed_by_label and inscribed_by_label[lbl] is not None:
                if lbl not in seen_labels:
                    seen_labels.append(lbl)

        if not seen_labels:
            out[pi] = None
            continue

        # One paragraph usually sits inside one bubble.  When seeds hit
        # more than one (multi-bubble paragraph β€” rare), pick the bubble
        # whose inscribed rect has the largest area; that's the one most
        # likely to be the actual speech bubble for the paragraph.
        best_lbl = max(
            seen_labels,
            key=lambda l: (
                (inscribed_by_label[l][2] * inscribed_by_label[l][3])  # type: ignore[index]
                if inscribed_by_label[l] is not None
                else 0.0
            ),
        )
        out[pi] = inscribed_by_label[best_lbl]

    return out


# --- YOLO segmentation backend (optional, opt-in) -------------------------

def _yolo_enabled() -> bool:
    """True when YOLO is both available *and* the user asked for it.

    The env-var gate is intentional: YOLO inference dwarfs OpenCV CC
    labelling in wall time, so we only run it when the user explicitly
    opts in.
    """
    if not _ULTRALYTICS_OK:
        return False
    raw = (os.environ.get("TP_BUBBLE_USE_YOLO", "") or "").strip().lower()
    return raw in ("1", "true", "yes", "on")


def _load_yolo_model(model_path: str) -> Any:
    """Cached YOLO loader.  Returns ``None`` on any failure so callers can
    fall back cleanly to the OpenCV path.
    """
    if not _ULTRALYTICS_OK or not model_path:
        return None
    cached = _yolo_model_cache.get(model_path)
    if cached is not None:
        return cached
    try:
        model = _YOLO(model_path)  # type: ignore[misc]
    except Exception:
        return None
    _yolo_model_cache[model_path] = model
    return model


def detect_bubble_bounds_by_paragraph_yolo(
    image: Image.Image,
    paragraphs: list[dict],
    img_w: int,
    img_h: int,
    model_path: str = "",
) -> dict[int, tuple[float, float, float, float] | None] | None:
    """Locate bubble bboxes with a YOLO-segmentation model.

    Returns the same shape as :func:`detect_bubble_bounds_by_paragraph`,
    or ``None`` when YOLO isn't available / the model can't load /
    inference fails β€” the orchestrator falls back to the OpenCV detector.

    Each paragraph's *centroid* (average of its item centers) is used to
    pick the smallest YOLO bbox that contains it.  Multiple paragraphs
    can land on the same bubble (multi-line speech) and they'll all map
    to that bubble's bbox, which is the desired behaviour.
    """
    if not paragraphs or img_w <= 0 or img_h <= 0:
        return None
    model_path = model_path or os.environ.get("TP_BUBBLE_YOLO_MODEL", "").strip()
    model = _load_yolo_model(model_path)
    if model is None:
        return None

    imgsz = int(os.environ.get("TP_BUBBLE_YOLO_IMGSZ", "640") or 640)
    conf = float(os.environ.get("TP_BUBBLE_YOLO_CONF", "0.25") or 0.25)

    try:
        np_img = np.asarray(image.convert("RGB"))
        results = model(np_img, imgsz=imgsz, conf=conf, verbose=False)
    except Exception:
        return None

    bubble_boxes: list[tuple[float, float, float, float]] = []
    try:
        for r in results:
            boxes = getattr(r, "boxes", None)
            if boxes is None:
                continue
            xyxy = getattr(boxes, "xyxy", None)
            if xyxy is None:
                continue
            arr = xyxy.cpu().numpy() if hasattr(xyxy, "cpu") else np.asarray(xyxy)
            for row in arr:
                x1, y1, x2, y2 = (float(v) for v in row[:4])
                if x2 > x1 and y2 > y1:
                    bubble_boxes.append((x1, y1, x2, y2))
    except Exception:
        return None

    if not bubble_boxes:
        return None

    out: dict[int, tuple[float, float, float, float] | None] = {}
    for i, para in enumerate(paragraphs):
        pi = int(para.get("para_index", i))
        seeds = _item_seed_points(para.get("items") or [], img_w, img_h)
        if not seeds:
            out[pi] = None
            continue

        cx = sum(s[0] for s in seeds) / len(seeds)
        cy = sum(s[1] for s in seeds) / len(seeds)

        # Pick the smallest bubble bbox containing the centroid β€” when
        # bubbles nest (thought bubbles inside frames), the inner wins.
        best: tuple[float, float, float, float] | None = None
        best_area = float("inf")
        for x1, y1, x2, y2 in bubble_boxes:
            if not (x1 <= cx <= x2 and y1 <= cy <= y2):
                continue
            area = (x2 - x1) * (y2 - y1)
            if area < best_area:
                best_area = area
                best = (x1, y1, x2 - x1, y2 - y1)
        out[pi] = best
    return out


def detect_bubble_bounds_combined(
    image: Image.Image,
    paragraphs: list[dict],
    img_w: int,
    img_h: int,
    prefer_yolo: bool | None = None,
    yolo_model_path: str = "",
) -> dict[int, tuple[float, float, float, float] | None]:
    """Detect bubble bounds with YOLO first, OpenCV as fallback.

    Strategy per paragraph:
    1. If YOLO is enabled and returned a bbox for this paragraph β†’ use it.
    2. Else if OpenCV found a connected component for this paragraph β†’ use it.
    3. Else β†’ ``None`` (renderer falls back to the text-only AABB).

    ``prefer_yolo`` defaults to the value of ``TP_BUBBLE_USE_YOLO``; set it
    explicitly to override on a per-call basis (e.g. in tests).
    """
    if prefer_yolo is None:
        prefer_yolo = _yolo_enabled()

    yolo_map: dict[int, tuple[float, float, float, float] | None] | None = None
    if prefer_yolo:
        yolo_map = detect_bubble_bounds_by_paragraph_yolo(
            image, paragraphs, img_w, img_h, model_path=yolo_model_path
        )

    opencv_map = detect_bubble_bounds_by_paragraph(image, paragraphs, img_w, img_h)

    if not yolo_map:
        return opencv_map

    merged: dict[int, tuple[float, float, float, float] | None] = {}
    for pi, ov in opencv_map.items():
        yv = yolo_map.get(pi)
        merged[pi] = yv if yv is not None else ov
    return merged


def attach_bubble_bounds(
    tree: dict,
    bubble_bounds_map: dict[int, tuple[float, float, float, float] | None],
) -> None:
    """Write each paragraph's bubble bounds into ``tree`` (mutates in place).

    The renderer reads ``para["bubble_bounds_px"]`` first and falls back
    to ``para["bounds_px"]`` (the text-only AABB) when it's missing.
    """
    if not isinstance(tree, dict):
        return
    for p in tree.get("paragraphs") or []:
        if not isinstance(p, dict):
            continue
        pi = int(p.get("para_index", -1))
        bounds = bubble_bounds_map.get(pi)
        if bounds is None:
            # Don't overwrite an existing value if the caller pre-populated one.
            p.setdefault("bubble_bounds_px", None)
            continue
        l, t, w, h = bounds
        p["bubble_bounds_px"] = [float(l), float(t), float(l + w), float(t + h)]