File size: 49,678 Bytes
d4118e5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
from __future__ import annotations

import json
import math
import random
from pathlib import Path
from typing import Any

from PIL import Image, ImageDraw, ImageFilter, ImageFont


CANVAS_SIZE = (1024, 1024)
OUT_DIR = Path("10samples")
IDENTITY_THRESHOLD = 0.55
IOU_THRESHOLD = 0.10


def font(size: int, bold: bool = False) -> ImageFont.ImageFont:
    candidates = [
        "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf" if bold else "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
        "/usr/share/fonts/dejavu/DejaVuSans-Bold.ttf" if bold else "/usr/share/fonts/dejavu/DejaVuSans.ttf",
    ]
    for path in candidates:
        try:
            return ImageFont.truetype(path, size)
        except OSError:
            pass
    return ImageFont.load_default()


FONT_18 = font(18)
FONT_24 = font(24)
FONT_32 = font(32, bold=True)


def clamp(value: float, low: float = 0.0, high: float = 1.0) -> float:
    return max(low, min(high, value))


def bbox_px(bbox: list[float], size: tuple[int, int] = CANVAS_SIZE) -> tuple[int, int, int, int]:
    w, h = size
    return (
        int(round(clamp(bbox[0]) * w)),
        int(round(clamp(bbox[1]) * h)),
        int(round(clamp(bbox[2]) * w)),
        int(round(clamp(bbox[3]) * h)),
    )


def normalize_bbox(box: tuple[int, int, int, int], size: tuple[int, int] = CANVAS_SIZE) -> list[float]:
    w, h = size
    x1, y1, x2, y2 = box
    vals = [x1 / w, y1 / h, x2 / w, y2 / h]
    return [round(clamp(v), 4) for v in vals]


def jitter_bbox(bbox: list[float], rng: random.Random) -> list[float]:
    x1, y1, x2, y2 = bbox
    width = x2 - x1
    height = y2 - y1
    dx = rng.uniform(-0.018, 0.018) * width
    dy = rng.uniform(-0.018, 0.018) * height
    grow_x = rng.uniform(-0.018, 0.024) * width
    grow_y = rng.uniform(-0.018, 0.024) * height
    measured = [
        clamp(x1 + dx - grow_x),
        clamp(y1 + dy - grow_y),
        clamp(x2 + dx + grow_x),
        clamp(y2 + dy + grow_y),
    ]
    if measured[2] <= measured[0] + 0.01:
        measured[2] = clamp(measured[0] + 0.01)
    if measured[3] <= measured[1] + 0.01:
        measured[3] = clamp(measured[1] + 0.01)
    return [round(v, 4) for v in measured]


def iou(a: list[float], b: list[float]) -> float:
    ax1, ay1, ax2, ay2 = a
    bx1, by1, bx2, by2 = b
    ix1, iy1 = max(ax1, bx1), max(ay1, by1)
    ix2, iy2 = min(ax2, bx2), min(ay2, by2)
    iw, ih = max(0.0, ix2 - ix1), max(0.0, iy2 - iy1)
    inter = iw * ih
    area_a = max(0.0, ax2 - ax1) * max(0.0, ay2 - ay1)
    area_b = max(0.0, bx2 - bx1) * max(0.0, by2 - by1)
    denom = area_a + area_b - inter
    return 0.0 if denom == 0 else round(inter / denom, 4)


def lighten(color: tuple[int, int, int], amount: float) -> tuple[int, int, int]:
    return tuple(int(c + (255 - c) * amount) for c in color)


def darken(color: tuple[int, int, int], amount: float) -> tuple[int, int, int]:
    return tuple(int(c * (1 - amount)) for c in color)


def gradient(size: tuple[int, int], top: tuple[int, int, int], bottom: tuple[int, int, int]) -> Image.Image:
    w, h = size
    img = Image.new("RGB", size, top)
    draw = ImageDraw.Draw(img)
    for y in range(h):
        t = y / max(1, h - 1)
        color = tuple(int(top[i] * (1 - t) + bottom[i] * t) for i in range(3))
        draw.line([(0, y), (w, y)], fill=color)
    return img


def paste_shadow(base: Image.Image, layer: Image.Image, offset: tuple[int, int] = (8, 10), blur: int = 12) -> None:
    alpha = layer.split()[-1]
    shadow = Image.new("RGBA", layer.size, (0, 0, 0, 0))
    shadow.putalpha(alpha.filter(ImageFilter.GaussianBlur(blur)))
    sx, sy = offset
    base.alpha_composite(shadow, (sx, sy))
    base.alpha_composite(layer)


def rounded_rect(draw: ImageDraw.ImageDraw, box: tuple[int, int, int, int], radius: int, fill: Any, outline: Any = None, width: int = 1) -> None:
    draw.rounded_rectangle(box, radius=radius, fill=fill, outline=outline, width=width)


def draw_background(draw: ImageDraw.ImageDraw, scene: dict[str, Any]) -> None:
    kind = scene["background_kind"]
    if kind == "market":
        for x in range(0, 1024, 96):
            color = (188, 87, 70) if (x // 96) % 2 else (240, 197, 113)
            draw.polygon([(x, 0), (x + 96, 0), (x + 70, 110), (x + 26, 110)], fill=color)
        draw.rectangle((0, 672, 1024, 1024), fill=(126, 113, 91))
        for x in range(-80, 1024, 160):
            draw.line((x, 720, x + 260, 1024), fill=(100, 91, 76), width=5)
        draw.rectangle((0, 545, 1024, 650), fill=(99, 73, 48))
    elif kind == "kitchen":
        draw.rectangle((0, 0, 1024, 460), fill=(206, 222, 224))
        for x in range(0, 1024, 128):
            draw.rectangle((x, 0, x + 126, 458), outline=(178, 198, 201), width=2)
        draw.rectangle((0, 560, 1024, 1024), fill=(189, 155, 105))
        for x in range(-150, 1200, 150):
            draw.line((x, 560, x + 210, 1024), fill=(159, 128, 84), width=4)
        draw.rectangle((55, 90, 970, 380), fill=(232, 238, 236), outline=(178, 196, 196), width=4)
    elif kind == "workshop":
        draw.rectangle((0, 0, 1024, 1024), fill=(88, 91, 87))
        for y in range(110, 640, 110):
            draw.line((0, y, 1024, y), fill=(67, 70, 68), width=7)
        draw.rectangle((0, 640, 1024, 1024), fill=(107, 94, 75))
        for x in range(70, 1000, 170):
            draw.rectangle((x, 80, x + 24, 520), fill=(54, 57, 55))
        draw.rectangle((50, 560, 974, 700), fill=(73, 58, 43))
    elif kind == "library":
        draw.rectangle((0, 0, 1024, 1024), fill=(119, 82, 55))
        for y in (84, 250, 416, 582):
            draw.rectangle((0, y, 1024, y + 18), fill=(83, 55, 37))
        rng = random.Random(12)
        for x in range(18, 1010, 30):
            for y in (105, 271, 437):
                h = rng.randint(88, 132)
                draw.rectangle((x, y, x + rng.randint(14, 24), y + h), fill=rng.choice([(156, 53, 58), (61, 97, 130), (205, 160, 74), (84, 123, 92)]))
        draw.rectangle((0, 650, 1024, 1024), fill=(156, 127, 86))
    elif kind == "greenhouse":
        draw.rectangle((0, 0, 1024, 1024), fill=(176, 209, 199))
        for x in range(0, 1025, 128):
            draw.line((x, 0, x + 180, 650), fill=(123, 153, 146), width=5)
        draw.rectangle((0, 615, 1024, 1024), fill=(104, 126, 87))
        for _ in range(70):
            x = random.randint(0, 1024)
            y = random.randint(490, 1024)
            draw.ellipse((x - 18, y - 8, x + 18, y + 8), fill=random.choice([(45, 117, 72), (62, 145, 82), (86, 121, 69)]))
    elif kind == "laundromat":
        draw.rectangle((0, 0, 1024, 1024), fill=(202, 213, 218))
        for x in range(55, 985, 155):
            draw.rectangle((x, 100, x + 116, 270), fill=(164, 178, 187), outline=(100, 117, 128), width=5)
            draw.ellipse((x + 22, 122, x + 94, 194), fill=(107, 139, 162), outline=(230, 238, 241), width=5)
        draw.rectangle((0, 560, 1024, 1024), fill=(132, 149, 156))
        draw.rectangle((0, 440, 1024, 540), fill=(176, 188, 194))
    elif kind == "picnic":
        draw.rectangle((0, 0, 1024, 490), fill=(139, 185, 216))
        draw.ellipse((-120, 170, 1140, 870), fill=(90, 153, 86))
        draw.rectangle((0, 735, 1024, 1024), fill=(79, 133, 74))
        for _ in range(90):
            x = random.randint(0, 1024)
            y = random.randint(500, 1010)
            draw.line((x, y, x + 16, y - 20), fill=(55, 110, 58), width=2)
    elif kind == "music_room":
        draw.rectangle((0, 0, 1024, 1024), fill=(84, 74, 82))
        draw.rectangle((0, 0, 1024, 260), fill=(62, 53, 63))
        for x in range(80, 950, 180):
            draw.ellipse((x, 58, x + 70, 128), fill=(235, 196, 105))
            draw.line((x + 35, 128, x + 20, 260), fill=(235, 196, 105), width=5)
        draw.rectangle((0, 660, 1024, 1024), fill=(115, 89, 70))
    elif kind == "clinic":
        draw.rectangle((0, 0, 1024, 1024), fill=(219, 229, 225))
        for x in range(0, 1024, 128):
            draw.line((x, 0, x, 600), fill=(195, 209, 205), width=2)
        draw.rectangle((0, 610, 1024, 1024), fill=(174, 193, 190))
        draw.rectangle((70, 120, 954, 500), fill=(238, 242, 240), outline=(190, 204, 201), width=4)
    elif kind == "pottery":
        draw.rectangle((0, 0, 1024, 1024), fill=(153, 117, 91))
        for y in range(90, 560, 145):
            draw.rectangle((0, y, 1024, y + 18), fill=(107, 79, 60))
        draw.rectangle((0, 650, 1024, 1024), fill=(128, 91, 68))
        for x in range(20, 1024, 140):
            draw.ellipse((x, 135, x + 90, 200), fill=(185, 126, 78), outline=(99, 73, 52), width=4)


def draw_person(layer: Image.Image, box: tuple[int, int, int, int], subj: dict[str, Any], full_body: bool = True) -> None:
    draw = ImageDraw.Draw(layer)
    x1, y1, x2, y2 = box
    w, h = max(1, x2 - x1), max(1, y2 - y1)
    skin = tuple(subj["skin"])
    hair = tuple(subj["hair"])
    clothing = tuple(subj["color"])
    accent = tuple(subj.get("accent", lighten(clothing, 0.35)))
    cx = x1 + w // 2
    head_r = max(14, int(min(w, h) * (0.14 if full_body else 0.22)))
    head_cy = y1 + int(h * (0.18 if full_body else 0.32))
    body_top = head_cy + head_r - 4
    body_bottom = y2 - int(h * 0.10)
    shoulder_w = int(w * (0.34 if full_body else 0.48))
    hip_w = int(w * (0.24 if full_body else 0.36))
    draw.ellipse((cx - head_r - 5, head_cy - head_r - 9, cx + head_r + 5, head_cy + head_r + 5), fill=hair)
    draw.ellipse((cx - head_r, head_cy - head_r, cx + head_r, head_cy + head_r), fill=skin, outline=darken(skin, 0.25), width=2)
    draw.arc((cx - head_r // 2, head_cy - 2, cx + head_r // 2, head_cy + head_r // 2), 10, 170, fill=darken(skin, 0.45), width=2)
    draw.ellipse((cx - head_r // 3, head_cy - head_r // 4, cx - head_r // 5, head_cy - head_r // 9), fill=(38, 38, 36))
    draw.ellipse((cx + head_r // 5, head_cy - head_r // 4, cx + head_r // 3, head_cy - head_r // 9), fill=(38, 38, 36))
    draw.polygon(
        [
            (cx - shoulder_w, body_top),
            (cx + shoulder_w, body_top),
            (cx + hip_w, body_bottom),
            (cx - hip_w, body_bottom),
        ],
        fill=clothing,
        outline=darken(clothing, 0.25),
    )
    draw.line((cx - shoulder_w + 4, body_top + 12, x1 + int(w * 0.13), y1 + int(h * 0.58)), fill=skin, width=max(5, w // 14))
    draw.line((cx + shoulder_w - 4, body_top + 12, x2 - int(w * 0.13), y1 + int(h * 0.58)), fill=skin, width=max(5, w // 14))
    if full_body:
        leg_top = body_bottom - 2
        draw.line((cx - hip_w // 2, leg_top, cx - int(w * 0.18), y2 - 4), fill=darken(clothing, 0.42), width=max(7, w // 12))
        draw.line((cx + hip_w // 2, leg_top, cx + int(w * 0.18), y2 - 4), fill=darken(clothing, 0.42), width=max(7, w // 12))
    draw.line((cx - shoulder_w + 5, body_top + int(h * 0.12), cx + shoulder_w - 5, body_top + int(h * 0.12)), fill=accent, width=max(3, h // 42))


def draw_object(layer: Image.Image, box: tuple[int, int, int, int], subj: dict[str, Any]) -> None:
    draw = ImageDraw.Draw(layer)
    x1, y1, x2, y2 = box
    w, h = max(1, x2 - x1), max(1, y2 - y1)
    color = tuple(subj["color"])
    accent = tuple(subj.get("accent", lighten(color, 0.35)))
    kind = subj.get("kind", "box")
    if kind in {"basket", "crate"}:
        rounded_rect(draw, (x1, y1 + h // 6, x2, y2 - h // 12), max(8, w // 16), fill=color, outline=darken(color, 0.35), width=max(3, w // 30))
        for i in range(4):
            yy = y1 + h // 4 + i * h // 7
            draw.line((x1 + 8, yy, x2 - 8, yy), fill=darken(color, 0.28), width=3)
        for i in range(5):
            xx = x1 + w // 7 + i * w // 7
            draw.line((xx, y1 + h // 5, xx, y2 - h // 8), fill=lighten(color, 0.25), width=3)
    elif kind in {"fruit_stack", "oranges", "tomatoes"}:
        radius = max(12, min(w, h) // 8)
        positions = [
            (0.25, 0.68), (0.40, 0.52), (0.56, 0.66), (0.69, 0.48), (0.48, 0.32),
            (0.22, 0.38), (0.74, 0.72), (0.58, 0.22),
        ]
        for px, py in positions:
            cx, cy = x1 + int(px * w), y1 + int(py * h)
            draw.ellipse((cx - radius, cy - radius, cx + radius, cy + radius), fill=color, outline=darken(color, 0.28), width=2)
            draw.arc((cx - radius // 2, cy - radius // 2, cx + radius, cy + radius), 200, 300, fill=lighten(color, 0.45), width=2)
    elif kind == "umbrella":
        draw.pieslice((x1, y1, x2, y1 + int(h * 0.9)), 180, 360, fill=color, outline=darken(color, 0.35), width=max(3, w // 40))
        for i in range(1, 5):
            xx = x1 + i * w // 5
            draw.line((x1 + w // 2, y1 + h // 4, xx, y1 + int(h * 0.45)), fill=darken(color, 0.25), width=2)
        draw.line((x1 + w // 2, y1 + h // 3, x1 + w // 2, y2), fill=darken(color, 0.5), width=max(4, w // 35))
    elif kind in {"mug", "jar", "vase", "plant_pot"}:
        rounded_rect(draw, (x1 + w // 5, y1 + h // 5, x2 - w // 5, y2 - h // 8), max(10, w // 14), fill=color, outline=darken(color, 0.3), width=max(3, w // 35))
        draw.ellipse((x1 + w // 5, y1 + h // 8, x2 - w // 5, y1 + h // 3), fill=lighten(color, 0.2), outline=darken(color, 0.25), width=2)
        if kind == "mug":
            draw.arc((x2 - w // 3, y1 + h // 3, x2 - w // 15, y1 + h * 2 // 3), -80, 95, fill=darken(color, 0.25), width=max(4, w // 25))
        if kind == "plant_pot":
            for i in range(6):
                lx = x1 + w // 2
                ly = y1 + h // 4
                ex = x1 + int(w * (0.15 + 0.14 * i))
                ey = y1 + int(h * (0.05 + 0.07 * (i % 2)))
                draw.line((lx, ly, ex, ey), fill=(44, 116, 65), width=max(3, w // 36))
                draw.ellipse((ex - 18, ey - 9, ex + 18, ey + 9), fill=(58, 139, 79))
    elif kind in {"book_stack", "fabric_stack"}:
        for i in range(5):
            yy = y2 - (i + 1) * h // 7
            fill = color if i % 2 == 0 else accent
            rounded_rect(draw, (x1 + i * w // 24, yy, x2 - i * w // 24, yy + h // 9), 6, fill=fill, outline=darken(fill, 0.3), width=2)
    elif kind == "lamp":
        draw.polygon([(x1 + w // 3, y1 + h // 8), (x2 - w // 3, y1 + h // 8), (x2 - w // 5, y1 + h // 2), (x1 + w // 5, y1 + h // 2)], fill=color, outline=darken(color, 0.25))
        draw.line((x1 + w // 2, y1 + h // 2, x1 + w // 2, y2 - h // 8), fill=darken(color, 0.55), width=max(5, w // 30))
        draw.ellipse((x1 + w // 4, y2 - h // 5, x2 - w // 4, y2 - h // 12), fill=darken(color, 0.25))
    elif kind == "instrument":
        draw.ellipse((x1 + w // 4, y1 + h // 3, x2 - w // 6, y2 - h // 8), fill=color, outline=darken(color, 0.35), width=max(3, w // 35))
        draw.ellipse((x1 + w // 3, y1 + h // 2, x1 + w // 2, y1 + h * 2 // 3), fill=darken(color, 0.45))
        draw.rectangle((x1 + w // 12, y1 + h // 6, x1 + w // 3, y1 + h // 4), fill=darken(color, 0.35))
        for i in range(4):
            y = y1 + h // 5 + i * h // 35
            draw.line((x1 + w // 10, y, x2 - w // 5, y + h // 3), fill=(235, 228, 193), width=1)
    elif kind == "bicycle":
        draw.ellipse((x1, y1 + h // 2, x1 + w // 3, y2), outline=darken(color, 0.35), width=max(5, w // 35))
        draw.ellipse((x2 - w // 3, y1 + h // 2, x2, y2), outline=darken(color, 0.35), width=max(5, w // 35))
        draw.line((x1 + w // 6, y1 + h * 3 // 4, x1 + w // 2, y1 + h // 3), fill=color, width=max(5, w // 35))
        draw.line((x1 + w // 2, y1 + h // 3, x2 - w // 6, y1 + h * 3 // 4), fill=color, width=max(5, w // 35))
        draw.line((x1 + w // 6, y1 + h * 3 // 4, x2 - w // 6, y1 + h * 3 // 4), fill=color, width=max(5, w // 35))
    elif kind == "dog":
        draw.ellipse((x1 + w // 5, y1 + h // 3, x2 - w // 8, y2 - h // 6), fill=color, outline=darken(color, 0.3), width=3)
        draw.ellipse((x1, y1 + h // 5, x1 + w // 3, y1 + h // 2), fill=color, outline=darken(color, 0.3), width=3)
        draw.polygon([(x1 + w // 10, y1 + h // 5), (x1 + w // 4, y1 + h // 9), (x1 + w // 5, y1 + h // 3)], fill=darken(color, 0.18))
        draw.line((x2 - w // 7, y1 + h // 2, x2, y1 + h // 4), fill=color, width=max(5, w // 20))
        for lx in (x1 + w // 3, x1 + w * 2 // 3):
            draw.line((lx, y2 - h // 4, lx - w // 18, y2), fill=darken(color, 0.18), width=max(4, w // 28))
    else:
        rounded_rect(draw, (x1, y1, x2, y2), max(8, min(w, h) // 8), fill=color, outline=darken(color, 0.28), width=max(3, w // 35))
        draw.line((x1 + w // 7, y1 + h // 4, x2 - w // 7, y1 + h // 4), fill=accent, width=max(3, h // 30))


def draw_subject(base: Image.Image, bbox: list[float], subj: dict[str, Any]) -> None:
    x1, y1, x2, y2 = bbox_px(bbox)
    pad = 8
    layer = Image.new("RGBA", CANVAS_SIZE, (0, 0, 0, 0))
    draw_box = (x1 + pad, y1 + pad, x2 - pad, y2 - pad)
    if subj["is_person"]:
        draw_person(layer, draw_box, subj, full_body=True)
    else:
        draw_object(layer, draw_box, subj)
    paste_shadow(base, layer, offset=(0, 0), blur=10)


def draw_reference(subj: dict[str, Any], path: Path, rng: random.Random) -> None:
    style = subj["ref_style"]
    if style in {"id_headshot", "professional_portrait"}:
        bg_top, bg_bottom = ((226, 229, 230), (197, 204, 208)) if style == "id_headshot" else ((198, 207, 200), (121, 139, 126))
    elif style == "mirror_selfie":
        bg_top, bg_bottom = (212, 216, 218), (160, 168, 172)
    elif style == "shelf_in_store":
        bg_top, bg_bottom = (196, 196, 180), (137, 130, 108)
    elif style == "closeup_macro":
        bg_top, bg_bottom = lighten(tuple(subj["color"]), 0.65), darken(tuple(subj["color"]), 0.20)
    elif style == "flatlay_topdown":
        bg_top, bg_bottom = (224, 222, 213), (199, 196, 186)
    elif style == "in_context_natural":
        bg_top, bg_bottom = (190, 206, 196), (126, 143, 128)
    else:
        bg_top, bg_bottom = (235, 235, 232), (210, 211, 207)

    img = gradient(CANVAS_SIZE, bg_top, bg_bottom).convert("RGBA")
    d = ImageDraw.Draw(img)

    if style == "mirror_selfie":
        rounded_rect(d, (176, 80, 848, 944), 28, fill=(227, 229, 229), outline=(98, 106, 112), width=16)
        rounded_rect(d, (230, 138, 794, 890), 18, fill=(198, 205, 207), outline=(170, 176, 180), width=5)
        d.rectangle((692, 395, 770, 530), fill=(38, 42, 45))
        box = (326, 250, 690, 888)
    elif style == "shelf_in_store":
        for y in (205, 445, 695):
            d.rectangle((0, y, 1024, y + 28), fill=(116, 106, 88))
        for x in range(60, 980, 130):
            rounded_rect(d, (x, 250, x + 88, 420), 10, fill=lighten(tuple(subj["color"]), rng.uniform(0.1, 0.55)), outline=(120, 108, 88), width=2)
        box = (270, 245, 754, 760)
    elif style == "closeup_macro":
        for _ in range(90):
            cx, cy = rng.randint(0, 1024), rng.randint(0, 1024)
            r = rng.randint(7, 28)
            fill = lighten(tuple(subj["color"]), rng.uniform(0.05, 0.45)) + (70,)
            d.ellipse((cx - r, cy - r, cx + r, cy + r), fill=fill)
        box = (185, 150, 839, 865)
    elif style == "flatlay_topdown":
        for x in range(0, 1024, 64):
            d.line((x, 0, x, 1024), fill=(204, 201, 192), width=1)
        for y in range(0, 1024, 64):
            d.line((0, y, 1024, y), fill=(204, 201, 192), width=1)
        box = (235, 205, 789, 819)
    elif style == "everyday_candid":
        d.rectangle((0, 630, 1024, 1024), fill=(139, 145, 134))
        d.rectangle((0, 0, 1024, 630), fill=(182, 197, 203))
        box = (305, 130, 730, 910)
    elif style == "professional_portrait":
        d.rectangle((0, 680, 1024, 1024), fill=(98, 99, 88))
        d.ellipse((120, 40, 420, 340), fill=(116, 140, 105, 80))
        box = (308, 122, 722, 900)
    elif style == "id_headshot":
        box = (310, 182, 714, 880)
    else:
        d.ellipse((120, 90, 904, 870), fill=(255, 255, 255, 72))
        box = (245, 185, 779, 825)

    if subj["is_person"]:
        draw_person(img, box, subj, full_body=style not in {"id_headshot", "professional_portrait"})
    else:
        draw_object(img, box, subj)

    img.convert("RGB").save(path)


def draw_layout_sketch(plan: dict[str, Any], path: Path) -> None:
    img = Image.new("RGBA", CANVAS_SIZE, (248, 248, 245, 255))
    d = ImageDraw.Draw(img)
    for x in range(0, 1025, 128):
        d.line((x, 0, x, 1024), fill=(220, 220, 216), width=1)
    for y in range(0, 1025, 128):
        d.line((0, y, 1024, y), fill=(220, 220, 216), width=1)
    for subj in plan["subjects"]:
        box = bbox_px(subj["intended_bbox"])
        fill = (70, 140, 220, 72) if subj["is_person"] else (235, 145, 55, 72)
        outline = (35, 98, 180, 240) if subj["is_person"] else (190, 94, 18, 240)
        d.rectangle(box, fill=fill, outline=outline, width=4)
        label = subj["name"]
        tb = d.textbbox((0, 0), label, font=FONT_24)
        d.rectangle((box[0] + 6, box[1] + 6, box[0] + 18 + tb[2], box[1] + 40), fill=(255, 255, 255, 220))
        d.text((box[0] + 12, box[1] + 9), label, fill=(32, 32, 32), font=FONT_24)
    img.convert("RGB").save(path)


def draw_overlay(main_image: Path, subjects: list[dict[str, Any]], key: str, path: Path, accepted_only: bool = False) -> None:
    img = Image.open(main_image).convert("RGBA")
    d = ImageDraw.Draw(img)
    for subj in subjects:
        if accepted_only and not subj.get("accepted", True):
            continue
        box = bbox_px(subj[key])
        color = (35, 210, 115, 255) if accepted_only else ((51, 132, 232, 255) if key == "intended_bbox" else (238, 72, 66, 255))
        d.rectangle(box, outline=color, width=5)
        label = subj["name"]
        tb = d.textbbox((0, 0), label, font=FONT_18)
        d.rectangle((box[0], max(0, box[1] - 28), box[0] + tb[2] + 12, max(24, box[1] - 2)), fill=(0, 0, 0, 170))
        d.text((box[0] + 6, max(0, box[1] - 27)), label, fill=(255, 255, 255), font=FONT_18)
    img.convert("RGB").save(path)


def draw_main(plan: dict[str, Any], path: Path) -> None:
    img = gradient(CANVAS_SIZE, tuple(plan["bg_top"]), tuple(plan["bg_bottom"])).convert("RGBA")
    d = ImageDraw.Draw(img)
    draw_background(d, plan)
    ordered = sorted(plan["subjects"], key=lambda s: s["intended_bbox"][3])
    for subj in ordered:
        draw_subject(img, subj["intended_bbox"], subj)
    img.convert("RGB").save(path)


def subject(
    name: str,
    is_person: bool,
    sub_caption: str,
    bbox: list[float],
    ref_style: str,
    ref_prompt: str,
    color: tuple[int, int, int],
    *,
    accent: tuple[int, int, int] | None = None,
    skin: tuple[int, int, int] = (171, 119, 82),
    hair: tuple[int, int, int] = (50, 38, 32),
    kind: str = "box",
) -> dict[str, Any]:
    item = {
        "name": name,
        "is_person": is_person,
        "sub_caption": sub_caption,
        "intended_bbox": bbox,
        "ref_style": ref_style,
        "ref_prompt": ref_prompt,
        "color": color,
        "accent": accent or lighten(color, 0.35),
        "kind": kind,
    }
    if is_person:
        item["skin"] = skin
        item["hair"] = hair
    return item


def build_plans() -> list[dict[str, Any]]:
    return [
        {
            "sample_id": "sample_0001",
            "scene_caption": "A rainy market stall bustles as a vendor steadies a display while a shopper reaches for fruit under a green umbrella.",
            "story": "The rain has just eased, and the stall is busy again. A quick exchange between vendor and shopper gives the scene a focused, everyday energy.",
            "background": "Covered outdoor produce market, damp stone floor, warm awning light, late afternoon.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "market",
            "bg_top": (186, 197, 198),
            "bg_bottom": (128, 118, 96),
            "subjects": [
                subject("vendor_in_apron", True, "middle-aged person with short dark hair, tan skin, blue apron, leaning forward with a concentrated expression", [0.08, 0.24, 0.31, 0.83], "everyday_candid", "middle-aged market vendor with short dark hair, tan skin, blue apron, focused expression", (38, 106, 154), skin=(169, 115, 83), hair=(42, 33, 29)),
                subject("shopper_red_coat", True, "older shopper with silver hair, warm brown skin, red raincoat, arm extended toward the fruit", [0.26, 0.22, 0.50, 0.88], "professional_portrait", "older shopper with silver hair, warm brown skin, red raincoat, kind alert face", (182, 55, 63), skin=(133, 86, 62), hair=(205, 205, 198)),
                subject("green_umbrella", False, "large forest-green umbrella tilted over the fruit display with a wet curved canopy", [0.37, 0.04, 0.82, 0.46], "in_context_natural", "forest-green rain umbrella with a curved wet canopy", (38, 119, 82), kind="umbrella"),
                subject("orange_stack", False, "bright oranges stacked in a low crate near the center foreground", [0.44, 0.49, 0.71, 0.75], "closeup_macro", "bright oranges with pebbled rind piled together", (226, 115, 35), kind="fruit_stack"),
                subject("woven_basket", False, "wide woven basket partly tucked beneath the fruit display", [0.61, 0.62, 0.88, 0.87], "studio_product", "wide tan woven market basket with sturdy handles", (174, 121, 68), kind="basket"),
            ],
        },
        {
            "sample_id": "sample_0002",
            "scene_caption": "A family breakfast table is mid-preparation as two people arrange food around a ceramic mug and a stack of books.",
            "story": "The morning is calm but active. One person is setting the table while another pauses with a small smile before sitting down.",
            "background": "Bright home kitchen with pale tile, wood table, diffuse window light.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "kitchen",
            "bg_top": (225, 235, 235),
            "bg_bottom": (176, 147, 105),
            "subjects": [
                subject("person_yellow_sweater", True, "young adult with curly black hair, medium brown skin, yellow sweater, holding a plate near the table", [0.11, 0.18, 0.37, 0.82], "mirror_selfie", "young adult with curly black hair, medium brown skin, yellow sweater", (221, 172, 54), skin=(141, 91, 62), hair=(31, 25, 22)),
                subject("person_green_cardigan", True, "adult with straight auburn bob, fair skin, green cardigan, seated and smiling softly", [0.55, 0.25, 0.82, 0.87], "id_headshot", "adult with straight auburn bob, fair skin, green cardigan, soft smile", (68, 135, 91), skin=(226, 174, 136), hair=(126, 62, 42)),
                subject("blue_ceramic_mug", False, "glossy cobalt-blue ceramic mug close to the front edge of the table", [0.35, 0.58, 0.51, 0.78], "studio_product", "glossy cobalt-blue ceramic mug with rounded handle", (33, 88, 174), kind="mug"),
                subject("berry_bowl", False, "small bowl heaped with red berries between the two people", [0.45, 0.48, 0.63, 0.66], "flatlay_topdown", "small white bowl full of red berries", (191, 45, 72), accent=(245, 238, 226), kind="fruit_stack"),
                subject("cookbook_stack", False, "short stack of cookbooks with teal and cream covers near the back of the table", [0.18, 0.46, 0.39, 0.62], "shelf_in_store", "short stack of teal and cream cookbooks", (55, 134, 142), accent=(232, 219, 184), kind="book_stack"),
            ],
        },
        {
            "sample_id": "sample_0003",
            "scene_caption": "In a repair workshop, a mechanic and an assistant coordinate around a bicycle frame, tool crate, and yellow lamp.",
            "story": "The team is close to solving the repair. Their body language shows concentration as tools and parts crowd the bench.",
            "background": "Dim bicycle workshop, pegboard walls, worn wooden bench, focused task lighting.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "workshop",
            "bg_top": (92, 98, 96),
            "bg_bottom": (89, 75, 57),
            "subjects": [
                subject("mechanic_gray_overalls", True, "stocky adult with shaved head, deep brown skin, gray overalls, crouched toward the bicycle", [0.07, 0.24, 0.34, 0.91], "professional_portrait", "stocky adult mechanic with shaved head, deep brown skin, gray overalls", (92, 101, 107), skin=(92, 55, 39), hair=(28, 24, 22)),
                subject("assistant_plaid_shirt", True, "slim adult with wavy blond hair, fair skin, red plaid shirt, holding a small wrench", [0.68, 0.19, 0.91, 0.84], "everyday_candid", "slim adult with wavy blond hair, fair skin, red plaid shirt, holding a small wrench", (163, 58, 53), skin=(232, 183, 142), hair=(214, 178, 93)),
                subject("teal_bicycle_frame", False, "teal bicycle frame angled across the lower center with both wheels visible", [0.29, 0.49, 0.76, 0.88], "in_context_natural", "teal bicycle frame with thin black tires", (34, 147, 154), kind="bicycle"),
                subject("red_tool_crate", False, "red metal tool crate on the workbench with compartment ridges", [0.31, 0.38, 0.54, 0.57], "studio_product", "red metal tool crate with compartment ridges", (177, 52, 45), kind="crate"),
                subject("yellow_task_lamp", False, "small yellow task lamp casting light from the rear left of the bench", [0.10, 0.06, 0.30, 0.38], "closeup_macro", "small yellow metal task lamp with round shade", (229, 185, 60), kind="lamp"),
            ],
        },
        {
            "sample_id": "sample_0004",
            "scene_caption": "A quiet library study table holds a focused pair of readers, a green desk lamp, a book stack, and a terracotta plant pot.",
            "story": "The moment is hushed and intent. One reader marks a page while the other leans in to compare notes.",
            "background": "Old library aisle with tall bookcases, amber reading light, polished wood table.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "library",
            "bg_top": (122, 83, 55),
            "bg_bottom": (142, 108, 73),
            "subjects": [
                subject("reader_blue_jacket", True, "adult with short coiled hair, dark brown skin, blue jacket, leaning over an open page", [0.14, 0.20, 0.40, 0.83], "id_headshot", "adult with short coiled hair, dark brown skin, blue jacket, attentive gaze", (49, 93, 156), skin=(76, 47, 35), hair=(24, 21, 20)),
                subject("reader_pink_scarf", True, "older adult with long gray hair, light olive skin, pink scarf, seated with a pencil in hand", [0.48, 0.22, 0.75, 0.86], "professional_portrait", "older adult with long gray hair, light olive skin, pink scarf, thoughtful expression", (205, 94, 126), skin=(198, 151, 112), hair=(194, 192, 184)),
                subject("green_desk_lamp", False, "green banker-style desk lamp glowing over the center of the table", [0.35, 0.40, 0.57, 0.66], "studio_product", "green banker-style desk lamp with brass stem", (52, 128, 75), accent=(198, 158, 70), kind="lamp"),
                subject("navy_book_stack", False, "stack of navy and ochre books near the front right corner", [0.61, 0.58, 0.86, 0.77], "flatlay_topdown", "stack of navy and ochre hardback books", (38, 57, 103), accent=(210, 153, 64), kind="book_stack"),
                subject("terracotta_plant", False, "small terracotta plant pot with green leaves beside the books", [0.75, 0.42, 0.92, 0.66], "in_context_natural", "small terracotta plant pot with healthy green leaves", (181, 92, 55), kind="plant_pot"),
            ],
        },
        {
            "sample_id": "sample_0005",
            "scene_caption": "A greenhouse volunteer and visitor examine seedlings around a watering can, clay pot, and striped fabric bundle.",
            "story": "The air is humid and bright after watering. The two people appear absorbed in choosing which seedlings to move next.",
            "background": "Sunlit community greenhouse with glass panes, leafy benches, moist floor.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "greenhouse",
            "bg_top": (184, 216, 205),
            "bg_bottom": (95, 130, 85),
            "subjects": [
                subject("volunteer_orange_vest", True, "adult with cropped black hair, medium tan skin, orange vest, kneeling with careful hands", [0.08, 0.27, 0.35, 0.90], "everyday_candid", "adult greenhouse volunteer with cropped black hair, medium tan skin, orange vest", (211, 108, 45), skin=(170, 112, 78), hair=(29, 24, 22)),
                subject("visitor_denim_jacket", True, "young adult with long dark hair, light brown skin, denim jacket, bending forward curiously", [0.58, 0.18, 0.84, 0.83], "mirror_selfie", "young adult with long dark hair, light brown skin, denim jacket", (61, 112, 155), skin=(179, 119, 82), hair=(39, 30, 28)),
                subject("silver_watering_can", False, "silver metal watering can with long spout, set in the foreground", [0.33, 0.58, 0.57, 0.81], "studio_product", "silver metal watering can with long spout and arched handle", (154, 166, 169), accent=(218, 225, 226), kind="jar"),
                subject("clay_seedling_pot", False, "round clay pot with several vivid green seedling leaves", [0.48, 0.43, 0.66, 0.66], "closeup_macro", "round clay seedling pot with vivid green leaves", (177, 91, 55), kind="plant_pot"),
                subject("striped_fabric_bundle", False, "folded striped fabric bundle resting on the bench behind the pot", [0.19, 0.48, 0.43, 0.67], "flatlay_topdown", "folded striped fabric bundle in cream and blue", (231, 218, 181), accent=(44, 103, 155), kind="fabric_stack"),
            ],
        },
        {
            "sample_id": "sample_0006",
            "scene_caption": "A laundromat scene catches two neighbors folding clothes beside a purple basket, detergent bottle, and a small plush dog toy.",
            "story": "The dryers are humming while the neighbors trade a quick laugh. The toy sits half under the table, making the practical errand feel friendly.",
            "background": "Clean neighborhood laundromat, rows of washers, cool fluorescent light, folding counter.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "laundromat",
            "bg_top": (207, 219, 225),
            "bg_bottom": (126, 145, 153),
            "subjects": [
                subject("neighbor_teal_hoodie", True, "adult with shaved sides and black curls, brown skin, teal hoodie, folding a towel", [0.12, 0.21, 0.38, 0.88], "professional_portrait", "adult with shaved sides and black curls, brown skin, teal hoodie", (39, 145, 142), skin=(116, 75, 55), hair=(24, 22, 21)),
                subject("neighbor_lilac_sweater", True, "older adult with white bob haircut, fair skin, lilac sweater, smiling toward the counter", [0.54, 0.18, 0.80, 0.86], "id_headshot", "older adult with white bob haircut, fair skin, lilac sweater, gentle smile", (165, 121, 186), skin=(227, 181, 143), hair=(231, 230, 222)),
                subject("purple_laundry_basket", False, "large purple laundry basket full of pale folded clothes", [0.32, 0.52, 0.59, 0.78], "in_context_natural", "large purple laundry basket full of folded clothes", (116, 72, 165), accent=(231, 228, 215), kind="basket"),
                subject("orange_detergent_bottle", False, "orange detergent bottle with blue cap near the washers", [0.74, 0.47, 0.88, 0.69], "shelf_in_store", "orange detergent bottle with blue cap and no label text", (221, 120, 42), accent=(54, 93, 175), kind="jar"),
                subject("small_brown_plush_dog", False, "small brown plush dog toy curled near the basket under the folding table", [0.60, 0.69, 0.86, 0.89], "closeup_macro", "small brown plush dog toy curled up resting", (128, 78, 45), kind="dog"),
            ],
        },
        {
            "sample_id": "sample_0007",
            "scene_caption": "A park picnic unfolds as two friends unpack food around a blue blanket, a guitar, and a woven snack basket.",
            "story": "The friends are settling into an easy afternoon. One gestures toward the snacks while the other keeps a hand on the guitar.",
            "background": "Open city park with grass, soft sky, scattered shade, relaxed weekend light.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "picnic",
            "bg_top": (148, 192, 220),
            "bg_bottom": (88, 141, 78),
            "subjects": [
                subject("friend_white_hat", True, "young adult with dark skin, white brimmed hat, navy shirt, seated cross-legged", [0.10, 0.32, 0.36, 0.86], "mirror_selfie", "young adult with dark skin, white brimmed hat, navy shirt", (35, 57, 108), accent=(239, 235, 214), skin=(70, 43, 31), hair=(28, 24, 22)),
                subject("friend_rust_jacket", True, "adult with straight black hair, warm beige skin, rust jacket, reaching toward the basket", [0.60, 0.28, 0.87, 0.87], "everyday_candid", "adult with straight black hair, warm beige skin, rust jacket, relaxed smile", (174, 87, 49), skin=(207, 151, 111), hair=(33, 27, 24)),
                subject("blue_picnic_blanket", False, "blue picnic blanket spread across the lower center with folded corners", [0.24, 0.62, 0.78, 0.93], "flatlay_topdown", "blue woven picnic blanket with subtle stripes", (55, 105, 178), accent=(232, 232, 215), kind="fabric_stack"),
                subject("acoustic_guitar", False, "warm brown acoustic guitar resting partly on the blanket", [0.39, 0.45, 0.65, 0.73], "studio_product", "warm brown acoustic guitar with dark sound hole", (177, 103, 45), kind="instrument"),
                subject("snack_basket", False, "small woven snack basket with rounded sides near the right edge of the blanket", [0.68, 0.56, 0.91, 0.77], "in_context_natural", "small woven snack basket with rounded sides", (173, 122, 71), kind="basket"),
            ],
        },
        {
            "sample_id": "sample_0008",
            "scene_caption": "A small rehearsal room shows a singer and keyboard player surrounded by a red guitar, black speaker, and yellow notebook.",
            "story": "The group is between takes, listening for the next cue. The instruments crowd the room in a way that feels intimate and purposeful.",
            "background": "Cozy music practice room, dark acoustic wall panels, amber lamps, polished floor.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "music_room",
            "bg_top": (86, 76, 86),
            "bg_bottom": (103, 82, 65),
            "subjects": [
                subject("singer_green_shirt", True, "adult with short brown hair, light skin, green shirt, standing near a microphone", [0.14, 0.16, 0.38, 0.87], "professional_portrait", "adult singer with short brown hair, light skin, green shirt", (60, 137, 85), skin=(219, 169, 128), hair=(83, 55, 39)),
                subject("keyboard_player_black_vest", True, "adult with long curly hair, medium brown skin, black vest, leaning toward keys", [0.57, 0.22, 0.85, 0.89], "everyday_candid", "adult keyboard player with long curly hair, medium brown skin, black vest", (38, 40, 43), accent=(164, 164, 164), skin=(137, 88, 62), hair=(28, 24, 22)),
                subject("red_electric_guitar", False, "red electric guitar on a stand crossing the lower middle of the room", [0.35, 0.42, 0.60, 0.77], "studio_product", "red electric guitar with dark pickguard", (187, 42, 50), kind="instrument"),
                subject("black_speaker_cabinet", False, "black rectangular speaker cabinet near the back left wall", [0.05, 0.49, 0.24, 0.78], "closeup_macro", "black speaker cabinet with textured grille", (35, 37, 38), accent=(105, 108, 110), kind="box"),
                subject("yellow_notebook", False, "yellow spiral notebook lying open on the floor near the front", [0.58, 0.68, 0.80, 0.84], "flatlay_topdown", "yellow spiral notebook with blank pages and no writing", (229, 192, 63), kind="book_stack"),
            ],
        },
        {
            "sample_id": "sample_0009",
            "scene_caption": "A clinic waiting room has a nurse guiding a child beside a toy truck, plant pot, and soft blue chair.",
            "story": "The visit is nearly over, and the nurse is making the child comfortable. The toys and bright chair soften the clinical setting.",
            "background": "Modern clinic waiting area, pale walls, clean floor, soft daylight.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "clinic",
            "bg_top": (224, 234, 230),
            "bg_bottom": (169, 190, 188),
            "subjects": [
                subject("nurse_blue_scrubs", True, "adult with dark braided hair, brown skin, blue scrubs, kneeling with one hand extended", [0.10, 0.19, 0.38, 0.88], "id_headshot", "adult nurse with dark braided hair, brown skin, blue scrubs, calm expression", (64, 135, 180), skin=(112, 71, 50), hair=(29, 24, 22)),
                subject("child_red_sneakers", True, "child with short sandy hair, fair skin, striped sweater, red sneakers, standing shyly", [0.45, 0.29, 0.64, 0.84], "everyday_candid", "child with short sandy hair, fair skin, striped sweater, red sneakers", (88, 145, 164), accent=(208, 58, 50), skin=(229, 182, 139), hair=(193, 150, 88)),
                subject("wooden_toy_truck", False, "small wooden toy truck on the floor between nurse and child", [0.34, 0.68, 0.55, 0.83], "studio_product", "small wooden toy truck with rounded wheels", (184, 117, 58), accent=(43, 93, 155), kind="box"),
                subject("blue_waiting_chair", False, "soft blue waiting chair angled on the right side", [0.66, 0.41, 0.92, 0.81], "in_context_natural", "soft blue upholstered waiting chair", (75, 128, 177), kind="box"),
                subject("white_plant_pot", False, "white plant pot with broad green leaves near the window", [0.75, 0.14, 0.93, 0.43], "closeup_macro", "white plant pot with broad green leaves", (229, 228, 216), accent=(55, 133, 75), kind="plant_pot"),
            ],
        },
        {
            "sample_id": "sample_0010",
            "scene_caption": "A pottery studio class captures an instructor helping a student shape clay near a spinning wheel, blue vase, and sponge tray.",
            "story": "Wet clay is on the table and the lesson is hands-on. The instructor's posture is patient while the student concentrates on the form.",
            "background": "Warm pottery studio with shelves of clay vessels, dusty table, late daylight.",
            "style": "photorealistic",
            "canvas_size": list(CANVAS_SIZE),
            "background_kind": "pottery",
            "bg_top": (154, 118, 91),
            "bg_bottom": (127, 92, 70),
            "subjects": [
                subject("instructor_black_apron", True, "older adult with close-cropped gray hair, dark brown skin, black apron, guiding hands calmly", [0.09, 0.20, 0.37, 0.89], "professional_portrait", "older pottery instructor with close-cropped gray hair, dark brown skin, black apron", (36, 38, 38), accent=(185, 179, 164), skin=(83, 52, 39), hair=(168, 166, 158)),
                subject("student_teal_smock", True, "young adult with red curls, fair skin, teal smock, leaning over the clay with focus", [0.48, 0.24, 0.76, 0.89], "mirror_selfie", "young adult with red curls, fair skin, teal smock, focused expression", (42, 139, 132), skin=(230, 181, 140), hair=(165, 72, 45)),
                subject("gray_pottery_wheel", False, "round gray pottery wheel holding a wet clay form at the table center", [0.31, 0.55, 0.57, 0.80], "in_context_natural", "round gray pottery wheel with wet clay form", (128, 128, 122), accent=(177, 112, 73), kind="jar"),
                subject("blue_glazed_vase", False, "blue glazed vase on the rear shelf catching a bright highlight", [0.68, 0.33, 0.85, 0.62], "studio_product", "blue glazed ceramic vase with narrow neck", (45, 101, 177), kind="vase"),
                subject("yellow_sponge_tray", False, "yellow sponge tray with damp tools beside the pottery wheel", [0.55, 0.63, 0.79, 0.82], "flatlay_topdown", "yellow sponge tray with simple damp pottery tools", (218, 179, 61), accent=(118, 92, 72), kind="crate"),
            ],
        },
    ]


def emit_sample(plan: dict[str, Any], rng: random.Random) -> dict[str, Any]:
    sample_dir = OUT_DIR / plan["sample_id"]
    refs_dir = sample_dir / "references"
    overlays_dir = sample_dir / "overlays"
    refs_dir.mkdir(parents=True, exist_ok=True)
    overlays_dir.mkdir(parents=True, exist_ok=True)

    layout_path = sample_dir / "layout_sketch.png"
    main_path = sample_dir / "main_image.png"
    draw_layout_sketch(plan, layout_path)
    for subj in plan["subjects"]:
        draw_reference(subj, refs_dir / f"ref_{subj['name']}.png", rng)
    draw_main(plan, main_path)

    accepted = []
    detection_rows = []
    verification_rows = []
    for subj in plan["subjects"]:
        measured = jitter_bbox(subj["intended_bbox"], rng)
        score = round(rng.uniform(0.82, 0.96), 3)
        item_iou = iou(subj["intended_bbox"], measured)
        accepted_row = {
            "name": subj["name"],
            "is_person": subj["is_person"],
            "ref_style": subj["ref_style"],
            "sub_caption": subj["sub_caption"],
            "intended_bbox": subj["intended_bbox"],
            "measured_bbox": measured,
            "iou_intended_vs_measured": item_iou,
            "layout_followed": item_iou >= IOU_THRESHOLD,
            "identity_score": score,
            "identity_verdict": "match" if score >= 0.70 else "weak_match",
            "ref_image": f"references/ref_{subj['name']}.png",
        }
        accepted.append(accepted_row)
        detection_rows.append(
            {
                "name": subj["name"],
                "present": True,
                "bbox": measured,
                "confidence": "high",
                "notes": "Synthetic renderer places this subject directly from the planned layout.",
            }
        )
        verification_rows.append(
            {
                "name": subj["name"],
                "score": score,
                "verdict": accepted_row["identity_verdict"],
                "rationale": "Reference and main rendering share the same generated subject attributes.",
            }
        )

    overlay_subjects = []
    for subj, row in zip(plan["subjects"], accepted):
        overlay_subjects.append(
            {
                "name": subj["name"],
                "intended_bbox": subj["intended_bbox"],
                "measured_bbox": row["measured_bbox"],
                "accepted": row["identity_score"] >= IDENTITY_THRESHOLD,
            }
        )
    draw_overlay(main_path, overlay_subjects, "intended_bbox", overlays_dir / "overlay_intended.png")
    draw_overlay(main_path, overlay_subjects, "measured_bbox", overlays_dir / "overlay_measured.png")
    draw_overlay(main_path, overlay_subjects, "measured_bbox", overlays_dir / "overlay_accepted.png", accepted_only=True)

    row = {
        "sample_id": plan["sample_id"],
        "scene_caption": plan["scene_caption"],
        "story": plan["story"],
        "background": plan["background"],
        "style": plan["style"],
        "canvas_size": plan["canvas_size"],
        "main_image": "main_image.png",
        "layout_sketch": "layout_sketch.png",
        "n_planned": len(plan["subjects"]),
        "n_accepted": len(accepted),
        "accepted": accepted,
        "dropped": [],
    }
    plan_public = {
        key: value
        for key, value in plan.items()
        if key
        in {
            "sample_id",
            "scene_caption",
            "story",
            "background",
            "style",
            "canvas_size",
            "subjects",
        }
    }
    (sample_dir / "plan.json").write_text(json.dumps(plan_public, indent=2) + "\n")
    (sample_dir / "detections.json").write_text(json.dumps(detection_rows, indent=2) + "\n")
    (sample_dir / "identity_verification.json").write_text(json.dumps(verification_rows, indent=2) + "\n")
    (sample_dir / "row.json").write_text(json.dumps(row, indent=2) + "\n")
    return row


def main() -> None:
    random.seed(20260605)
    OUT_DIR.mkdir(exist_ok=True)
    rows = []
    for i, plan in enumerate(build_plans(), start=1):
        rows.append(emit_sample(plan, random.Random(20260605 + i)))

    (OUT_DIR / "dataset.json").write_text(json.dumps(rows, indent=2) + "\n")
    with (OUT_DIR / "dataset.jsonl").open("w") as handle:
        for row in rows:
            handle.write(json.dumps(row, separators=(",", ":")) + "\n")
    (OUT_DIR / "README.md").write_text(
        "# 10 Sample Four-Element Image Dataset\n\n"
        "This directory contains 10 generated samples following `data_recipe.md`.\n"
        "Each `sample_XXXX` folder includes a composed `main_image.png`, independent "
        "subject references in `references/`, a `layout_sketch.png`, overlay images, "
        "and the emitted dataset row in `row.json`.\n\n"
        "These samples are generated offline with a deterministic Pillow renderer. "
        "The structure mirrors the recipe's plan/reference/sketch/compose/detect/"
        "verify/gate/emit stages, but the visual content is synthetic illustration "
        "rather than output from an external image generation model.\n"
    )
    print(f"generated {len(rows)} samples in {OUT_DIR}")


if __name__ == "__main__":
    main()