File size: 32,797 Bytes
b485bdf
 
 
 
 
 
071ea55
 
 
 
b485bdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
071ea55
 
 
 
 
 
 
 
 
 
 
b485bdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
071ea55
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b485bdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
071ea55
 
 
 
 
 
 
b485bdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8741801
 
 
 
 
 
 
 
 
b485bdf
071ea55
 
 
 
 
 
 
 
b485bdf
 
 
 
 
 
 
 
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
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
"""
Oppai ONNX tagger — single-file launcher.

First run: creates `.venv/`, installs requirements, re-execs inside the venv,
then starts a local Gradio web UI on http://127.0.0.1:7860 .

If the user has no Python at all, `run.bat` instead drops a portable Python
into `.pyenv/` and uses that. In that case `app.py` skips venv creation and
installs requirements directly into the portable Python's site-packages.

Subsequent runs: skip install (marker file) and start the UI immediately.

For most users:
    Just run `run.bat` (or `py app.py`). The launcher will show a numbered
    menu so you can pick an existing model or download one — no flags
    required. Press Enter to accept the highlighted default at any prompt.

Advanced flags:
    py app.py --reinstall              # force re-install of requirements
    py app.py --model-dir <folder>     # skip the menu, load a specific folder

Models live in folders next to this script. Any folder containing
`model.onnx`, `selected_tags.csv`, and `preprocessing.json` is treated as a
model and will appear in the launcher menu and the UI's model picker. You
can also download variants from https://huggingface.co/Grio43/OppaiOracle
directly from the menu or from the UI.
"""

from __future__ import annotations

import os
import subprocess
import sys
import venv
from pathlib import Path

ROOT = Path(__file__).resolve().parent
VENV_DIR = ROOT / ".venv"
MARKER = VENV_DIR / ".bootstrapped"

# Portable Python lives here when run.bat had to download one because no
# system Python was available. When app.py is launched by .pyenv/python.exe,
# we install requirements directly into that interpreter — no venv needed.
PYENV_DIR = ROOT / ".pyenv"
PYENV_MARKER = PYENV_DIR / ".bootstrapped"

# Microsoft Visual C++ 2015-2022 redistributable (x64). onnxruntime's native
# DLLs depend on this; if it's missing the import fails with
# "DLL load failed while importing onnxruntime_pybind11_state".
VC_REDIST_URL = "https://aka.ms/vs/17/release/vc_redist.x64.exe"

# Default folder if it exists; otherwise the first auto-discovered folder
# next to this script is used. Override with --model-dir or the UI picker.
DEFAULT_MODEL_DIR = ROOT / "V1.1_onnx"

# Variants published on HuggingFace that are usable with this ONNX runtime.
# First entry is the recommended default in interactive prompts.
HF_REPO_ID = "Grio43/OppaiOracle"
HF_VARIANTS = ["V1.1_onnx", "V1_onnx"]
HF_VARIANT_DESC = {
    "V1.1_onnx": "448×448, higher accuracy",
    "V1_onnx":   "320×320, smaller and faster",
}

REQUIREMENTS = [
    "onnxruntime>=1.20",
    "pillow>=10.0",
    "numpy>=1.26,<3",
    "gradio>=4.44",
    "huggingface_hub>=0.24",
]


# ---------------------------------------------------------------------------
# Bootstrap
# ---------------------------------------------------------------------------

def _venv_python() -> Path:
    if os.name == "nt":
        return VENV_DIR / "Scripts" / "python.exe"
    return VENV_DIR / "bin" / "python"


def _in_target_venv() -> bool:
    # Belt-and-suspenders: compare both sys.executable and sys.prefix against
    # the target venv. Windows Store Python uses reparse points that can make
    # Path.resolve() on sys.executable return a path that differs from the
    # venv's python.exe even when running inside it; sys.prefix is more
    # reliable for that case. Either match counts as "in venv".
    try:
        target_py = _venv_python().resolve()
    except OSError:
        target_py = None
    try:
        target_dir = VENV_DIR.resolve()
    except OSError:
        target_dir = None
    try:
        if target_py is not None and Path(sys.executable).resolve() == target_py:
            return True
    except OSError:
        pass
    try:
        if target_dir is not None and Path(sys.prefix).resolve() == target_dir:
            return True
    except OSError:
        pass
    return False


def _bootstrap(force_reinstall: bool) -> None:
    if not VENV_DIR.exists():
        print(f"[bootstrap] Creating virtualenv at {VENV_DIR} ...")
        venv.EnvBuilder(with_pip=True, clear=False, upgrade_deps=False).create(VENV_DIR)

    py = _venv_python()
    needs_install = force_reinstall or not MARKER.exists()
    if needs_install:
        print("[bootstrap] Upgrading pip ...")
        subprocess.check_call([str(py), "-m", "pip", "install", "--upgrade", "pip"])
        print(f"[bootstrap] Installing: {', '.join(REQUIREMENTS)}")
        subprocess.check_call([str(py), "-m", "pip", "install", *REQUIREMENTS])
        MARKER.write_text("ok\n", encoding="utf-8")
    else:
        print("[bootstrap] Requirements already installed (delete .venv/.bootstrapped to redo).")

    args = [a for a in sys.argv[1:] if a != "--reinstall"]
    print("[bootstrap] Re-launching inside venv ...\n")
    sys.exit(subprocess.call([str(py), str(Path(__file__).resolve()), *args]))


# ---------------------------------------------------------------------------
# Portable-Python bootstrap (run.bat path when no system Python exists)
# ---------------------------------------------------------------------------

def _running_portable_python() -> bool:
    """True when sys.executable lives inside ROOT/.pyenv/."""
    try:
        return Path(sys.executable).resolve().parent == PYENV_DIR.resolve()
    except OSError:
        return False


def _bootstrap_portable(force_reinstall: bool) -> None:
    """Install pip + requirements directly into the embedded Python.

    Embeddable Python ships without pip and disables `import site` by default,
    so we fix both before pip-installing the rest. No re-exec is needed —
    we're already running inside the target interpreter.
    """
    py = Path(sys.executable)

    # 1) Enable site-packages by uncommenting `import site` in pythonXY._pth.
    pth_files = sorted(py.parent.glob("python*._pth"))
    for pth in pth_files:
        try:
            text = pth.read_text(encoding="utf-8")
        except OSError:
            continue
        new_text = text
        for needle in ("#import site", "# import site"):
            if needle in new_text:
                new_text = new_text.replace(needle, "import site")
        if new_text != text:
            try:
                pth.write_text(new_text, encoding="utf-8")
                print(f"[bootstrap] Enabled site-packages in {pth.name}.")
            except OSError as e:
                print(f"[bootstrap] Could not edit {pth}: {e}")

    # 2) Install pip if missing.
    has_pip = subprocess.call(
        [str(py), "-m", "pip", "--version"],
        stdout=subprocess.DEVNULL,
        stderr=subprocess.DEVNULL,
    ) == 0
    if not has_pip:
        print("[bootstrap] Installing pip into portable Python ...")
        import urllib.request
        getpip = py.parent / "_get-pip.py"
        try:
            urllib.request.urlretrieve("https://bootstrap.pypa.io/get-pip.py", getpip)
            subprocess.check_call([str(py), str(getpip), "--no-warn-script-location"])
        finally:
            try:
                getpip.unlink()
            except OSError:
                pass

    # 3) Install requirements (skipped if marker says we already did it).
    needs_install = force_reinstall or not PYENV_MARKER.exists()
    if needs_install:
        print("[bootstrap] Upgrading pip ...")
        subprocess.check_call([str(py), "-m", "pip", "install", "--upgrade", "pip"])
        print(f"[bootstrap] Installing: {', '.join(REQUIREMENTS)}")
        subprocess.check_call([str(py), "-m", "pip", "install", *REQUIREMENTS])
        PYENV_MARKER.write_text("ok\n", encoding="utf-8")
    else:
        print("[bootstrap] Requirements already installed (delete .pyenv/.bootstrapped to redo).")


# ---------------------------------------------------------------------------
# onnxruntime / Visual C++ runtime recovery
# ---------------------------------------------------------------------------

def _looks_like_vcredist_failure(err: BaseException) -> bool:
    """Heuristic: did onnxruntime fail to load because VC++ runtime DLLs
    aren't on the box? The exact wording varies slightly by Windows version
    but always mentions the failed DLL by name."""
    if os.name != "nt":
        return False
    msg = str(err)
    return "DLL load failed" in msg and "onnxruntime_pybind11_state" in msg


def _install_vcredist_with_uac() -> bool:
    """Download vc_redist.x64.exe and run it elevated. Returns True on
    apparent success (exit 0 or 3010 = success-pending-reboot)."""
    if os.name != "nt":
        return False
    import urllib.request
    dest = ROOT / ".vc_redist.x64.exe"
    print(f"[runtime] Downloading {VC_REDIST_URL} ...")
    try:
        urllib.request.urlretrieve(VC_REDIST_URL, dest)
    except Exception as e:  # noqa: BLE001
        print(f"[runtime] Download failed: {e}")
        return False

    # Start-Process -Verb RunAs is what triggers the UAC prompt; -Wait blocks
    # until the installer exits so we can read its exit code. Apostrophes in
    # the path are escaped by doubling for PowerShell's single-quote syntax.
    ps_dest = str(dest).replace("'", "''")
    ps_cmd = (
        "$ErrorActionPreference='Stop';"
        f"$p = Start-Process -FilePath '{ps_dest}' "
        "-ArgumentList '/install','/quiet','/norestart' "
        "-Verb RunAs -Wait -PassThru;"
        "exit $p.ExitCode"
    )
    print("[runtime] Launching installer — accept the UAC prompt to continue.")
    try:
        rc = subprocess.call(
            ["powershell", "-NoProfile", "-ExecutionPolicy", "Bypass", "-Command", ps_cmd]
        )
    except OSError as e:
        print(f"[runtime] Could not launch installer: {e}")
        rc = -1
    finally:
        try:
            dest.unlink()
        except OSError:
            pass

    if rc in (0, 3010):
        if rc == 3010:
            print("[runtime] Install succeeded but Windows requested a reboot.")
        return True
    if rc == 1602 or rc == 1223:
        print("[runtime] Install was cancelled at the UAC prompt.")
    else:
        print(f"[runtime] Installer exited with code {rc}.")
    return False


def _recover_and_relaunch_after_vcredist() -> None:
    """Try to install the VC++ redistributable, then re-exec the launcher so
    the onnxruntime DLL gets a fresh load attempt. Windows caches failed DLL
    loads within a process, so re-importing in-place doesn't work."""
    print()
    print("=" * 60)
    print("  onnxruntime failed to load")
    print("=" * 60)
    print("This almost always means the Microsoft Visual C++ 2015-2022")
    print("Redistributable (x64) is missing on this Windows machine.")
    print("Trying to install it for you now — you'll see a UAC prompt.")
    print()

    if not _install_vcredist_with_uac():
        print()
        print("Manual fix:")
        print(f"  1. Download {VC_REDIST_URL}")
        print("  2. Run it (accept the UAC prompt)")
        print("  3. Re-run this script")
        print()
        print("If install keeps failing, your CPU may lack AVX2 support, which")
        print("modern onnxruntime wheels require. Most CPUs from 2014+ have it.")
        sys.exit(1)

    args = [a for a in sys.argv[1:] if a != "--reinstall"]
    print("\n[runtime] Visual C++ runtime installed. Re-launching the app ...\n")
    rc = subprocess.call([sys.executable, str(Path(__file__).resolve()), *args])
    sys.exit(rc)


# ---------------------------------------------------------------------------
# App
# ---------------------------------------------------------------------------

REQUIRED_FILES = ("model.onnx", "selected_tags.csv", "preprocessing.json")


def _discover_model_dirs() -> list[Path]:
    """Return every subdirectory of ROOT that looks like a usable model folder."""
    out: list[Path] = []
    if not ROOT.exists():
        return out
    for sub in sorted(ROOT.iterdir(), key=lambda p: p.name.lower()):
        if not sub.is_dir():
            continue
        if all((sub / f).exists() for f in REQUIRED_FILES):
            out.append(sub)
    return out


def _variant_rank(name: str) -> int:
    try:
        return HF_VARIANTS.index(name)
    except ValueError:
        return len(HF_VARIANTS)


def _is_tty() -> bool:
    try:
        return sys.stdin.isatty() and sys.stdout.isatty()
    except (AttributeError, OSError):
        return False


def _prompt_choice(prompt: str, options: list[tuple[str, str]], default_idx: int = 0) -> str | None:
    """Show a numbered terminal menu. Returns the chosen option's value, or None on EOF.

    options: list of (display_text, value).
    """
    if not options:
        return None
    if not _is_tty():
        return options[default_idx][1]

    print()
    print(prompt)
    for i, (display, _) in enumerate(options, 1):
        marker = "  <- press Enter for this" if i - 1 == default_idx else ""
        print(f"  {i}) {display}{marker}")
    while True:
        try:
            raw = input(f"Choice [1-{len(options)}, default {default_idx + 1}]: ").strip()
        except EOFError:
            return options[default_idx][1]
        if not raw:
            return options[default_idx][1]
        try:
            idx = int(raw) - 1
        except ValueError:
            print(f"  Please enter a number 1-{len(options)}.")
            continue
        if 0 <= idx < len(options):
            return options[idx][1]
        print(f"  Out of range. Pick 1-{len(options)}.")


def _download_variant(variant: str) -> Path | None:
    """Download a HuggingFace variant into ROOT/<variant>. Returns the folder on success."""
    try:
        from huggingface_hub import snapshot_download
    except ImportError:
        print("[app] huggingface_hub is not installed — re-run with --reinstall.")
        return None

    print(f"[app] Downloading '{variant}' from huggingface.co/{HF_REPO_ID} ...")
    try:
        snapshot_download(
            repo_id=HF_REPO_ID,
            allow_patterns=[f"{variant}/*"],
            local_dir=str(ROOT),
        )
    except Exception as e:  # noqa: BLE001
        print(f"[app] Download failed: {e}")
        return None

    target = ROOT / variant
    missing = [f for f in REQUIRED_FILES if not (target / f).exists()]
    if missing:
        print(f"[app] Download finished but {target} is missing: {', '.join(missing)}")
        return None
    return target


def _interactive_pick_model() -> Path | None:
    """Show a friendly menu so non-technical users can pick or download a model.

    Returns the chosen model directory, or None if the user wants to start the
    UI without loading anything (they can pick from the web UI then).
    """
    discovered = _discover_model_dirs()
    discovered.sort(key=lambda p: (_variant_rank(p.name), p.name.lower()))
    discovered_names = {p.name for p in discovered}

    options: list[tuple[str, str]] = []
    actions: list[tuple[str, str]] = []  # parallel list of (action, payload)

    for p in discovered:
        desc = HF_VARIANT_DESC.get(p.name, "model folder")
        options.append((f"Use {p.name}  ({desc})", str(p)))
        actions.append(("load", str(p)))

    for v in HF_VARIANTS:
        if v in discovered_names:
            continue
        desc = HF_VARIANT_DESC.get(v, "")
        suffix = f"  ({desc})" if desc else ""
        options.append((f"Download {v} from HuggingFace{suffix}", v))
        actions.append(("download", v))

    options.append(("Open the web UI without loading anything (pick later from the page)", "skip"))
    actions.append(("skip", ""))

    if not _is_tty() and discovered:
        return discovered[0]
    if not _is_tty():
        return None

    print()
    print("=" * 50)
    print("  Oppai ONNX Tagger")
    print("=" * 50)
    if discovered:
        print(f"Found {len(discovered)} model folder(s) next to app.py.")
    else:
        print("No model folders found yet next to app.py.")
        print(f"Pick a variant to download from huggingface.co/{HF_REPO_ID}.")

    chosen = _prompt_choice("What would you like to do?", options, default_idx=0)
    if chosen is None:
        return None

    idx = next(i for i, (_, v) in enumerate(options) if v == chosen)
    action, payload = actions[idx]
    if action == "load":
        return Path(payload)
    if action == "download":
        return _download_variant(payload)
    return None  # skip


def _resolve_initial_model(cli_dir: str | None) -> Path | None:
    if cli_dir:
        p = Path(cli_dir).expanduser().resolve()
        if not p.is_dir():
            print(f"[app] --model-dir not a directory: {p}")
            return None
        missing = [f for f in REQUIRED_FILES if not (p / f).exists()]
        if missing:
            print(f"[app] --model-dir is missing required files: {', '.join(missing)}")
            return None
        return p
    return _interactive_pick_model()


def _run_app() -> None:
    import argparse
    import csv
    import json

    import numpy as np
    try:
        import onnxruntime as ort
    except ImportError as e:
        if _looks_like_vcredist_failure(e):
            _recover_and_relaunch_after_vcredist()
            sys.exit(1)  # only reached if recover failed without exiting
        raise
    import gradio as gr
    from PIL import Image

    parser = argparse.ArgumentParser(add_help=False)
    parser.add_argument("--model-dir", type=str, default=None)
    cli_args, _ = parser.parse_known_args()

    cat_names = {0: "general", 1: "artist", 3: "copyright", 4: "character", 5: "meta"}
    inv_cat_names = {v: k for k, v in cat_names.items()}

    # Mutable holder so the UI can swap models without restarting the process.
    state: dict = {
        "session": None,
        "tag_names": [],
        "categories": [],
        "skip_mask": None,
        "image_size": 0,
        "pad_color": (0, 0, 0),
        "mean": None,
        "std": None,
        "breakeven_threshold": None,
        "model_dir": None,
        "providers": [],
    }

    def _ort_providers() -> list[str]:
        available = ort.get_available_providers()
        if "DmlExecutionProvider" in available:
            return ["DmlExecutionProvider", "CPUExecutionProvider"]
        if "CUDAExecutionProvider" in available:
            return ["CUDAExecutionProvider", "CPUExecutionProvider"]
        return ["CPUExecutionProvider"]

    def load_model(model_dir: Path) -> str:
        model_dir = Path(model_dir).expanduser().resolve()
        if not model_dir.is_dir():
            raise FileNotFoundError(f"not a directory: {model_dir}")
        missing = [f for f in REQUIRED_FILES if not (model_dir / f).exists()]
        if missing:
            raise FileNotFoundError(
                f"{model_dir} is missing required files: {', '.join(missing)}"
            )

        tag_names: list[str] = []
        categories: list[int] = []
        with (model_dir / "selected_tags.csv").open(encoding="utf-8") as f:
            for row in csv.DictReader(f):
                tag_names.append(row["name"])
                categories.append(int(row["category"]))
        n_tags = len(tag_names)

        skip_mask = np.zeros(n_tags, dtype=bool)
        for i, name in enumerate(tag_names):
            if name in ("<PAD>", "<UNK>"):
                skip_mask[i] = True

        with (model_dir / "preprocessing.json").open(encoding="utf-8") as f:
            preproc = json.load(f)
        image_size = int(preproc["image_size"])
        pad_color = tuple(int(c) for c in preproc["pad_color_rgb"])
        mean = np.array(preproc["normalize_mean"], dtype=np.float32).reshape(3, 1, 1)
        std = np.array(preproc["normalize_std"], dtype=np.float32).reshape(3, 1, 1)

        # Calibrated breakeven (precision = recall) lives in pr_thresholds.json.
        # It is tuned for whole-eval-set precision and is far too strict for
        # interactive single-image tagging, so we surface it only as a hint.
        breakeven_threshold = None
        thr_path = model_dir / "pr_thresholds.json"
        if thr_path.exists():
            try:
                with thr_path.open(encoding="utf-8") as f:
                    thr_data = json.load(f)
                breakeven_threshold = float(thr_data["micro"]["pr_breakeven"]["threshold"])
            except (OSError, KeyError, ValueError, json.JSONDecodeError):
                pass

        providers = _ort_providers()
        sess_opts = ort.SessionOptions()
        sess_opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
        print(f"[app] Loading {model_dir / 'model.onnx'} ({image_size}×{image_size}) ...")
        print(f"[app] Providers: {providers}")
        session = ort.InferenceSession(
            str(model_dir / "model.onnx"), sess_options=sess_opts, providers=providers
        )

        state.update(
            session=session,
            tag_names=tag_names,
            categories=categories,
            skip_mask=skip_mask,
            image_size=image_size,
            pad_color=pad_color,
            mean=mean,
            std=std,
            breakeven_threshold=breakeven_threshold,
            model_dir=model_dir,
            providers=providers,
        )
        return _status_md()

    def _status_md() -> str:
        if state["session"] is None:
            return (
                "**No model loaded.** Drop an ONNX model folder next to "
                "`app.py`, or use the **Download from HuggingFace** section below."
            )
        try:
            display = state["model_dir"].relative_to(ROOT)
        except ValueError:
            display = state["model_dir"]
        parts = [
            f"**Loaded:** `{display}`",
            f"{state['image_size']}×{state['image_size']}",
            f"{len(state['tag_names'])} tags",
            f"providers: {', '.join(state['providers'])}",
        ]
        if state["breakeven_threshold"] is not None:
            parts.append(f"P=R breakeven: {state['breakeven_threshold']:.3f}")
        return " — ".join(parts)

    def _dropdown_choices() -> list[tuple[str, str]]:
        out = []
        for p in _discover_model_dirs():
            try:
                label = str(p.relative_to(ROOT))
            except ValueError:
                label = p.name
            out.append((label, str(p)))
        return out

    def _current_value() -> str | None:
        return str(state["model_dir"]) if state["model_dir"] else None

    # Initial load (CLI override > default folder > first discovered)
    initial = _resolve_initial_model(cli_args.model_dir)
    if initial is not None:
        try:
            load_model(initial)
        except Exception as e:  # noqa: BLE001
            print(f"[app] Initial model load failed: {e!r}")
    else:
        print("[app] No model folder found yet — pick or download one in the UI.")

    def letterbox(img: Image.Image):
        img = img.convert("RGB")
        w, h = img.size
        size = state["image_size"]
        scale = min(size / w, size / h)
        nw, nh = max(1, int(round(w * scale))), max(1, int(round(h * scale)))
        resized = img.resize((nw, nh), Image.BICUBIC)
        canvas = Image.new("RGB", (size, size), state["pad_color"])
        x0 = (size - nw) // 2
        y0 = (size - nh) // 2
        canvas.paste(resized, (x0, y0))
        mask = np.ones((size, size), dtype=bool)  # True = padded
        mask[y0:y0 + nh, x0:x0 + nw] = False
        return canvas, mask

    def preprocess(img: Image.Image):
        canvas, mask = letterbox(img)
        arr = np.asarray(canvas, dtype=np.float32) / 255.0
        arr = arr.transpose(2, 0, 1)  # CHW
        arr = (arr - state["mean"]) / state["std"]
        return arr.astype(np.float32), mask

    def predict(image, threshold: float, max_tags, category_filter):
        if state["session"] is None:
            return "", "*no model loaded — pick or download one above*"
        if image is None:
            return "", "*upload an image to start*"
        try:
            max_tags_i = int(max_tags) if max_tags is not None else 0
            if max_tags_i <= 0:
                return "", "*no tags above threshold*"

            # An empty list means "no categories selected" -> show nothing.
            # `None` (event before component initialized) means "no filter".
            if category_filter is None:
                keep_cats = None
            else:
                keep_cats = {inv_cat_names[c] for c in category_filter if c in inv_cat_names}
                if not keep_cats:
                    return "", "*no tags above threshold*"

            pixel_values, padding_mask = preprocess(image)
            outputs = state["session"].run(
                ["probabilities"],
                {
                    "pixel_values": pixel_values[None, ...],
                    "padding_mask": padding_mask[None, ...],
                },
            )
            probs = outputs[0][0].astype(np.float32)
            probs[state["skip_mask"]] = -1.0  # never surface PAD/UNK

            order = np.argsort(-probs)
            results = []
            tag_names = state["tag_names"]
            categories = state["categories"]
            for idx in order:
                p = float(probs[idx])
                if p < threshold:
                    break
                cat = categories[idx]
                if keep_cats is not None and cat not in keep_cats:
                    continue
                results.append((tag_names[idx], p, cat))
                if len(results) >= max_tags_i:
                    break

            if not results:
                return "", "*no tags above threshold*"

            comma = ", ".join(name.replace("_", " ") for name, _, _ in results)
            lines = ["| # | Tag | Confidence | Category |", "|---|---|---|---|"]
            for i, (name, p, cat) in enumerate(results, 1):
                lines.append(f"| {i} | `{name}` | {p:.3f} | {cat_names.get(cat, str(cat))} |")
            return comma, "\n".join(lines)
        except Exception as e:  # noqa: BLE001 — keep Gradio toast away
            print(f"[app] predict() error: {e!r}")
            return "", f"*error during inference: {e}*"

    # --- UI callbacks ------------------------------------------------------

    def on_refresh():
        choices = _dropdown_choices()
        return gr.update(choices=choices, value=_current_value()), _status_md()

    def on_load(dropdown_value: str | None, custom_path: str):
        target = (custom_path or "").strip() or dropdown_value
        if not target:
            return gr.update(), _status_md(), "Pick a model folder or paste a path first."
        try:
            load_model(Path(target))
        except Exception as e:  # noqa: BLE001
            return gr.update(), _status_md(), f"Load failed: {e}"
        choices = _dropdown_choices()
        return (
            gr.update(choices=choices, value=_current_value()),
            _status_md(),
            f"Loaded `{Path(target).name}`.",
        )

    def on_download(variant: str, progress=gr.Progress(track_tqdm=True)):
        if not variant:
            return gr.update(), _status_md(), "Pick a variant first."
        try:
            from huggingface_hub import snapshot_download
        except ImportError:
            return (
                gr.update(),
                _status_md(),
                "huggingface_hub is not installed — re-run `app.py --reinstall`.",
            )
        progress(0, desc=f"Downloading {variant} from {HF_REPO_ID} ...")
        try:
            snapshot_download(
                repo_id=HF_REPO_ID,
                allow_patterns=[f"{variant}/*"],
                local_dir=str(ROOT),
            )
        except Exception as e:  # noqa: BLE001
            return gr.update(), _status_md(), f"Download failed: {e}"

        target = ROOT / variant
        msg = f"Downloaded `{variant}`."
        if all((target / f).exists() for f in REQUIRED_FILES):
            try:
                load_model(target)
                msg += f" Loaded `{variant}`."
            except Exception as e:  # noqa: BLE001
                msg += f" Load failed: {e}"
        choices = _dropdown_choices()
        return gr.update(choices=choices, value=_current_value()), _status_md(), msg

    # --- UI layout ---------------------------------------------------------

    with gr.Blocks(title="Oppai ONNX Tagger") as demo:
        gr.Markdown(
            "# Oppai ONNX Tagger\n"
            "Upload an image and tweak the threshold / max tags. "
            "Pick a model below or download one from "
            "[Grio43/OppaiOracle](https://huggingface.co/Grio43/OppaiOracle)."
        )

        with gr.Accordion("Model", open=True):
            with gr.Row():
                model_dd = gr.Dropdown(
                    choices=_dropdown_choices(),
                    value=_current_value(),
                    label="Detected model folders (next to app.py)",
                    interactive=True,
                    scale=3,
                )
                refresh_btn = gr.Button("Refresh", scale=1)
            with gr.Row():
                custom_path = gr.Textbox(
                    label="…or paste a custom model folder path (overrides dropdown)",
                    placeholder=r"e.g. C:\models\my_onnx_folder",
                    scale=4,
                )
                load_btn = gr.Button("Load", variant="primary", scale=1)
            with gr.Row():
                hf_dd = gr.Dropdown(
                    choices=HF_VARIANTS,
                    value=HF_VARIANTS[0],
                    label=f"Download a variant from {HF_REPO_ID}",
                    scale=3,
                )
                download_btn = gr.Button("Download", scale=1)
            status_md = gr.Markdown(_status_md())
            action_msg = gr.Markdown("")

        with gr.Row():
            with gr.Column(scale=1):
                inp = gr.Image(type="pil", label="Image", height=448)
                threshold = gr.Slider(
                    0.0, 1.0,
                    value=0.35,
                    step=0.005,
                    label="Threshold (interactive default 0.35; calibrated breakeven shown above)",
                )
                max_tags = gr.Slider(1, 200, value=50, step=1, label="Max tags")
                cats = gr.CheckboxGroup(
                    choices=list(cat_names.values()),
                    value=list(cat_names.values()),
                    label="Categories to include",
                )
                btn = gr.Button("Tag image", variant="primary")
            with gr.Column(scale=1):
                tags_out = gr.Textbox(
                    label="Tags (comma-separated, underscores → spaces)",
                    lines=5,
                )
                table_out = gr.Markdown(label="Per-tag detail")

        refresh_btn.click(on_refresh, outputs=[model_dd, status_md])
        load_btn.click(on_load, inputs=[model_dd, custom_path], outputs=[model_dd, status_md, action_msg])
        download_btn.click(on_download, inputs=[hf_dd], outputs=[model_dd, status_md, action_msg])

        ev_inputs = [inp, threshold, max_tags, cats]
        ev_outputs = [tags_out, table_out]
        btn.click(predict, ev_inputs, ev_outputs)
        inp.change(predict, ev_inputs, ev_outputs)
        threshold.release(predict, ev_inputs, ev_outputs)
        max_tags.release(predict, ev_inputs, ev_outputs)
        cats.change(predict, ev_inputs, ev_outputs)

    # CPU inference is ~1-3s per image; cap concurrency so spammed slider
    # changes queue serially instead of fighting for the same model session.
    demo.queue(default_concurrency_limit=1).launch(
        server_name="127.0.0.1", server_port=7860, inbrowser=True
    )


# ---------------------------------------------------------------------------
# Entrypoint
# ---------------------------------------------------------------------------

def main() -> None:
    # On Windows, the console codepage is often cp1252/cp932/etc., not UTF-8.
    # Our messages contain em-dashes and × — `errors="replace"` keeps them from
    # crashing the bootstrap with UnicodeEncodeError on those consoles.
    for stream in (sys.stdout, sys.stderr):
        try:
            stream.reconfigure(errors="replace")
        except (AttributeError, OSError):
            pass

    force = "--reinstall" in sys.argv[1:]

    if _running_portable_python():
        # run.bat downloaded an embeddable Python because the user had none.
        # Install requirements directly into it; no venv re-exec needed.
        _bootstrap_portable(force_reinstall=force)
        _run_app()
        return

    if not _in_target_venv():
        _bootstrap(force_reinstall=force)
        return  # _bootstrap re-execs and exits
    _run_app()


if __name__ == "__main__":
    main()