File size: 35,833 Bytes
fcf5483 | 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 | # Deep Shrink Hires.fix (RU++ v2.3.1 LTS + Experimental, FIXED)
# Исправления: статические _block_scale/_curve_weight; синхронизация instance→class после process().
from dataclasses import dataclass
from typing import List, Optional, Dict, Any, Tuple
import json, os
import gradio as gr
import torch
import torch.nn.functional as F
import modules.devices as devices
import modules.scripts as scripts
import modules.script_callbacks as script_callbacks
import modules.sd_unet as sd_unet
import modules.shared as shared
from ldm.modules.diffusionmodules.openaimodel import Upsample, Downsample, ResBlock
from ldm.modules.diffusionmodules.util import timestep_embedding
# -------------------------- Утилиты --------------------------
def _to_scalar(x) -> float:
if isinstance(x, torch.Tensor):
return float(x.item())
try:
return float(x)
except Exception:
return 0.0
def _clamp(v: float, lo: float, hi: float) -> float:
return max(lo, min(hi, v))
def _get_or_last(seq: List[float], index: int, default: float) -> float:
if not seq:
return default
return seq[index] if index < len(seq) else seq[-1]
def _safe_size(h: torch.Tensor, scale_factor: float, min_feat: int) -> Tuple[int, int]:
hi, wi = h.shape[-2], h.shape[-1]
ho = max(2, int(round(hi * scale_factor)))
wo = max(2, int(round(wi * scale_factor)))
if ho < min_feat or wo < min_feat:
return hi, wi
return ho, wo
def _interpolate(img: torch.Tensor, size: Tuple[int, int], mode: str, antialias: bool) -> torch.Tensor:
if size == img.shape[-2:]:
return img
dtype = img.dtype
try:
out = F.interpolate(
img.float(),
size=size,
mode=mode,
align_corners=False if mode in ("bilinear", "bicubic") else None,
antialias=(antialias if mode in ("bilinear", "bicubic") else False),
)
except TypeError:
out = F.interpolate(
img.float(),
size=size,
mode=mode,
align_corners=False if mode in ("bilinear", "bicubic") else None,
)
return out.to(dtype)
def _resize(h: torch.Tensor, scale: float, mode: str, antialias: bool, min_feat: int) -> torch.Tensor:
if scale == 1.0:
return h
size = _safe_size(h, scale, min_feat)
return _interpolate(h, size, mode, antialias)
def _parse_number_list(text: str, as_int: bool = False) -> List[float]:
text = (text or "").replace("\n", " ")
vals: List[float] = []
for chunk in text.split(";"):
s = chunk.strip()
if not s:
continue
if "/" in s:
a, b = s.split("/", 1)
v = float(a.strip()) / float(b.strip())
else:
v = float(s)
vals.append(v)
if not vals:
raise ValueError("Список пуст.")
return [int(round(v)) for v in vals] if as_int else vals
# -------------------------- Данные --------------------------
@dataclass
class DSHFAction:
enable: bool
timestep: float
depth: int
scale: float
# -------------------------- Скрипт --------------------------
class DSHF(scripts.Script):
# Счётчики текущего прохода (класс-поля нужны U-Net'у)
currentBlock: int = 0
currentConv: int = 0
currentTimestep: float = 1000.0
# Глобальные параметры (класс-поля: U-Net читает их как DSHF.*)
enabled: bool = True
interp_method: str = "bicubic"
interp_antialias: bool = True
channels_last: bool = False
min_feature_size: int = 8
# Кривая
curve_enable: bool = False
curve_type: str = "linear"
curve_t_start: float = 800.0
curve_t_end: float = 200.0
curve_scale_start: float = 1.0
curve_scale_end: float = 1.0
curve_alpha: float = 0.5
auto_end_enable: bool = False
auto_end_strength: float = 0.35
# Пороговые правила
actions: List[DSHFAction] = []
# Experimental (класс-поля для доступа из U-Net)
exp_section_enable: bool = False
exp_enable: bool = False
exp_timestep: float = 625.0
exp_scales: List[float] = [1.0]
exp_dilations: List[int] = [1]
exp_in_muls: List[float] = [1.0]
exp_out_muls: List[float] = [1.0]
exp_cfg_muls: List[float] = [1.0]
# ---------------- UI ----------------
def title(self): return "Deep Shrink Hires.fix (RU++ v2.3.1)"
def show(self, is_img2img): return scripts.AlwaysVisible
def ui(self, is_img2img):
with gr.Tabs():
with gr.TabItem("Настройки"):
Enable_Ext = gr.Checkbox(value=True, label="Включить расширение")
with gr.Accordion("Пороги (до 8 правил)", open=True):
Rule_Count = gr.Slider(1, 8, value=2, step=1, label="Сколько правил использовать")
rules = []
defaults = [
(True, 625, 3, 2.0),
(True, 0, 3, 2.0),
(False, 900, 3, 2.0),
(False, 650, 3, 2.0),
(False, 900, 3, 2.0),
(False, 650, 3, 2.0),
(False, 900, 3, 2.0),
(False, 650, 3, 2.0),
]
for i, (en, ts, dp, sc) in enumerate(defaults, start=1):
with gr.Row():
rules.append((
gr.Checkbox(value=en, label=f"Правило {i}"),
gr.Number(value=ts, label=f"Timestep {i}"),
gr.Number(value=dp, label=f"Глубина блока {i}", precision=0),
gr.Number(value=sc, label=f"Масштаб {i}"),
))
with gr.Accordion("Глобальная кривая масштаба", open=False):
Curve_Enable = gr.Checkbox(value=False, label="Включить кривую")
Curve_Type = gr.Dropdown(choices=["linear", "cosine", "sigmoid"], value="linear", label="Тип")
with gr.Row():
Curve_t_start = gr.Number(value=800, label="t_start")
Curve_t_end = gr.Number(value=200, label="t_end")
with gr.Row():
Curve_scale_start = gr.Number(value=1.0, label="scale_start")
Curve_scale_end = gr.Number(value=1.0, label="scale_end")
Curve_alpha = gr.Slider(0.0, 1.0, value=0.5, step=0.05, label="alpha компенсации")
Min_feature = gr.Slider(2, 64, value=8, step=1, label="Мин. размер фичей")
with gr.Row():
Auto_end_enable = gr.Checkbox(value=False, label="Автоподбор scale_end по целевому разрешению")
Auto_end_strength = gr.Slider(0.0, 1.0, value=0.35, step=0.05, label="Сила автоподбора")
with gr.Accordion("Выполнение", open=False):
Interp_Method = gr.Dropdown(choices=["nearest", "bilinear", "bicubic", "area"], value="bicubic", label="Интерполяция")
Interp_AA = gr.Checkbox(value=True, label="Антиалиасинг для bilinear/bicubic")
Channels_Last = gr.Checkbox(value=False, label="Оптимизация channels_last")
with gr.Accordion("Экспериментальные (пер-свёрточные)", open=False):
Exp_Section_Enable = gr.Checkbox(value=False, label="Включить секцию")
with gr.Group(visible=False) as ExpGrp:
Exp_Enable = gr.Checkbox(value=False, label="Активировать экспериментальное ядро")
Exp_Timestep = gr.Number(value=625, label="Пороговой timestep для эксперимента")
Exp_Scales = gr.Textbox(value="1", lines=2, label="Масштабы по свёрткам")
Exp_Dilations = gr.Textbox(value="1", lines=1, label="Дилатации по свёрткам (целые)")
Exp_InMuls = gr.Textbox(value="1", lines=1, label="Входные умножители по блокам")
Exp_OutMuls = gr.Textbox(value="1", lines=1, label="Выходные умножители по блокам")
Exp_CFGMuls = gr.Textbox(value="1", lines=1, label="CFG-множители по блокам")
Exp_Section_Enable.change(lambda v: gr.update(visible=bool(v)), Exp_Section_Enable, ExpGrp)
with gr.Accordion("Пресеты", open=False):
Preset_Action = gr.Dropdown(choices=["Сохранить", "Загрузить"], value="Загрузить", label="Действие")
Preset_Name = gr.Textbox(value="", label="Имя пресета")
Preset_JSON = gr.Textbox(value="", lines=6, label="Профиль в JSON (для импорта/экспорта)")
with gr.TabItem("Справка"):
gr.Markdown("""
**LTS** — безопасное масштабирование только на границах блоков.
**Experimental** — опционально: пер-свёрточные масштабы, дилатации, In/Out и CFG-мультипликаторы.
""")
flat = [Enable_Ext, Rule_Count]
for row in rules:
flat += list(row)
flat += [
Curve_Enable, Curve_Type, Curve_t_start, Curve_t_end,
Curve_scale_start, Curve_scale_end, Curve_alpha, Min_feature,
Auto_end_enable, Auto_end_strength,
Interp_Method, Interp_AA, Channels_Last,
Exp_Section_Enable, Exp_Enable, Exp_Timestep,
Exp_Scales, Exp_Dilations, Exp_InMuls, Exp_OutMuls, Exp_CFGMuls,
Preset_Action, Preset_Name, Preset_JSON,
]
return flat
# ---------------- Исполнение ----------------
def _reset_instance_defaults(self):
# только для локальной сборки входных значений; класс-поля перезапишем ниже
self._inst_actions: List[DSHFAction] = []
self._inst_curve_enable = False
self._inst_curve_type = "linear"
self._inst_curve_t_start = 800.0
self._inst_curve_t_end = 200.0
self._inst_curve_scale_start = 1.0
self._inst_curve_scale_end = 1.0
self._inst_curve_alpha = 0.5
self._inst_auto_end_enable = False
self._inst_auto_end_strength = 0.35
self._inst_interp_method = "bicubic"
self._inst_interp_antialias = True
self._inst_channels_last = False
self._inst_min_feature_size = 8
self._inst_exp_section_enable = False
self._inst_exp_enable = False
self._inst_exp_timestep = 625.0
self._inst_exp_scales = [1.0]
self._inst_exp_dilations = [1]
self._inst_exp_in_muls = [1.0]
self._inst_exp_out_muls = [1.0]
self._inst_exp_cfg_muls = [1.0]
@staticmethod
def _curve_weight(ts: float) -> Optional[float]:
if not DSHF.curve_enable:
return None
t0, t1 = DSHF.curve_t_start, DSHF.curve_t_end
if t0 == t1:
return DSHF.curve_scale_end
x = _clamp((ts - t1) / (t0 - t1), 0.0, 1.0)
if DSHF.curve_type == "linear":
w = x
elif DSHF.curve_type == "cosine":
w = 0.5 - 0.5 * torch.cos(torch.tensor(x) * torch.pi).item()
else:
w = 1.0 / (1.0 + torch.exp(torch.tensor(-10.0 * (x - 0.5)))).item()
s = DSHF.curve_scale_start + (DSHF.curve_scale_end - DSHF.curve_scale_start) * w
return _clamp(float(s), 0.25, 4.0)
@staticmethod
def _block_scale(depth: int, ts: float) -> Optional[float]:
rule_scale: Optional[float] = None
for a in DSHF.actions:
if a.enable and a.depth == depth and a.timestep <= ts:
rule_scale = a.scale
break
curve_scale = DSHF._curve_weight(ts)
if rule_scale is None and curve_scale is None:
return None
if rule_scale is None:
return curve_scale
if curve_scale is None:
return rule_scale
return _clamp(rule_scale * curve_scale, 0.25, 4.0)
def _auto_scale_end(self, p) -> Optional[float]:
if not self._inst_auto_end_enable or not self._inst_curve_enable:
return None
try:
bw, bh = int(getattr(p, "width", 0)), int(getattr(p, "height", 0))
if bw <= 0 or bh <= 0:
return None
tw, th = bw, bh
if getattr(p, "enable_hr", False):
hrx = int(getattr(p, "hr_resize_x", 0))
hry = int(getattr(p, "hr_resize_y", 0))
hrs = float(getattr(p, "hr_scale", 0.0) or 0.0)
if hrx > 0 and hry > 0:
tw, th = hrx, hry
elif hrs > 0.0:
tw, th = int(round(bw * hrs)), int(round(bh * hrs))
r = ((max(1, tw * th)) / max(1, bw * bh)) ** 0.5
if r <= 1.0:
return None
return _clamp(1.0 + _clamp(self._inst_auto_end_strength, 0.0, 1.0) * (r - 1.0), 1.0, 1.7)
except Exception:
return None
def process(self, p, *args):
# Используем только с нашим UNet
if not isinstance(sd_unet.current_unet, DSHF.DeepShrinkHiresFixUNet):
return
it = iter(args)
def nxt(): return next(it)
# Глобальный тумблер
enabled = bool(nxt())
if not enabled:
DSHF.enabled = False
return
# Сбор значений в instance-поле, чтобы не держать мусор в классе
self._reset_instance_defaults()
# ---- Правила ----
rule_count = int(_clamp(_to_scalar(nxt()), 1, 8))
tmp_rules: List[DSHFAction] = []
for _ in range(8):
en = bool(nxt()); ts = _to_scalar(nxt()); dp = int(_to_scalar(nxt())); sc = float(_to_scalar(nxt()))
tmp_rules.append(DSHFAction(en, ts, dp, sc))
self._inst_actions = tmp_rules[:rule_count]
# ---- Кривая ----
self._inst_curve_enable = bool(nxt())
self._inst_curve_type = str(nxt())
self._inst_curve_t_start = _to_scalar(nxt())
self._inst_curve_t_end = _to_scalar(nxt())
self._inst_curve_scale_start = float(_to_scalar(nxt()))
self._inst_curve_scale_end = float(_to_scalar(nxt()))
self._inst_curve_alpha = float(_clamp(_to_scalar(nxt()), 0.0, 1.0))
self._inst_min_feature_size = int(_clamp(_to_scalar(nxt()), 2, 256))
self._inst_auto_end_enable = bool(nxt())
self._inst_auto_end_strength = float(_clamp(_to_scalar(nxt()), 0.0, 1.0))
# ---- Выполнение ----
self._inst_interp_method = str(nxt())
self._inst_interp_antialias = bool(nxt())
self._inst_channels_last = bool(nxt())
# ---- Experimental ----
self._inst_exp_section_enable = bool(nxt())
self._inst_exp_enable = bool(nxt())
self._inst_exp_timestep = _to_scalar(nxt())
try: self._inst_exp_scales = list(map(float, _parse_number_list(str(nxt()), as_int=False)))
except Exception: self._inst_exp_scales = [1.0]
try: self._inst_exp_dilations = list(map(int, _parse_number_list(str(nxt()), as_int=True)))
except Exception: self._inst_exp_dilations = [1]
try: self._inst_exp_in_muls = list(map(float, _parse_number_list(str(nxt()), as_int=False)))
except Exception: self._inst_exp_in_muls = [1.0]
try: self._inst_exp_out_muls = list(map(float, _parse_number_list(str(nxt()), as_int=False)))
except Exception: self._inst_exp_out_muls = [1.0]
try: self._inst_exp_cfg_muls = list(map(float, _parse_number_list(str(nxt()), as_int=False)))
except Exception: self._inst_exp_cfg_muls = [1.0]
# ---- Пресеты ----
preset_action = str(nxt() or "")
preset_name = str(nxt() or "").strip()
preset_json = str(nxt() or "").strip()
preset_path = os.path.join(os.path.dirname(__file__), "dshf_presets.json")
if preset_action == "Сохранить" and preset_name:
data = self._export_profile_instance()
try:
cur = {"version": 1, "presets": {}}
if os.path.exists(preset_path):
with open(preset_path, "r", encoding="utf-8") as f:
cur = json.load(f)
cur["presets"][preset_name] = data
with open(preset_path, "w", encoding="utf-8") as f:
json.dump(cur, f, ensure_ascii=False, indent=2)
print(f"[DSHF] Пресет сохранён: {preset_name}")
except Exception as e:
print(f"[DSHF] Не удалось сохранить пресет: {e}")
elif preset_action == "Загрузить":
if preset_json:
try:
prof = json.loads(preset_json)
self._apply_profile_instance(prof)
print("[DSHF] Профиль применён из JSON")
except Exception as e:
print(f"[DSHF] Ошибка JSON: {e}")
elif preset_name:
try:
with open(preset_path, "r", encoding="utf-8") as f:
cur = json.load(f)
prof = cur.get("presets", {}).get(preset_name)
if prof:
self._apply_profile_instance(prof)
print(f"[DSHF] Профиль загружен: {preset_name}")
except Exception as e:
print(f"[DSHF] Не удалось загрузить пресет: {e}")
# Автоподбор конца кривой
auto = self._auto_scale_end(p)
if auto is not None:
self._inst_curve_scale_end = float(auto)
# -------- СИНХРОНИЗАЦИЯ instance → class (то, что читает U-Net) --------
DSHF.enabled = True
DSHF.actions = list(self._inst_actions)
DSHF.curve_enable = bool(self._inst_curve_enable)
DSHF.curve_type = str(self._inst_curve_type)
DSHF.curve_t_start = float(self._inst_curve_t_start)
DSHF.curve_t_end = float(self._inst_curve_t_end)
DSHF.curve_scale_start = float(self._inst_curve_scale_start)
DSHF.curve_scale_end = float(self._inst_curve_scale_end)
DSHF.curve_alpha = float(self._inst_curve_alpha)
DSHF.min_feature_size = int(self._inst_min_feature_size)
DSHF.auto_end_enable = bool(self._inst_auto_end_enable)
DSHF.auto_end_strength = float(self._inst_auto_end_strength)
DSHF.interp_method = str(self._inst_interp_method)
DSHF.interp_antialias = bool(self._inst_interp_antialias)
DSHF.channels_last = bool(self._inst_channels_last)
# experimental
DSHF.exp_section_enable = bool(self._inst_exp_section_enable)
DSHF.exp_enable = bool(self._inst_exp_section_enable and self._inst_exp_enable)
DSHF.exp_timestep = float(self._inst_exp_timestep)
DSHF.exp_scales = list(self._inst_exp_scales)
DSHF.exp_dilations = list(self._inst_exp_dilations)
DSHF.exp_in_muls = list(self._inst_exp_in_muls)
DSHF.exp_out_muls = list(self._inst_exp_out_muls)
DSHF.exp_cfg_muls = list(self._inst_exp_cfg_muls)
# --------- Профили (instance-вариант, чтобы сохранять то, что в UI) ---------
def _export_profile_instance(self) -> Dict[str, Any]:
return {
"actions": [a.__dict__ for a in self._inst_actions],
"curve": dict(
enable=self._inst_curve_enable, type=self._inst_curve_type,
t_start=self._inst_curve_t_start, t_end=self._inst_curve_t_end,
scale_start=self._inst_curve_scale_start, scale_end=self._inst_curve_scale_end,
alpha=self._inst_curve_alpha, min_feature=self._inst_min_feature_size,
auto_end_enable=self._inst_auto_end_enable, auto_end_strength=self._inst_auto_end_strength,
),
"runtime": dict(
interp_method=self._inst_interp_method, interp_antialias=self._inst_interp_antialias,
channels_last=self._inst_channels_last
),
"experimental": dict(
section_enable=self._inst_exp_section_enable, enable=self._inst_exp_enable, timestep=self._inst_exp_timestep,
scales=self._inst_exp_scales, dilations=self._inst_exp_dilations,
in_muls=self._inst_exp_in_muls, out_muls=self._inst_exp_out_muls, cfg_muls=self._inst_exp_cfg_muls
),
}
def _apply_profile_instance(self, data: Dict[str, Any]) -> None:
try:
self._inst_actions = [DSHFAction(bool(a.get("enable", True)),
float(a.get("timestep", 0)),
int(a.get("depth", 0)),
float(a.get("scale", 1.0)))
for a in data.get("actions", [])]
c = data.get("curve", {})
self._inst_curve_enable = bool(c.get("enable", False))
self._inst_curve_type = str(c.get("type", "linear"))
self._inst_curve_t_start = float(c.get("t_start", 800))
self._inst_curve_t_end = float(c.get("t_end", 200))
self._inst_curve_scale_start = float(c.get("scale_start", 1.0))
self._inst_curve_scale_end = float(c.get("scale_end", 1.0))
self._inst_curve_alpha = float(_clamp(float(c.get("alpha", 0.5)), 0.0, 1.0))
self._inst_min_feature_size = int(_clamp(float(c.get("min_feature", 8)), 2, 256))
self._inst_auto_end_enable = bool(c.get("auto_end_enable", False))
self._inst_auto_end_strength = float(_clamp(float(c.get("auto_end_strength", 0.35)), 0.0, 1.0))
r = data.get("runtime", {})
self._inst_interp_method = str(r.get("interp_method", self._inst_interp_method))
self._inst_interp_antialias = bool(r.get("interp_antialias", self._inst_interp_antialias))
self._inst_channels_last = bool(r.get("channels_last", self._inst_channels_last))
e = data.get("experimental", {})
self._inst_exp_section_enable = bool(e.get("section_enable", False))
self._inst_exp_enable = bool(e.get("enable", False))
self._inst_exp_timestep = float(e.get("timestep", 625))
self._inst_exp_scales = list(map(float, e.get("scales", [1.0])))
self._inst_exp_dilations = list(map(int, e.get("dilations", [1])))
self._inst_exp_in_muls = list(map(float, e.get("in_muls", [1.0])))
self._inst_exp_out_muls = list(map(float, e.get("out_muls", [1.0])))
self._inst_exp_cfg_muls = list(map(float, e.get("cfg_muls", [1.0])))
except Exception as ex:
print(f"[DSHF] Ошибка применения профиля: {ex}")
# --------- Обёртки Conv2d (работают только при включённом experimental) ---------
class DSHF_Scale(torch.nn.Module):
def __init__(self, conv2d_ref: List[torch.nn.Conv2d], get_rt):
super().__init__()
self.conv2d_ref = conv2d_ref
self.get_rt = get_rt # -> (mode, aa, min_feat)
def forward(self, h: torch.Tensor):
if not DSHF.exp_enable or DSHF.currentTimestep < DSHF.exp_timestep:
return h
mode, aa, min_feat = self.get_rt()
idx = DSHF.currentConv
pre_scale = 1.0 / _get_or_last(DSHF.exp_scales, idx, 1.0)
if pre_scale != 1.0:
h = _resize(h, pre_scale, mode, aa, min_feat)
conv = self.conv2d_ref[0]
k = conv.kernel_size if isinstance(conv.kernel_size, tuple) else (conv.kernel_size, conv.kernel_size)
if max(k) > 1:
dil = int(_get_or_last(DSHF.exp_dilations, idx, 1))
conv.dilation = (dil, dil); conv.padding = (dil, dil)
else:
conv.dilation = (1, 1); conv.padding = (0, 0)
return h
class DSHF_Unscale(torch.nn.Module):
def __init__(self, get_rt):
super().__init__()
self.get_rt = get_rt
def forward(self, h: torch.Tensor):
if not DSHF.exp_enable or DSHF.currentTimestep < DSHF.exp_timestep:
DSHF.currentConv += 1; return h
mode, aa, min_feat = self.get_rt()
idx = DSHF.currentConv
post_scale = _get_or_last(DSHF.exp_scales, idx, 1.0)
if post_scale != 1.0:
h = _resize(h, post_scale, mode, aa, min_feat)
alpha = float(DSHF.curve_alpha)
if alpha != 0.0:
h = h * (post_scale ** alpha)
DSHF.currentConv += 1
return h
class DSHF_InMul(torch.nn.Module):
def forward(self, h: torch.Tensor):
if not DSHF.exp_enable or DSHF.currentTimestep < DSHF.exp_timestep:
return h
mul = _get_or_last(DSHF.exp_in_muls, DSHF.currentBlock, 1.0)
return h if mul == 1.0 else h * float(mul)
class DSHF_OutMul(torch.nn.Module):
def forward(self, h: torch.Tensor):
if not DSHF.exp_enable or DSHF.currentTimestep < DSHF.exp_timestep:
return h
mul = _get_or_last(DSHF.exp_out_muls, DSHF.currentBlock, 1.0)
return h if mul == 1.0 else h * float(mul)
# ---------------- Подменённый U-Net ----------------
class DeepShrinkHiresFixUNet(sd_unet.SdUnet):
def __init__(self, _model):
super().__init__()
self.model = _model.to(devices.device)
getter = lambda: (DSHF.interp_method, DSHF.interp_antialias, DSHF.min_feature_size)
# Оборачивание слоёв
for i, input_block in enumerate(self.model.input_blocks):
for j, layer in enumerate(input_block):
if isinstance(layer, ResBlock):
for k, in_layer in enumerate(layer.in_layers):
if isinstance(in_layer, torch.nn.Conv2d):
self.model.input_blocks[i][j].in_layers[k] = torch.nn.Sequential(
DSHF.DSHF_Scale([in_layer], getter), in_layer, DSHF.DSHF_Unscale(getter), DSHF.DSHF_InMul()
)
for k, out_layer in enumerate(layer.out_layers):
if isinstance(out_layer, torch.nn.Conv2d):
self.model.input_blocks[i][j].out_layers[k] = torch.nn.Sequential(
DSHF.DSHF_Scale([out_layer], getter), out_layer, DSHF.DSHF_Unscale(getter), DSHF.DSHF_OutMul()
)
else:
if isinstance(layer, torch.nn.Conv2d):
self.model.input_blocks[i][j] = torch.nn.Sequential(
DSHF.DSHF_Scale([layer], getter), layer, DSHF.DSHF_Unscale(getter)
)
if isinstance(layer, Downsample):
layer.op = torch.nn.Sequential(DSHF.DSHF_Scale([layer.op], getter), layer.op, DSHF.DSHF_Unscale(getter))
if isinstance(layer, Upsample) and hasattr(layer, "conv") and isinstance(layer.conv, torch.nn.Conv2d):
layer.conv = torch.nn.Sequential(DSHF.DSHF_Scale([layer.conv], getter), layer.conv, DSHF.DSHF_Unscale(getter))
for j, layer in enumerate(self.model.middle_block):
if isinstance(layer, ResBlock):
for k, in_layer in enumerate(layer.in_layers):
if isinstance(in_layer, torch.nn.Conv2d):
self.model.middle_block[j].in_layers[k] = torch.nn.Sequential(
DSHF.DSHF_Scale([in_layer], getter), in_layer, DSHF.DSHF_Unscale(getter), DSHF.DSHF_InMul()
)
for k, out_layer in enumerate(layer.out_layers):
if isinstance(out_layer, torch.nn.Conv2d):
self.model.middle_block[j].out_layers[k] = torch.nn.Sequential(
DSHF.DSHF_Scale([out_layer], getter), out_layer, DSHF.DSHF_Unscale(getter), DSHF.DSHF_OutMul()
)
else:
if isinstance(layer, torch.nn.Conv2d):
self.model.middle_block[j] = torch.nn.Sequential(
DSHF.DSHF_Scale([layer], getter), layer, DSHF.DSHF_Unscale(getter)
)
if isinstance(layer, Downsample):
layer.op = torch.nn.Sequential(DSHF.DSHF_Scale([layer.op], getter), layer.op, DSHF.DSHF_Unscale(getter))
if isinstance(layer, Upsample) and hasattr(layer, "conv") and isinstance(layer.conv, torch.nn.Conv2d):
layer.conv = torch.nn.Sequential(DSHF.DSHF_Scale([layer.conv], getter), layer.conv, DSHF.DSHF_Unscale(getter))
for i, output_block in enumerate(self.model.output_blocks):
for j, layer in enumerate(output_block):
if isinstance(layer, ResBlock):
for k, in_layer in enumerate(layer.in_layers):
if isinstance(in_layer, torch.nn.Conv2d):
self.model.output_blocks[i][j].in_layers[k] = torch.nn.Sequential(
DSHF.DSHF_Scale([in_layer], getter), in_layer, DSHF.DSHF_Unscale(getter), DSHF.DSHF_InMul()
)
for k, out_layer in enumerate(layer.out_layers):
if isinstance(out_layer, torch.nn.Conv2d):
self.model.output_blocks[i][j].out_layers[k] = torch.nn.Sequential(
DSHF.DSHF_Scale([out_layer], getter), out_layer, DSHF.DSHF_Unscale(getter), DSHF.DSHF_OutMul()
)
else:
if isinstance(layer, torch.nn.Conv2d):
self.model.output_blocks[i][j] = torch.nn.Sequential(
DSHF.DSHF_Scale([layer], getter), layer, DSHF.DSHF_Unscale(getter)
)
if isinstance(layer, Downsample):
layer.op = torch.nn.Sequential(DSHF.DSHF_Scale([layer.op], getter), layer.op, DSHF.DSHF_Unscale(getter))
if isinstance(layer, Upsample) and hasattr(layer, "conv") and isinstance(layer.conv, torch.nn.Conv2d):
layer.conv = torch.nn.Sequential(DSHF.DSHF_Scale([layer.conv], getter), layer.conv, DSHF.DSHF_Unscale(getter))
for i, module in enumerate(self.model.out):
if isinstance(module, torch.nn.Conv2d):
self.model.out[i] = torch.nn.Sequential(DSHF.DSHF_Scale([module], getter), module, DSHF.DSHF_Unscale(getter))
def forward(self, x, timesteps, context, y=None, **kwargs):
assert (y is not None) == (self.model.num_classes is not None), "must specify y iff class-conditional"
if DSHF.channels_last:
x = x.contiguous(memory_format=torch.channels_last)
hs = []
emb = self.model.time_embed(timestep_embedding(timesteps, self.model.model_channels, repeat_only=False))
if self.model.num_classes is not None:
assert y.shape[0] == x.shape[0]
emb = emb + self.model.label_emb(y)
h = x.type(self.model.dtype)
depth = 0
DSHF.currentBlock = 0
DSHF.currentConv = 0
DSHF.currentTimestep = float(timesteps.detach().float().min().item())
# Входные блоки
for module in self.model.input_blocks:
s = DSHF._block_scale(depth, DSHF.currentTimestep)
if s is not None:
h = _resize(h, 1.0 / float(s), DSHF.interp_method, DSHF.interp_antialias, DSHF.min_feature_size)
context_tmp = context
if DSHF.exp_enable and DSHF.currentTimestep >= DSHF.exp_timestep:
cfg_mul = _get_or_last(DSHF.exp_cfg_muls, DSHF.currentBlock, 1.0)
if cfg_mul != 1.0:
context_tmp = context * float(cfg_mul)
h = module(h, emb, context_tmp)
hs.append(h)
depth += 1
DSHF.currentBlock += 1
# Средний блок
s = DSHF._block_scale(depth, DSHF.currentTimestep)
if s is not None:
h = _resize(h, 1.0 / float(s), DSHF.interp_method, DSHF.interp_antialias, DSHF.min_feature_size)
context_tmp = context
if DSHF.exp_enable and DSHF.currentTimestep >= DSHF.exp_timestep:
cfg_mul = _get_or_last(DSHF.exp_cfg_muls, DSHF.currentBlock, 1.0)
if cfg_mul != 1.0:
context_tmp = context * float(cfg_mul)
h = self.model.middle_block(h, emb, context_tmp)
s = DSHF._block_scale(depth, DSHF.currentTimestep)
if s is not None:
h = _resize(h, float(s), DSHF.interp_method, DSHF.interp_antialias, DSHF.min_feature_size)
DSHF.currentBlock += 1
# Выходные блоки
for module in self.model.output_blocks:
depth -= 1
h = torch.cat([h, hs.pop()], dim=1)
context_tmp = context
if DSHF.exp_enable and DSHF.currentTimestep >= DSHF.exp_timestep:
cfg_mul = _get_or_last(DSHF.exp_cfg_muls, DSHF.currentBlock, 1.0)
if cfg_mul != 1.0:
context_tmp = context * float(cfg_mul)
h = module(h, emb, context_tmp)
s = DSHF._block_scale(depth, DSHF.currentTimestep)
if s is not None:
h = _resize(h, float(s), DSHF.interp_method, DSHF.interp_antialias, DSHF.min_feature_size)
DSHF.currentBlock += 1
h = h.type(x.dtype)
return self.model.id_predictor(h) if self.model.predict_codebook_ids else self.model.out(h)
# Регистрация варианта U-Net
DeepShrinkHiresFixUNetOption = sd_unet.SdUnetOption()
DeepShrinkHiresFixUNetOption.label = "Deep Shrink Hires.fix"
DeepShrinkHiresFixUNetOption.create_unet = lambda: DSHF.DeepShrinkHiresFixUNet(shared.sd_model.model.diffusion_model)
script_callbacks.on_list_unets(lambda unets: unets.append(DeepShrinkHiresFixUNetOption))
|