Spaces:
Sleeping
Sleeping
File size: 56,054 Bytes
4450790 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 |
import math
from scipy import integrate
import torch
from torch import nn
import torchsde
from tqdm.auto import trange, tqdm
from . import utils
from . import deis
import comfy.model_patcher
import comfy.model_sampling
def append_zero(x):
return torch.cat([x, x.new_zeros([1])])
def get_sigmas_karras(n, sigma_min, sigma_max, rho=7., device='cpu'):
"""Constructs the noise schedule of Karras et al. (2022)."""
ramp = torch.linspace(0, 1, n, device=device)
min_inv_rho = sigma_min ** (1 / rho)
max_inv_rho = sigma_max ** (1 / rho)
sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
return append_zero(sigmas).to(device)
def get_sigmas_exponential(n, sigma_min, sigma_max, device='cpu'):
"""Constructs an exponential noise schedule."""
sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), n, device=device).exp()
return append_zero(sigmas)
def get_sigmas_polyexponential(n, sigma_min, sigma_max, rho=1., device='cpu'):
"""Constructs an polynomial in log sigma noise schedule."""
ramp = torch.linspace(1, 0, n, device=device) ** rho
sigmas = torch.exp(ramp * (math.log(sigma_max) - math.log(sigma_min)) + math.log(sigma_min))
return append_zero(sigmas)
def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'):
"""Constructs a continuous VP noise schedule."""
t = torch.linspace(1, eps_s, n, device=device)
sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1)
return append_zero(sigmas)
def get_sigmas_laplace(n, sigma_min, sigma_max, mu=0., beta=0.5, device='cpu'):
"""Constructs the noise schedule proposed by Tiankai et al. (2024). """
epsilon = 1e-5 # avoid log(0)
x = torch.linspace(0, 1, n, device=device)
clamp = lambda x: torch.clamp(x, min=sigma_min, max=sigma_max)
lmb = mu - beta * torch.sign(0.5-x) * torch.log(1 - 2 * torch.abs(0.5-x) + epsilon)
sigmas = clamp(torch.exp(lmb))
return sigmas
def to_d(x, sigma, denoised):
"""Converts a denoiser output to a Karras ODE derivative."""
return (x - denoised) / utils.append_dims(sigma, x.ndim)
def get_ancestral_step(sigma_from, sigma_to, eta=1.):
"""Calculates the noise level (sigma_down) to step down to and the amount
of noise to add (sigma_up) when doing an ancestral sampling step."""
if not eta:
return sigma_to, 0.
sigma_up = min(sigma_to, eta * (sigma_to ** 2 * (sigma_from ** 2 - sigma_to ** 2) / sigma_from ** 2) ** 0.5)
sigma_down = (sigma_to ** 2 - sigma_up ** 2) ** 0.5
return sigma_down, sigma_up
def default_noise_sampler(x):
return lambda sigma, sigma_next: torch.randn_like(x)
class BatchedBrownianTree:
"""A wrapper around torchsde.BrownianTree that enables batches of entropy."""
def __init__(self, x, t0, t1, seed=None, **kwargs):
self.cpu_tree = True
if "cpu" in kwargs:
self.cpu_tree = kwargs.pop("cpu")
t0, t1, self.sign = self.sort(t0, t1)
w0 = kwargs.get('w0', torch.zeros_like(x))
if seed is None:
seed = torch.randint(0, 2 ** 63 - 1, []).item()
self.batched = True
try:
assert len(seed) == x.shape[0]
w0 = w0[0]
except TypeError:
seed = [seed]
self.batched = False
if self.cpu_tree:
self.trees = [torchsde.BrownianTree(t0.cpu(), w0.cpu(), t1.cpu(), entropy=s, **kwargs) for s in seed]
else:
self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
@staticmethod
def sort(a, b):
return (a, b, 1) if a < b else (b, a, -1)
def __call__(self, t0, t1):
t0, t1, sign = self.sort(t0, t1)
if self.cpu_tree:
w = torch.stack([tree(t0.cpu().float(), t1.cpu().float()).to(t0.dtype).to(t0.device) for tree in self.trees]) * (self.sign * sign)
else:
w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
return w if self.batched else w[0]
class BrownianTreeNoiseSampler:
"""A noise sampler backed by a torchsde.BrownianTree.
Args:
x (Tensor): The tensor whose shape, device and dtype to use to generate
random samples.
sigma_min (float): The low end of the valid interval.
sigma_max (float): The high end of the valid interval.
seed (int or List[int]): The random seed. If a list of seeds is
supplied instead of a single integer, then the noise sampler will
use one BrownianTree per batch item, each with its own seed.
transform (callable): A function that maps sigma to the sampler's
internal timestep.
"""
def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x, cpu=False):
self.transform = transform
t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max))
self.tree = BatchedBrownianTree(x, t0, t1, seed, cpu=cpu)
def __call__(self, sigma, sigma_next):
t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next))
return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
@torch.no_grad()
def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
"""Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
if s_churn > 0:
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
sigma_hat = sigmas[i] * (gamma + 1)
else:
gamma = 0
sigma_hat = sigmas[i]
if gamma > 0:
eps = torch.randn_like(x) * s_noise
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
dt = sigmas[i + 1] - sigma_hat
# Euler method
x = x + d * dt
return x
@torch.no_grad()
def sample_euler_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
if isinstance(model.inner_model.inner_model.model_sampling, comfy.model_sampling.CONST):
return sample_euler_ancestral_RF(model, x, sigmas, extra_args, callback, disable, eta, s_noise, noise_sampler)
"""Ancestral sampling with Euler method steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d = to_d(x, sigmas[i], denoised)
# Euler method
dt = sigma_down - sigmas[i]
x = x + d * dt
if sigmas[i + 1] > 0:
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
return x
@torch.no_grad()
def sample_euler_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1.0, s_noise=1., noise_sampler=None):
"""Ancestral sampling with Euler method steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
# sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
downstep_ratio = 1 + (sigmas[i+1]/sigmas[i] - 1) * eta
sigma_down = sigmas[i+1] * downstep_ratio
alpha_ip1 = 1 - sigmas[i+1]
alpha_down = 1 - sigma_down
renoise_coeff = (sigmas[i+1]**2 - sigma_down**2*alpha_ip1**2/alpha_down**2)**0.5
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
# Euler method
sigma_down_i_ratio = sigma_down / sigmas[i]
x = sigma_down_i_ratio * x + (1 - sigma_down_i_ratio) * denoised
if sigmas[i + 1] > 0 and eta > 0:
x = (alpha_ip1/alpha_down) * x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * renoise_coeff
return x
@torch.no_grad()
def sample_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
"""Implements Algorithm 2 (Heun steps) from Karras et al. (2022)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
if s_churn > 0:
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
sigma_hat = sigmas[i] * (gamma + 1)
else:
gamma = 0
sigma_hat = sigmas[i]
sigma_hat = sigmas[i] * (gamma + 1)
if gamma > 0:
eps = torch.randn_like(x) * s_noise
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
dt = sigmas[i + 1] - sigma_hat
if sigmas[i + 1] == 0:
# Euler method
x = x + d * dt
else:
# Heun's method
x_2 = x + d * dt
denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
d_prime = (d + d_2) / 2
x = x + d_prime * dt
return x
@torch.no_grad()
def sample_dpm_2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
"""A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
if s_churn > 0:
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
sigma_hat = sigmas[i] * (gamma + 1)
else:
gamma = 0
sigma_hat = sigmas[i]
if gamma > 0:
eps = torch.randn_like(x) * s_noise
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
if sigmas[i + 1] == 0:
# Euler method
dt = sigmas[i + 1] - sigma_hat
x = x + d * dt
else:
# DPM-Solver-2
sigma_mid = sigma_hat.log().lerp(sigmas[i + 1].log(), 0.5).exp()
dt_1 = sigma_mid - sigma_hat
dt_2 = sigmas[i + 1] - sigma_hat
x_2 = x + d * dt_1
denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
d_2 = to_d(x_2, sigma_mid, denoised_2)
x = x + d_2 * dt_2
return x
@torch.no_grad()
def sample_dpm_2_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
if isinstance(model.inner_model.inner_model.model_sampling, comfy.model_sampling.CONST):
return sample_dpm_2_ancestral_RF(model, x, sigmas, extra_args, callback, disable, eta, s_noise, noise_sampler)
"""Ancestral sampling with DPM-Solver second-order steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d = to_d(x, sigmas[i], denoised)
if sigma_down == 0:
# Euler method
dt = sigma_down - sigmas[i]
x = x + d * dt
else:
# DPM-Solver-2
sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp()
dt_1 = sigma_mid - sigmas[i]
dt_2 = sigma_down - sigmas[i]
x_2 = x + d * dt_1
denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
d_2 = to_d(x_2, sigma_mid, denoised_2)
x = x + d_2 * dt_2
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
return x
@torch.no_grad()
def sample_dpm_2_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""Ancestral sampling with DPM-Solver second-order steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
downstep_ratio = 1 + (sigmas[i+1]/sigmas[i] - 1) * eta
sigma_down = sigmas[i+1] * downstep_ratio
alpha_ip1 = 1 - sigmas[i+1]
alpha_down = 1 - sigma_down
renoise_coeff = (sigmas[i+1]**2 - sigma_down**2*alpha_ip1**2/alpha_down**2)**0.5
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d = to_d(x, sigmas[i], denoised)
if sigma_down == 0:
# Euler method
dt = sigma_down - sigmas[i]
x = x + d * dt
else:
# DPM-Solver-2
sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp()
dt_1 = sigma_mid - sigmas[i]
dt_2 = sigma_down - sigmas[i]
x_2 = x + d * dt_1
denoised_2 = model(x_2, sigma_mid * s_in, **extra_args)
d_2 = to_d(x_2, sigma_mid, denoised_2)
x = x + d_2 * dt_2
x = (alpha_ip1/alpha_down) * x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * renoise_coeff
return x
def linear_multistep_coeff(order, t, i, j):
if order - 1 > i:
raise ValueError(f'Order {order} too high for step {i}')
def fn(tau):
prod = 1.
for k in range(order):
if j == k:
continue
prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
return prod
return integrate.quad(fn, t[i], t[i + 1], epsrel=1e-4)[0]
@torch.no_grad()
def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, order=4):
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigmas_cpu = sigmas.detach().cpu().numpy()
ds = []
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
d = to_d(x, sigmas[i], denoised)
ds.append(d)
if len(ds) > order:
ds.pop(0)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
cur_order = min(i + 1, order)
coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
return x
class PIDStepSizeController:
"""A PID controller for ODE adaptive step size control."""
def __init__(self, h, pcoeff, icoeff, dcoeff, order=1, accept_safety=0.81, eps=1e-8):
self.h = h
self.b1 = (pcoeff + icoeff + dcoeff) / order
self.b2 = -(pcoeff + 2 * dcoeff) / order
self.b3 = dcoeff / order
self.accept_safety = accept_safety
self.eps = eps
self.errs = []
def limiter(self, x):
return 1 + math.atan(x - 1)
def propose_step(self, error):
inv_error = 1 / (float(error) + self.eps)
if not self.errs:
self.errs = [inv_error, inv_error, inv_error]
self.errs[0] = inv_error
factor = self.errs[0] ** self.b1 * self.errs[1] ** self.b2 * self.errs[2] ** self.b3
factor = self.limiter(factor)
accept = factor >= self.accept_safety
if accept:
self.errs[2] = self.errs[1]
self.errs[1] = self.errs[0]
self.h *= factor
return accept
class DPMSolver(nn.Module):
"""DPM-Solver. See https://arxiv.org/abs/2206.00927."""
def __init__(self, model, extra_args=None, eps_callback=None, info_callback=None):
super().__init__()
self.model = model
self.extra_args = {} if extra_args is None else extra_args
self.eps_callback = eps_callback
self.info_callback = info_callback
def t(self, sigma):
return -sigma.log()
def sigma(self, t):
return t.neg().exp()
def eps(self, eps_cache, key, x, t, *args, **kwargs):
if key in eps_cache:
return eps_cache[key], eps_cache
sigma = self.sigma(t) * x.new_ones([x.shape[0]])
eps = (x - self.model(x, sigma, *args, **self.extra_args, **kwargs)) / self.sigma(t)
if self.eps_callback is not None:
self.eps_callback()
return eps, {key: eps, **eps_cache}
def dpm_solver_1_step(self, x, t, t_next, eps_cache=None):
eps_cache = {} if eps_cache is None else eps_cache
h = t_next - t
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
x_1 = x - self.sigma(t_next) * h.expm1() * eps
return x_1, eps_cache
def dpm_solver_2_step(self, x, t, t_next, r1=1 / 2, eps_cache=None):
eps_cache = {} if eps_cache is None else eps_cache
h = t_next - t
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
s1 = t + r1 * h
u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
x_2 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / (2 * r1) * h.expm1() * (eps_r1 - eps)
return x_2, eps_cache
def dpm_solver_3_step(self, x, t, t_next, r1=1 / 3, r2=2 / 3, eps_cache=None):
eps_cache = {} if eps_cache is None else eps_cache
h = t_next - t
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
s1 = t + r1 * h
s2 = t + r2 * h
u1 = x - self.sigma(s1) * (r1 * h).expm1() * eps
eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', u1, s1)
u2 = x - self.sigma(s2) * (r2 * h).expm1() * eps - self.sigma(s2) * (r2 / r1) * ((r2 * h).expm1() / (r2 * h) - 1) * (eps_r1 - eps)
eps_r2, eps_cache = self.eps(eps_cache, 'eps_r2', u2, s2)
x_3 = x - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / r2 * (h.expm1() / h - 1) * (eps_r2 - eps)
return x_3, eps_cache
def dpm_solver_fast(self, x, t_start, t_end, nfe, eta=0., s_noise=1., noise_sampler=None):
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
if not t_end > t_start and eta:
raise ValueError('eta must be 0 for reverse sampling')
m = math.floor(nfe / 3) + 1
ts = torch.linspace(t_start, t_end, m + 1, device=x.device)
if nfe % 3 == 0:
orders = [3] * (m - 2) + [2, 1]
else:
orders = [3] * (m - 1) + [nfe % 3]
for i in range(len(orders)):
eps_cache = {}
t, t_next = ts[i], ts[i + 1]
if eta:
sd, su = get_ancestral_step(self.sigma(t), self.sigma(t_next), eta)
t_next_ = torch.minimum(t_end, self.t(sd))
su = (self.sigma(t_next) ** 2 - self.sigma(t_next_) ** 2) ** 0.5
else:
t_next_, su = t_next, 0.
eps, eps_cache = self.eps(eps_cache, 'eps', x, t)
denoised = x - self.sigma(t) * eps
if self.info_callback is not None:
self.info_callback({'x': x, 'i': i, 't': ts[i], 't_up': t, 'denoised': denoised})
if orders[i] == 1:
x, eps_cache = self.dpm_solver_1_step(x, t, t_next_, eps_cache=eps_cache)
elif orders[i] == 2:
x, eps_cache = self.dpm_solver_2_step(x, t, t_next_, eps_cache=eps_cache)
else:
x, eps_cache = self.dpm_solver_3_step(x, t, t_next_, eps_cache=eps_cache)
x = x + su * s_noise * noise_sampler(self.sigma(t), self.sigma(t_next))
return x
def dpm_solver_adaptive(self, x, t_start, t_end, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None):
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
if order not in {2, 3}:
raise ValueError('order should be 2 or 3')
forward = t_end > t_start
if not forward and eta:
raise ValueError('eta must be 0 for reverse sampling')
h_init = abs(h_init) * (1 if forward else -1)
atol = torch.tensor(atol)
rtol = torch.tensor(rtol)
s = t_start
x_prev = x
accept = True
pid = PIDStepSizeController(h_init, pcoeff, icoeff, dcoeff, 1.5 if eta else order, accept_safety)
info = {'steps': 0, 'nfe': 0, 'n_accept': 0, 'n_reject': 0}
while s < t_end - 1e-5 if forward else s > t_end + 1e-5:
eps_cache = {}
t = torch.minimum(t_end, s + pid.h) if forward else torch.maximum(t_end, s + pid.h)
if eta:
sd, su = get_ancestral_step(self.sigma(s), self.sigma(t), eta)
t_ = torch.minimum(t_end, self.t(sd))
su = (self.sigma(t) ** 2 - self.sigma(t_) ** 2) ** 0.5
else:
t_, su = t, 0.
eps, eps_cache = self.eps(eps_cache, 'eps', x, s)
denoised = x - self.sigma(s) * eps
if order == 2:
x_low, eps_cache = self.dpm_solver_1_step(x, s, t_, eps_cache=eps_cache)
x_high, eps_cache = self.dpm_solver_2_step(x, s, t_, eps_cache=eps_cache)
else:
x_low, eps_cache = self.dpm_solver_2_step(x, s, t_, r1=1 / 3, eps_cache=eps_cache)
x_high, eps_cache = self.dpm_solver_3_step(x, s, t_, eps_cache=eps_cache)
delta = torch.maximum(atol, rtol * torch.maximum(x_low.abs(), x_prev.abs()))
error = torch.linalg.norm((x_low - x_high) / delta) / x.numel() ** 0.5
accept = pid.propose_step(error)
if accept:
x_prev = x_low
x = x_high + su * s_noise * noise_sampler(self.sigma(s), self.sigma(t))
s = t
info['n_accept'] += 1
else:
info['n_reject'] += 1
info['nfe'] += order
info['steps'] += 1
if self.info_callback is not None:
self.info_callback({'x': x, 'i': info['steps'] - 1, 't': s, 't_up': s, 'denoised': denoised, 'error': error, 'h': pid.h, **info})
return x, info
@torch.no_grad()
def sample_dpm_fast(model, x, sigma_min, sigma_max, n, extra_args=None, callback=None, disable=None, eta=0., s_noise=1., noise_sampler=None):
"""DPM-Solver-Fast (fixed step size). See https://arxiv.org/abs/2206.00927."""
if sigma_min <= 0 or sigma_max <= 0:
raise ValueError('sigma_min and sigma_max must not be 0')
with tqdm(total=n, disable=disable) as pbar:
dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
if callback is not None:
dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
return dpm_solver.dpm_solver_fast(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), n, eta, s_noise, noise_sampler)
@torch.no_grad()
def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callback=None, disable=None, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1., noise_sampler=None, return_info=False):
"""DPM-Solver-12 and 23 (adaptive step size). See https://arxiv.org/abs/2206.00927."""
if sigma_min <= 0 or sigma_max <= 0:
raise ValueError('sigma_min and sigma_max must not be 0')
with tqdm(disable=disable) as pbar:
dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update)
if callback is not None:
dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info})
x, info = dpm_solver.dpm_solver_adaptive(x, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise, noise_sampler)
if return_info:
return x, info
return x
@torch.no_grad()
def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
if isinstance(model.inner_model.inner_model.model_sampling, comfy.model_sampling.CONST):
return sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args, callback, disable, eta, s_noise, noise_sampler)
"""Ancestral sampling with DPM-Solver++(2S) second-order steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigma_down == 0:
# Euler method
d = to_d(x, sigmas[i], denoised)
dt = sigma_down - sigmas[i]
x = x + d * dt
else:
# DPM-Solver++(2S)
t, t_next = t_fn(sigmas[i]), t_fn(sigma_down)
r = 1 / 2
h = t_next - t
s = t + r * h
x_2 = (sigma_fn(s) / sigma_fn(t)) * x - (-h * r).expm1() * denoised
denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_2
# Noise addition
if sigmas[i + 1] > 0:
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
return x
@torch.no_grad()
def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""Ancestral sampling with DPM-Solver++(2S) second-order steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda lbda: (lbda.exp() + 1) ** -1
lambda_fn = lambda sigma: ((1-sigma)/sigma).log()
# logged_x = x.unsqueeze(0)
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
downstep_ratio = 1 + (sigmas[i+1]/sigmas[i] - 1) * eta
sigma_down = sigmas[i+1] * downstep_ratio
alpha_ip1 = 1 - sigmas[i+1]
alpha_down = 1 - sigma_down
renoise_coeff = (sigmas[i+1]**2 - sigma_down**2*alpha_ip1**2/alpha_down**2)**0.5
# sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigmas[i + 1] == 0:
# Euler method
d = to_d(x, sigmas[i], denoised)
dt = sigma_down - sigmas[i]
x = x + d * dt
else:
# DPM-Solver++(2S)
if sigmas[i] == 1.0:
sigma_s = 0.9999
else:
t_i, t_down = lambda_fn(sigmas[i]), lambda_fn(sigma_down)
r = 1 / 2
h = t_down - t_i
s = t_i + r * h
sigma_s = sigma_fn(s)
# sigma_s = sigmas[i+1]
sigma_s_i_ratio = sigma_s / sigmas[i]
u = sigma_s_i_ratio * x + (1 - sigma_s_i_ratio) * denoised
D_i = model(u, sigma_s * s_in, **extra_args)
sigma_down_i_ratio = sigma_down / sigmas[i]
x = sigma_down_i_ratio * x + (1 - sigma_down_i_ratio) * D_i
# print("sigma_i", sigmas[i], "sigma_ip1", sigmas[i+1],"sigma_down", sigma_down, "sigma_down_i_ratio", sigma_down_i_ratio, "sigma_s_i_ratio", sigma_s_i_ratio, "renoise_coeff", renoise_coeff)
# Noise addition
if sigmas[i + 1] > 0 and eta > 0:
x = (alpha_ip1/alpha_down) * x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * renoise_coeff
# logged_x = torch.cat((logged_x, x.unsqueeze(0)), dim=0)
return x
@torch.no_grad()
def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
"""DPM-Solver++ (stochastic)."""
if len(sigmas) <= 1:
return x
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
seed = extra_args.get("seed", None)
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigmas[i + 1] == 0:
# Euler method
d = to_d(x, sigmas[i], denoised)
dt = sigmas[i + 1] - sigmas[i]
x = x + d * dt
else:
# DPM-Solver++
t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
h = t_next - t
s = t + h * r
fac = 1 / (2 * r)
# Step 1
sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta)
s_ = t_fn(sd)
x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised
x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su
denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
# Step 2
sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta)
t_next_ = t_fn(sd)
denoised_d = (1 - fac) * denoised + fac * denoised_2
x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d
x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su
return x
@torch.no_grad()
def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=None):
"""DPM-Solver++(2M)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()
old_denoised = None
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
h = t_next - t
if old_denoised is None or sigmas[i + 1] == 0:
x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised
else:
h_last = t - t_fn(sigmas[i - 1])
r = h_last / h
denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised
x = (sigma_fn(t_next) / sigma_fn(t)) * x - (-h).expm1() * denoised_d
old_denoised = denoised
return x
@torch.no_grad()
def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
"""DPM-Solver++(2M) SDE."""
if len(sigmas) <= 1:
return x
if solver_type not in {'heun', 'midpoint'}:
raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
seed = extra_args.get("seed", None)
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
old_denoised = None
h_last = None
h = None
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigmas[i + 1] == 0:
# Denoising step
x = denoised
else:
# DPM-Solver++(2M) SDE
t, s = -sigmas[i].log(), -sigmas[i + 1].log()
h = s - t
eta_h = eta * h
x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised
if old_denoised is not None:
r = h_last / h
if solver_type == 'heun':
x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)
elif solver_type == 'midpoint':
x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)
if eta:
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
old_denoised = denoised
h_last = h
return x
@torch.no_grad()
def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""DPM-Solver++(3M) SDE."""
if len(sigmas) <= 1:
return x
seed = extra_args.get("seed", None)
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
denoised_1, denoised_2 = None, None
h, h_1, h_2 = None, None, None
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigmas[i + 1] == 0:
# Denoising step
x = denoised
else:
t, s = -sigmas[i].log(), -sigmas[i + 1].log()
h = s - t
h_eta = h * (eta + 1)
x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised
if h_2 is not None:
r0 = h_1 / h
r1 = h_2 / h
d1_0 = (denoised - denoised_1) / r0
d1_1 = (denoised_1 - denoised_2) / r1
d1 = d1_0 + (d1_0 - d1_1) * r0 / (r0 + r1)
d2 = (d1_0 - d1_1) / (r0 + r1)
phi_2 = h_eta.neg().expm1() / h_eta + 1
phi_3 = phi_2 / h_eta - 0.5
x = x + phi_2 * d1 - phi_3 * d2
elif h_1 is not None:
r = h_1 / h
d = (denoised - denoised_1) / r
phi_2 = h_eta.neg().expm1() / h_eta + 1
x = x + phi_2 * d
if eta:
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
denoised_1, denoised_2 = denoised, denoised_1
h_1, h_2 = h, h_1
return x
@torch.no_grad()
def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
if len(sigmas) <= 1:
return x
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
@torch.no_grad()
def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
if len(sigmas) <= 1:
return x
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
@torch.no_grad()
def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
if len(sigmas) <= 1:
return x
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
return sample_dpmpp_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=r)
def DDPMSampler_step(x, sigma, sigma_prev, noise, noise_sampler):
alpha_cumprod = 1 / ((sigma * sigma) + 1)
alpha_cumprod_prev = 1 / ((sigma_prev * sigma_prev) + 1)
alpha = (alpha_cumprod / alpha_cumprod_prev)
mu = (1.0 / alpha).sqrt() * (x - (1 - alpha) * noise / (1 - alpha_cumprod).sqrt())
if sigma_prev > 0:
mu += ((1 - alpha) * (1. - alpha_cumprod_prev) / (1. - alpha_cumprod)).sqrt() * noise_sampler(sigma, sigma_prev)
return mu
def generic_step_sampler(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None, step_function=None):
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
x = step_function(x / torch.sqrt(1.0 + sigmas[i] ** 2.0), sigmas[i], sigmas[i + 1], (x - denoised) / sigmas[i], noise_sampler)
if sigmas[i + 1] != 0:
x *= torch.sqrt(1.0 + sigmas[i + 1] ** 2.0)
return x
@torch.no_grad()
def sample_ddpm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None):
return generic_step_sampler(model, x, sigmas, extra_args, callback, disable, noise_sampler, DDPMSampler_step)
@torch.no_grad()
def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler=None):
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
x = denoised
if sigmas[i + 1] > 0:
x = model.inner_model.inner_model.model_sampling.noise_scaling(sigmas[i + 1], noise_sampler(sigmas[i], sigmas[i + 1]), x)
return x
@torch.no_grad()
def sample_heunpp2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
# From MIT licensed: https://github.com/Carzit/sd-webui-samplers-scheduler/
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
s_end = sigmas[-1]
for i in trange(len(sigmas) - 1, disable=disable):
gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
eps = torch.randn_like(x) * s_noise
sigma_hat = sigmas[i] * (gamma + 1)
if gamma > 0:
x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
dt = sigmas[i + 1] - sigma_hat
if sigmas[i + 1] == s_end:
# Euler method
x = x + d * dt
elif sigmas[i + 2] == s_end:
# Heun's method
x_2 = x + d * dt
denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
w = 2 * sigmas[0]
w2 = sigmas[i+1]/w
w1 = 1 - w2
d_prime = d * w1 + d_2 * w2
x = x + d_prime * dt
else:
# Heun++
x_2 = x + d * dt
denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
dt_2 = sigmas[i + 2] - sigmas[i + 1]
x_3 = x_2 + d_2 * dt_2
denoised_3 = model(x_3, sigmas[i + 2] * s_in, **extra_args)
d_3 = to_d(x_3, sigmas[i + 2], denoised_3)
w = 3 * sigmas[0]
w2 = sigmas[i + 1] / w
w3 = sigmas[i + 2] / w
w1 = 1 - w2 - w3
d_prime = w1 * d + w2 * d_2 + w3 * d_3
x = x + d_prime * dt
return x
#From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
#under Apache 2 license
def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
x_next = x
buffer_model = []
for i in trange(len(sigmas) - 1, disable=disable):
t_cur = sigmas[i]
t_next = sigmas[i + 1]
x_cur = x_next
denoised = model(x_cur, t_cur * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d_cur = (x_cur - denoised) / t_cur
order = min(max_order, i+1)
if order == 1: # First Euler step.
x_next = x_cur + (t_next - t_cur) * d_cur
elif order == 2: # Use one history point.
x_next = x_cur + (t_next - t_cur) * (3 * d_cur - buffer_model[-1]) / 2
elif order == 3: # Use two history points.
x_next = x_cur + (t_next - t_cur) * (23 * d_cur - 16 * buffer_model[-1] + 5 * buffer_model[-2]) / 12
elif order == 4: # Use three history points.
x_next = x_cur + (t_next - t_cur) * (55 * d_cur - 59 * buffer_model[-1] + 37 * buffer_model[-2] - 9 * buffer_model[-3]) / 24
if len(buffer_model) == max_order - 1:
for k in range(max_order - 2):
buffer_model[k] = buffer_model[k+1]
buffer_model[-1] = d_cur
else:
buffer_model.append(d_cur)
return x_next
#From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
#under Apache 2 license
def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
x_next = x
t_steps = sigmas
buffer_model = []
for i in trange(len(sigmas) - 1, disable=disable):
t_cur = sigmas[i]
t_next = sigmas[i + 1]
x_cur = x_next
denoised = model(x_cur, t_cur * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d_cur = (x_cur - denoised) / t_cur
order = min(max_order, i+1)
if order == 1: # First Euler step.
x_next = x_cur + (t_next - t_cur) * d_cur
elif order == 2: # Use one history point.
h_n = (t_next - t_cur)
h_n_1 = (t_cur - t_steps[i-1])
coeff1 = (2 + (h_n / h_n_1)) / 2
coeff2 = -(h_n / h_n_1) / 2
x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1])
elif order == 3: # Use two history points.
h_n = (t_next - t_cur)
h_n_1 = (t_cur - t_steps[i-1])
h_n_2 = (t_steps[i-1] - t_steps[i-2])
temp = (1 - h_n / (3 * (h_n + h_n_1)) * (h_n * (h_n + h_n_1)) / (h_n_1 * (h_n_1 + h_n_2))) / 2
coeff1 = (2 + (h_n / h_n_1)) / 2 + temp
coeff2 = -(h_n / h_n_1) / 2 - (1 + h_n_1 / h_n_2) * temp
coeff3 = temp * h_n_1 / h_n_2
x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1] + coeff3 * buffer_model[-2])
elif order == 4: # Use three history points.
h_n = (t_next - t_cur)
h_n_1 = (t_cur - t_steps[i-1])
h_n_2 = (t_steps[i-1] - t_steps[i-2])
h_n_3 = (t_steps[i-2] - t_steps[i-3])
temp1 = (1 - h_n / (3 * (h_n + h_n_1)) * (h_n * (h_n + h_n_1)) / (h_n_1 * (h_n_1 + h_n_2))) / 2
temp2 = ((1 - h_n / (3 * (h_n + h_n_1))) / 2 + (1 - h_n / (2 * (h_n + h_n_1))) * h_n / (6 * (h_n + h_n_1 + h_n_2))) \
* (h_n * (h_n + h_n_1) * (h_n + h_n_1 + h_n_2)) / (h_n_1 * (h_n_1 + h_n_2) * (h_n_1 + h_n_2 + h_n_3))
coeff1 = (2 + (h_n / h_n_1)) / 2 + temp1 + temp2
coeff2 = -(h_n / h_n_1) / 2 - (1 + h_n_1 / h_n_2) * temp1 - (1 + (h_n_1 / h_n_2) + (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3)))) * temp2
coeff3 = temp1 * h_n_1 / h_n_2 + ((h_n_1 / h_n_2) + (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3))) * (1 + h_n_2 / h_n_3)) * temp2
coeff4 = -temp2 * (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3))) * h_n_1 / h_n_2
x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1] + coeff3 * buffer_model[-2] + coeff4 * buffer_model[-3])
if len(buffer_model) == max_order - 1:
for k in range(max_order - 2):
buffer_model[k] = buffer_model[k+1]
buffer_model[-1] = d_cur.detach()
else:
buffer_model.append(d_cur.detach())
return x_next
#From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
#under Apache 2 license
@torch.no_grad()
def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=3, deis_mode='tab'):
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
x_next = x
t_steps = sigmas
coeff_list = deis.get_deis_coeff_list(t_steps, max_order, deis_mode=deis_mode)
buffer_model = []
for i in trange(len(sigmas) - 1, disable=disable):
t_cur = sigmas[i]
t_next = sigmas[i + 1]
x_cur = x_next
denoised = model(x_cur, t_cur * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d_cur = (x_cur - denoised) / t_cur
order = min(max_order, i+1)
if t_next <= 0:
order = 1
if order == 1: # First Euler step.
x_next = x_cur + (t_next - t_cur) * d_cur
elif order == 2: # Use one history point.
coeff_cur, coeff_prev1 = coeff_list[i]
x_next = x_cur + coeff_cur * d_cur + coeff_prev1 * buffer_model[-1]
elif order == 3: # Use two history points.
coeff_cur, coeff_prev1, coeff_prev2 = coeff_list[i]
x_next = x_cur + coeff_cur * d_cur + coeff_prev1 * buffer_model[-1] + coeff_prev2 * buffer_model[-2]
elif order == 4: # Use three history points.
coeff_cur, coeff_prev1, coeff_prev2, coeff_prev3 = coeff_list[i]
x_next = x_cur + coeff_cur * d_cur + coeff_prev1 * buffer_model[-1] + coeff_prev2 * buffer_model[-2] + coeff_prev3 * buffer_model[-3]
if len(buffer_model) == max_order - 1:
for k in range(max_order - 2):
buffer_model[k] = buffer_model[k+1]
buffer_model[-1] = d_cur.detach()
else:
buffer_model.append(d_cur.detach())
return x_next
@torch.no_grad()
def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
extra_args = {} if extra_args is None else extra_args
temp = [0]
def post_cfg_function(args):
temp[0] = args["uncond_denoised"]
return args["denoised"]
model_options = extra_args.get("model_options", {}).copy()
extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
sigma_hat = sigmas[i]
denoised = model(x, sigma_hat * s_in, **extra_args)
d = to_d(x, sigma_hat, temp[0])
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
# Euler method
x = denoised + d * sigmas[i + 1]
return x
@torch.no_grad()
def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""Ancestral sampling with Euler method steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
temp = [0]
def post_cfg_function(args):
temp[0] = args["uncond_denoised"]
return args["denoised"]
model_options = extra_args.get("model_options", {}).copy()
extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
d = to_d(x, sigmas[i], temp[0])
# Euler method
x = denoised + d * sigma_down
if sigmas[i + 1] > 0:
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
return x
@torch.no_grad()
def sample_dpmpp_2s_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""Ancestral sampling with DPM-Solver++(2S) second-order steps."""
extra_args = {} if extra_args is None else extra_args
noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
temp = [0]
def post_cfg_function(args):
temp[0] = args["uncond_denoised"]
return args["denoised"]
model_options = extra_args.get("model_options", {}).copy()
extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
if sigma_down == 0:
# Euler method
d = to_d(x, sigmas[i], temp[0])
x = denoised + d * sigma_down
else:
# DPM-Solver++(2S)
t, t_next = t_fn(sigmas[i]), t_fn(sigma_down)
# r = torch.sinh(1 + (2 - eta) * (t_next - t) / (t - t_fn(sigma_up))) works only on non-cfgpp, weird
r = 1 / 2
h = t_next - t
s = t + r * h
x_2 = (sigma_fn(s) / sigma_fn(t)) * (x + (denoised - temp[0])) - (-h * r).expm1() * denoised
denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
x = (sigma_fn(t_next) / sigma_fn(t)) * (x + (denoised - temp[0])) - (-h).expm1() * denoised_2
# Noise addition
if sigmas[i + 1] > 0:
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
return x
@torch.no_grad()
def sample_dpmpp_2m_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
"""DPM-Solver++(2M)."""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
t_fn = lambda sigma: sigma.log().neg()
old_uncond_denoised = None
uncond_denoised = None
def post_cfg_function(args):
nonlocal uncond_denoised
uncond_denoised = args["uncond_denoised"]
return args["denoised"]
model_options = extra_args.get("model_options", {}).copy()
extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
h = t_next - t
if old_uncond_denoised is None or sigmas[i + 1] == 0:
denoised_mix = -torch.exp(-h) * uncond_denoised
else:
h_last = t - t_fn(sigmas[i - 1])
r = h_last / h
denoised_mix = -torch.exp(-h) * uncond_denoised - torch.expm1(-h) * (1 / (2 * r)) * (denoised - old_uncond_denoised)
x = denoised + denoised_mix + torch.exp(-h) * x
old_uncond_denoised = uncond_denoised
return x
|