-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathrk_sampler_beta.py
585 lines (437 loc) · 33.1 KB
/
rk_sampler_beta.py
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
import torch
from tqdm.auto import trange
import gc
from .rk_method_beta import RK_Method_Beta, RK_NoiseSampler
from .rk_guide_func_beta import *
from .helper import get_extra_options_kv, extra_options_flag, lagrange_interpolation
from .phi_functions import Phi
from .res4lyf import RESplain
from .config import MAX_STEPS
def init_implicit_sampling(RK, x_0, x_, eps_, eps_prev_, data_, eps, denoised, denoised_prev, denoised_prev2, step, sigmas, h, s_, extra_options):
sigma = sigmas[step]
if extra_options_flag("implicit_skip_model_call_at_start", extra_options) and denoised.sum() + eps.sum() != 0:
eps_[0], data_[0] = eps.clone(), denoised.clone()
eps_[0] = RK.get_epsilon_anchored(x_0, denoised, sigma)
if denoised_prev2.sum() != 0:
sratio = sigma - s_[0]
data_[0] = denoised + sratio * (denoised - denoised_prev2)
elif extra_options_flag("implicit_full_skip_model_call_at_start", extra_options) and denoised.sum() + eps.sum() != 0:
eps_[0], data_[0] = eps.clone(), denoised.clone()
eps_[0] = RK.get_epsilon_anchored(x_0, denoised, sigma)
if denoised_prev2.sum() != 0:
for r in range(RK.rows):
sratio = sigma - s_[r]
data_[r] = denoised + sratio * (denoised - denoised_prev2)
eps_[r] = RK.get_epsilon_anchored(x_0, data_[r], s_[r])
elif extra_options_flag("implicit_lagrange_skip_model_call_at_start", extra_options) and denoised.sum() + eps.sum() != 0:
if denoised_prev2.sum() != 0:
sigma_prev = sigmas[step-1]
h_prev = sigma - sigma_prev
w = h / h_prev
substeps_prev = len(RK.C[:-1])
for r in range(RK.rows):
sratio = sigma - s_[r]
data_[r] = lagrange_interpolation([0,1], [denoised_prev2, denoised], 1 + w*RK.C[r]).squeeze(0) + denoised_prev2 - denoised
eps_[r] = RK.get_epsilon_anchored(x_0, data_[r], s_[r])
if extra_options_flag("implicit_lagrange_skip_model_call_at_start_0_only", extra_options):
for r in range(RK.rows):
eps_ [r] = eps_ [0].clone() * s_[0] / s_[r]
data_[r] = denoised.clone()
else:
eps_[0], data_[0] = eps.clone(), denoised.clone()
eps_[0] = RK.get_epsilon_anchored(x_0, denoised, sigma)
elif extra_options_flag("implicit_lagrange_init", extra_options) and denoised.sum() + eps.sum() != 0:
sigma_prev = sigmas[step-1]
h_prev = sigma - sigma_prev
w = h / h_prev
substeps_prev = len(RK.C[:-1])
z_prev_ = eps_.clone()
for r in range (substeps_prev):
z_prev_[r] = h * RK.zum(r, eps_) # u,v not implemented for lagrange guess for implicit
zi_1 = lagrange_interpolation(RK.C[:-1], z_prev_[:substeps_prev], RK.C[0]).squeeze(0) # + x_prev - x_0"""
x_[0] = x_0 + zi_1
else:
eps_[0], data_[0] = RK(x_[0], sigma, x_0, sigma)
if not extra_options_flag("implicit_lagrange_init", extra_options) \
and not extra_options_flag("radaucycle", extra_options) \
and not extra_options_flag("implicit_full_skip_model_call_at_start", extra_options) \
and not extra_options_flag("implicit_lagrange_skip_model_call_at_start", extra_options):
for r in range(RK.rows):
eps_ [r] = eps_ [0].clone() * sigma / s_[r]
data_[r] = data_[0].clone()
x_, eps_ = RK.newton_iter(x_0, x_, eps_, eps_prev_, data_, s_, 0, h, sigmas, step, "init", extra_options)
return x_, eps_, data_
@torch.no_grad()
def sample_rk_beta(model, x, sigmas, extra_args=None, callback=None, disable=None, noise_sampler_type="gaussian", noise_sampler_type_substep="gaussian", noise_mode_sde="hard", noise_seed=-1, rk_type="res_2m", implicit_sampler_name="explicit_full",
eta=0.0, s_noise=1., s_noise_substep=1., d_noise=1., alpha=-1.0, alpha_substep=-1.0, k=1.0, k_substep=1.0, c1=0.0, c2=0.5, c3=1.0, implicit_steps_diag=0, implicit_steps_full=0,
LGW_MASK_RESCALE_MIN=True, sigmas_override=None, sampler_mode="standard", epsilon_scales=None,regional_conditioning_weights=None, sde_noise=[],
extra_options="",
etas=None, etas_substep=None, s_noises=None, s_noises_substep=None, momentums=None, guides=None, cfgpp=0.0, cfg_cw = 1.0,regional_conditioning_floors=None, frame_weights_grp=None, eta_substep=0.0, noise_mode_sde_substep="hard",
noise_boost_step=0.0, noise_boost_substep=0.0, overshoot=0.0, overshoot_substep=0.0, overshoot_mode="hard", overshoot_mode_substep="hard", BONGMATH=True, noise_anchor=1.0,
implicit_type="predictor-corrector", implicit_type_substeps="predictor-corrector",
):
extra_args = {} if extra_args is None else extra_args
default_dtype = getattr(torch, get_extra_options_kv("default_dtype", "", extra_options), x.dtype)
model_device = x.device
work_device = 'cpu' if extra_options_flag("work_device_cpu", extra_options) else model_device
if hasattr(model.inner_model.inner_model.diffusion_model, "raw_x") and not extra_options_flag("ignore_raw_x", extra_options):
if model.inner_model.inner_model.diffusion_model.raw_x is not None:
x = model.inner_model.inner_model.diffusion_model.raw_x.clone()
del model.inner_model.inner_model.diffusion_model.raw_x
RESplain("Continuing from raw latent from previous sampler.", debug=False)
x = x.to(dtype=default_dtype, device=work_device)
sigmas = sigmas.to(dtype=default_dtype, device=work_device)
if hasattr(model.inner_model.inner_model.diffusion_model, "last_seed") and noise_seed < 0:
if model.inner_model.inner_model.diffusion_model.last_seed is not None:
noise_seed = model.inner_model.inner_model.diffusion_model.last_seed + 1
del model.inner_model.inner_model.diffusion_model.last_seed
RESplain("Updated noise_seed to:", noise_seed, " by continuing from last sampler.", debug=False)
if noise_seed < 0:
noise_seed = torch.initial_seed()+1
RESplain("Set noise_seed to:", noise_seed, " using torch.initial_seed()+1", debug=False)
c1 = float(get_extra_options_kv("c1", str(c1), extra_options))
c2 = float(get_extra_options_kv("c2", str(c2), extra_options))
c3 = float(get_extra_options_kv("c3", str(c3), extra_options))
guide_skip_steps = int(get_extra_options_kv("guide_skip_steps", 0, extra_options))
rk_swap_step = int(get_extra_options_kv("rk_swap_step", str(MAX_STEPS), extra_options))
rk_swap_print = extra_options_flag("rk_swap_print", extra_options)
rk_swap_threshold = float(get_extra_options_kv("rk_swap_threshold", "0.0", extra_options))
rk_swap_type = get_extra_options_kv("rk_swap_type", "", extra_options)
pseudoimplicit_step_weights = get_extra_options_list("pseudoimplicit_step_weights", "", extra_options).split(",")
if pseudoimplicit_step_weights[0]:
pseudoimplicit_step_weights = [float(pseudoimplicit_step_weights[_]) for _ in range(len(pseudoimplicit_step_weights))]
else:
pseudoimplicit_step_weights = [1. for _ in range(max(implicit_steps_diag, implicit_steps_full)+1)]
pseudoimplicit_row_weights = get_extra_options_list("pseudoimplicit_row_weights", "", extra_options).split(",")
if pseudoimplicit_row_weights[0]:
pseudoimplicit_row_weights = [float(pseudoimplicit_row_weights[_]) for _ in range(len(pseudoimplicit_row_weights))]
else:
pseudoimplicit_row_weights = [1. for _ in range(100)]
cfg_cw = float(get_extra_options_kv("cfg_cw", str(cfg_cw), extra_options))
# SETUP SAMPLER
if implicit_sampler_name not in ("use_explicit", "none"):
rk_type = implicit_sampler_name
RESplain("rk_type:", rk_type)
if implicit_sampler_name == "none":
implicit_steps_diag = implicit_steps_full = 0
RK = RK_Method_Beta.create(model, rk_type, noise_anchor, model_device=model_device, work_device=work_device, dtype=default_dtype, extra_options=extra_options)
RK.extra_args = RK.init_cfg_channelwise(x, cfg_cw, **extra_args)
RK.extra_args['model_options']['transformer_options']['regional_conditioning_weight'] = 0.0
RK.extra_args['model_options']['transformer_options']['regional_conditioning_floor'] = 0.0
# SETUP SIGMAS
NS = RK_NoiseSampler(RK, model, device=work_device, dtype=default_dtype, extra_options=extra_options)
sigmas, UNSAMPLE = NS.prepare_sigmas(sigmas, sigmas_override, d_noise, sampler_mode)
SDE_NOISE_EXTERNAL = False
if sde_noise is not None:
if len(sde_noise) > 0 and sigmas[1] > sigmas[2]:
SDE_NOISE_EXTERNAL = True
sigma_up_total = torch.zeros_like(sigmas[0])
for i in range(len(sde_noise)-1):
sigma_up_total += sigmas[i+1]
etas = torch.full_like(sigmas, eta / sigma_up_total)
NS.init_noise_samplers(x, noise_seed, noise_sampler_type, noise_sampler_type_substep, noise_mode_sde, noise_mode_sde_substep, overshoot_mode, overshoot_mode_substep, noise_boost_step, noise_boost_substep, alpha, alpha_substep, k, k_substep)
data_ = None
eps_ = None
eps = torch.zeros_like(x, dtype=default_dtype, device=work_device)
denoised = torch.zeros_like(x, dtype=default_dtype, device=work_device)
denoised_prev = torch.zeros_like(x, dtype=default_dtype, device=work_device)
denoised_prev2 = torch.zeros_like(x, dtype=default_dtype, device=work_device)
x_ = None
eps_prev_ = None
denoised_data_prev = None
denoised_data_prev2 = None
h_prev = None
# SETUP GUIDES
LG = LatentGuide(model, sigmas, UNSAMPLE, LGW_MASK_RESCALE_MIN, extra_options, device=work_device, dtype=default_dtype, frame_weights_grp=frame_weights_grp)
x = LG.init_guides(x, RK.IMPLICIT, guides, NS.noise_sampler)
if torch.norm(LG.mask - torch.ones_like(LG.mask)) != 0 and (LG.y0.sum() == 0 or LG.y0_inv.sum() == 0):
SKIP_PSEUDO = True
RESplain("skipping pseudo...")
if LG.y0.sum() == 0:
SKIP_PSEUDO_Y = "y0"
elif LG.y0_inv.sum() == 0:
SKIP_PSEUDO_Y = "y0_inv"
else:
SKIP_PSEUDO = False
if LG.y0.sum() != 0 and LG.y0_inv.sum() != 0:
denoised_prev = LG.mask * LG.y0 + (1-LG.mask) * LG.y0_inv
elif LG.y0.sum() != 0:
denoised_prev = LG.y0
elif LG.y0_inv.sum() != 0:
denoised_prev = LG.y0_inv
if extra_options_flag("pseudo_mix_strength", extra_options):
orig_y0 = LG.y0.clone()
orig_y0_inv = LG.y0_inv.clone()
gc.collect()
# BEGIN SAMPLING LOOP
num_steps = len(sigmas)-2 if sigmas[-1] == 0 else len(sigmas)-1
for step in trange(num_steps, disable=disable):
sigma, sigma_next = sigmas[step], sigmas[step+1]
if regional_conditioning_weights is not None:
RK.extra_args['model_options']['transformer_options']['regional_conditioning_weight'] = regional_conditioning_weights[step]
RK.extra_args['model_options']['transformer_options']['regional_conditioning_floor'] = regional_conditioning_floors [step]
epsilon_scale = float(epsilon_scales[step]) if epsilon_scales is not None else None
eta = etas [step] if etas is not None else eta
eta_substep = etas_substep [step] if etas_substep is not None else eta_substep
s_noise = s_noises [step] if s_noises is not None else s_noise
s_noise_substep = s_noises_substep[step] if s_noises_substep is not None else s_noise_substep
NS.set_sde_step(sigma, sigma_next, eta, overshoot, s_noise)
RK.set_coeff(rk_type, NS.h, c1, c2, c3, step, sigmas, NS.sigma_down)
NS.set_substep_list(RK)
recycled_stages = max(RK.multistep_stages, RK.hybrid_stages)
if step == 0 or step == guide_skip_steps:
x_, data_, eps_, eps_prev_ = (torch.zeros( RK.rows+2, *x.shape, dtype=default_dtype, device=work_device) for _ in range(4))
data_prev_ = torch.zeros(max(RK.rows+2, 4), *x.shape, dtype=default_dtype, device=work_device)
recycled_stages = len(data_prev_)-1
sde_noise_t = None
if SDE_NOISE_EXTERNAL:
if step >= len(sde_noise):
SDE_NOISE_EXTERNAL=False
else:
sde_noise_t = sde_noise[step]
x_[0] = x.clone()
# PRENOISE METHOD HERE!
x_0 = x_[0].clone()
# RECYCLE STAGES FOR MULTISTEP
if RK.multistep_stages > 0 or RK.hybrid_stages > 0:
for ms in range(len(eps_)):
eps_[ms] = RK.get_epsilon_anchored(x_0, data_prev_[ms], sigma)
eps_prev_ = eps_.clone()
# INITIALIZE IMPLICIT SAMPLING
if RK.IMPLICIT:
x_, eps_, data_ = init_implicit_sampling(RK, x_0, x_, eps_, eps_prev_, data_, eps, denoised, denoised_prev, denoised_prev2, step, sigmas, NS.h, NS.s_, extra_options)
# BEGIN FULLY IMPLICIT LOOP
for full_iter in range(implicit_steps_full + 1):
if RK.IMPLICIT:
x_, eps_ = RK.newton_iter(x_0, x_, eps_, eps_prev_, data_, NS.s_, 0, NS.h, sigmas, step, "init", extra_options)
# PREPARE FULLY PSEUDOIMPLICIT GUIDES
if step > 0 or not SKIP_PSEUDO:
if full_iter > 0 and extra_options_flag("fully_implicit_reupdate_x", extra_options):
x_[0] = NS.sigma_from_to(x_0, x, sigma, sigma_next, NS.s_[0])
x_0 = NS.sigma_from_to(x_0, x, sigma, sigma_next, sigma)
if extra_options_flag("fully_pseudo_init", extra_options) and full_iter == 0:
guide_mode_tmp = LG.guide_mode
LG.guide_mode = "fully_" + LG.guide_mode
x_0, x_, eps_ = LG.prepare_fully_pseudoimplicit_guides_substep(x_0, x_, eps_, eps_prev_, data_, denoised_prev, 0, step, sigmas, eta_substep, overshoot_substep, s_noise_substep, NS, RK, pseudoimplicit_row_weights, pseudoimplicit_step_weights, full_iter, BONGMATH, extra_options)
if extra_options_flag("fully_pseudo_init", extra_options) and full_iter == 0:
LG.guide_mode = guide_mode_tmp
# TABLEAU LOOP
for row in range(RK.rows - RK.multistep_stages - RK.row_offset + 1):
for diag_iter in range(implicit_steps_diag+1):
if noise_sampler_type_substep == "brownian" and (full_iter > 0 or diag_iter > 0):
eta_substep = 0.
NS.set_sde_substep(row, RK.multistep_stages, eta_substep, overshoot_substep, s_noise_substep, full_iter, diag_iter, implicit_steps_full, implicit_steps_diag)
# PRENOISE METHOD HERE!
# A-TABLEAU
if row < RK.rows:
# PREPARE PSEUDOIMPLICIT GUIDES
if step > 0 or not SKIP_PSEUDO:
x_0, x_, eps_, x_row_pseudoimplicit, sub_sigma_pseudoimplicit = LG.process_pseudoimplicit_guides_substep(x_0, x_, eps_, eps_prev_, data_, denoised_prev, row, step, sigmas, NS, RK, pseudoimplicit_row_weights, pseudoimplicit_step_weights, full_iter, BONGMATH, extra_options)
# PREPARE MODEL CALL
if LG.guide_mode in {"pseudoimplicit","pseudoimplicit_cw", "pseudoimplicit_projection", "pseudoimplicit_projection_cw","fully_pseudoimplicit", "fully_pseudoimplicit_projection","fully_pseudoimplicit_cw", "fully_pseudoimplicit_projection_cw"} and (step > 0 or not SKIP_PSEUDO) and (LG.lgw[step] > 0 or LG.lgw_inv[step] > 0) and x_row_pseudoimplicit is not None:
x_tmp = x_row_pseudoimplicit
s_tmp = sub_sigma_pseudoimplicit
# Fully implicit iteration (explicit only) # or... Fully implicit iteration (implicit only... not standard)
elif (full_iter > 0 and RK.row_offset == 1 and row == 0) or (full_iter > 0 and RK.row_offset == 0 and row == 0 and extra_options_flag("fully_implicit_update_x", extra_options)):
if extra_options_flag("fully_explicit_pogostick_eta", extra_options):
super_alpha_ratio, super_sigma_down, super_sigma_up = NS.get_sde_coeff(sigma, sigma_next, None, eta)
x = super_alpha_ratio * x + super_sigma_up * NS.noise_sampler(sigma=sigma_next, sigma_next=sigma)
x_tmp = x
s_tmp = sigma
elif extra_options_flag("enable_fully_explicit_lagrange_rebound1", extra_options):
substeps_prev = len(RK.C[:-1])
x_tmp = lagrange_interpolation(RK.C[1:-1], x_[1:substeps_prev], RK.C[0]).squeeze(0)
elif extra_options_flag("enable_fully_explicit_lagrange_rebound2", extra_options):
substeps_prev = len(RK.C[:-1])
x_tmp = lagrange_interpolation(RK.C[1:], x_[1:substeps_prev+1], RK.C[0]).squeeze(0)
elif extra_options_flag("enable_fully_explicit_rebound1", extra_options): # 17630, faded dots, just crap
eps_tmp, denoised_tmp = RK(x, sigma_next, x, sigma_next)
eps_tmp = (x - denoised_tmp) / sigma_next
x_[0] = denoised_tmp + sigma * eps_tmp
x_0 = x_[0]
x_tmp = x_[0]
s_tmp = sigma
elif implicit_type == "rebound":
eps_tmp, denoised_tmp = RK(x, sigma_next, x_0, sigma)
eps_tmp = (x - denoised_tmp) / sigma_next
x = denoised_tmp + sigma * eps_tmp
x_tmp = x
s_tmp = sigma
elif implicit_type == "bongmath" and (NS.sub_sigma_up > 0 or NS.sub_sigma_up_eta > 0):
if BONGMATH:
x_tmp = x_[row]
s_tmp = NS.s_[row]
else:
x_tmp = NS.sigma_from_to(x_0, x, sigma, sigma_next, sigma)
s_tmp = sigma
else:
x_tmp = x
s_tmp = sigma_next
# All others
else:
# three potential toggle options: force rebound/model call, force PC style, force pogostick style
if diag_iter > 0: # Diagonally implicit iteration (explicit or implicit)
if extra_options_flag("diag_explicit_pogostick_eta", extra_options):
super_alpha_ratio, super_sigma_down, super_sigma_up = NS.get_sde_coeff(NS.s_[row], NS.s_[row+RK.row_offset+RK.multistep_stages], None, eta)
x_[row+RK.row_offset] = super_alpha_ratio * x_[row+RK.row_offset] + super_sigma_up * NS.noise_sampler(sigma=NS.s_[row+RK.row_offset+RK.multistep_stages], sigma_next=NS.s_[row])
x_tmp = x_[row+RK.row_offset]
s_tmp = sigma
elif implicit_type_substeps == "rebound":
eps_[row], data_[row] = RK(x_[row+RK.row_offset], NS.s_[row+RK.row_offset+RK.multistep_stages], x_0, sigma)
x_ = RK.update_substep(x_0, x_, eps_, eps_prev_, row, RK.row_offset, NS.h_new, NS.h_new_orig, extra_options)
x_[row+RK.row_offset] = NS.rebound_overshoot_substep(x_0, x_[row+RK.row_offset])
x_[row+RK.row_offset] = NS.sigma_from_to(x_0, x_[row+RK.row_offset], sigma, NS.s_[row+RK.row_offset+RK.multistep_stages], NS.s_[row])
x_tmp = x_[row+RK.row_offset]
s_tmp = NS.s_[row]
elif implicit_type_substeps == "bongmath" and (NS.sub_sigma_up > 0 or NS.sub_sigma_up_eta > 0) and not extra_options_flag("disable_diag_explicit_bongmath_rebound", extra_options):
if BONGMATH:
x_tmp = x_[row]
s_tmp = NS.s_[row]
else:
x_tmp = NS.sigma_from_to(x_0, x_[row+RK.row_offset], sigma, NS.s_[row+RK.row_offset+RK.multistep_stages], NS.s_[row])
s_tmp = NS.s_[row]
else:
x_tmp = x_[row+RK.row_offset]
s_tmp = NS.s_[row+RK.row_offset+RK.multistep_stages]
else:
x_tmp = x_[row]
s_tmp = NS.sub_sigma
if RK.IMPLICIT:
if not extra_options_flag("disable_implicit_guide_preproc", extra_options):
eps_, x_ = LG.process_guides_substep(x_0, x_, eps_, data_, row, step, sigma, sigma_next, NS.sigma_down, NS.s_, epsilon_scale, RK, extra_options)
eps_prev_, x_ = LG.process_guides_substep(x_0, x_, eps_prev_, data_, row, step, sigma, sigma_next, NS.sigma_down, NS.s_, epsilon_scale, RK, extra_options)
if row == 0 and (extra_options_flag("implicit_lagrange_init", extra_options) or extra_options_flag("radaucycle", extra_options)):
pass
else:
x_[row+RK.row_offset] = x_0 + NS.h_new * RK.zum(row+RK.row_offset, eps_, eps_prev_)
x_[row+RK.row_offset] = NS.rebound_overshoot_substep(x_0, x_[row+RK.row_offset])
if row > 0:
x_[row+RK.row_offset] = NS.swap_noise_substep(x_0, x_[row+RK.row_offset])
if BONGMATH and step < sigmas.shape[0]-1 and not extra_options_flag("disable_implicit_prebong", extra_options):
x_0, x_, eps_ = RK.bong_iter(x_0, x_, eps_, eps_prev_, data_, sigma, NS.s_, row, RK.row_offset, NS.h, extra_options) # TRY WITH h_new ??
x_tmp = x_[row+RK.row_offset]
# MODEL CALL
if RK.IMPLICIT and row == 0 and (extra_options_flag("implicit_lazy_recycle_first_model_call_at_start", extra_options) or extra_options_flag("radaucycle", extra_options)):
pass
else:
if s_tmp == 0:
break
x_, eps_ = RK.newton_iter(x_0, x_, eps_, eps_prev_, data_, NS.s_, row, NS.h, sigmas, step, "pre", extra_options) # will this do anything? not x_tmp
eps_[row], data_[row] = RK(x_tmp, s_tmp, x_0, sigma)
if extra_options_flag("bong2m", extra_options) and RK.multistep_stages > 0 and step < len(sigmas)-4:
h_no_eta = -torch.log(sigmas[step+1]/sigmas[step])
h_prev1_no_eta = -torch.log(sigmas[step]/sigmas[step-1])
c2_prev = (-h_prev1_no_eta / h_no_eta).item()
eps_prev = denoised_data_prev - x_0
φ = Phi(h_prev, [0.,c2_prev])
a2_1 = c2_prev * φ(1,2)
for i in range(100):
x_prev = x_0 - h_prev * (a2_1 * eps_prev)
eps_prev = denoised_data_prev - x_prev
eps_[1] = eps_prev
if extra_options_flag("bong3m", extra_options) and RK.multistep_stages > 0 and step < len(sigmas)-10:
h_no_eta = -torch.log(sigmas[step+1]/sigmas[step])
h_prev1_no_eta = -torch.log(sigmas[step]/sigmas[step-1])
h_prev2_no_eta = -torch.log(sigmas[step]/sigmas[step-2])
c2_prev = (-h_prev1_no_eta / h_no_eta).item()
c3_prev = (-h_prev2_no_eta / h_no_eta).item()
eps_prev2 = denoised_data_prev2 - x_0
eps_prev = denoised_data_prev - x_0
φ = Phi(h_prev1_no_eta, [0.,c2_prev, c3_prev])
a2_1 = c2_prev * φ(1,2)
for i in range(100):
x_prev = x_0 - h_prev1_no_eta * (a2_1 * eps_prev)
eps_prev = denoised_data_prev2 - x_prev
eps_[1] = eps_prev
φ = Phi(h_prev2_no_eta, [0.,c3_prev, c3_prev])
def calculate_gamma(c2_prev, c3_prev):
return (3*(c3_prev**3) - 2*c3_prev) / (c2_prev*(2 - 3*c2_prev))
gamma = calculate_gamma(c2_prev, c3_prev)
a2_1 = c2_prev * φ(1,2)
a3_2 = gamma * c2_prev * φ(2,2) + (c3_prev ** 2 / c2_prev) * φ(2, 3)
a3_1 = c3_prev * φ(1,3) - a3_2
for i in range(100):
x_prev2 = x_0 - h_prev2_no_eta * (a3_1 * eps_prev + a3_2 * eps_prev2)
x_prev = x_prev2 + h_prev2_no_eta * (a2_1 * eps_prev)
eps_prev2 = denoised_data_prev - x_prev2
eps_prev = denoised_data_prev2 - x_prev
eps_[2] = eps_prev2
# GUIDE
#if not UNSAMPLE:
if not extra_options_flag("disable_guides_eps_substep", extra_options):
eps_, x_ = LG.process_guides_substep(x_0, x_, eps_, data_, row, step, NS.sigma, NS.sigma_next, NS.sigma_down, NS.s_, epsilon_scale, RK, extra_options)
if not extra_options_flag("disable_guides_eps_prev_substep", extra_options):
eps_prev_, x_ = LG.process_guides_substep(x_0, x_, eps_prev_, data_, row, step, NS.sigma, NS.sigma_next, NS.sigma_down, NS.s_, epsilon_scale, RK, extra_options)
if (full_iter == 0 and diag_iter == 0) or extra_options_flag("newton_iter_post_use_on_implicit_steps", extra_options):
x_, eps_ = RK.newton_iter(x_0, x_, eps_, eps_prev_, data_, NS.s_, row, NS.h, sigmas, step, "post", extra_options)
# UPDATE
x_ = RK.update_substep(x_0, x_, eps_, eps_prev_, row, RK.row_offset, NS.h_new, NS.h_new_orig, extra_options)
x_[row+RK.row_offset] = NS.rebound_overshoot_substep(x_0, x_[row+RK.row_offset])
if not RK.IMPLICIT and NS.noise_mode_sde_substep != "hard_sq":
x_[row+RK.row_offset] = NS.swap_noise_substep(x_0, x_[row+RK.row_offset])
if BONGMATH and NS.s_[row] > RK.sigma_min and NS.h < RK.sigma_max/2 and diag_iter == implicit_steps_diag and not extra_options_flag("disable_terminal_bongmath", extra_options):
if step == 0 and UNSAMPLE:
pass
else:
x_0, x_, eps_ = RK.bong_iter(x_0, x_, eps_, eps_prev_, data_, sigma, NS.s_, row, RK.row_offset, NS.h, extra_options)
x_next = x_[RK.rows - RK.multistep_stages - RK.row_offset + 1]
x_next = NS.rebound_overshoot_step(x_0, x_next)
eps = (x_0 - x_next) / (sigma - sigma_next)
denoised = x_0 - sigma * eps
x_next = LG.process_guides_poststep(x_next, denoised, eps, step, extra_options)
x = NS.swap_noise_step(x_0, x_next)
preview_callback(x, eps, denoised, x_, eps_, data_, step, sigma, sigma_next, callback, extra_options)
h_prev = NS.h
x_prev = x_0
denoised_prev2 = denoised_prev
denoised_prev = denoised
data_prev_[0] = data_[0]
for ms in range(recycled_stages):
data_prev_[recycled_stages - ms] = data_prev_[recycled_stages - ms - 1]
rk_type = RK.swap_rk_type_at_step_or_threshold(x_0, data_prev_, NS.sigma_down, sigmas, step, RK, rk_swap_step, rk_swap_threshold, rk_swap_type, rk_swap_print)
denoised_data_prev2 = denoised_data_prev
denoised_data_prev = data_[0]
if SKIP_PSEUDO:
if SKIP_PSEUDO_Y == "y0":
LG.y0 = denoised
LG.HAS_LATENT_GUIDE_INV = True
else:
LG.y0_inv = denoised
LG.HAS_LATENT_GUIDE_INV = True
if extra_options_flag("pseudo_mix_strength", extra_options):
pseudo_mix_strength = float(get_extra_options_kv("pseudo_mix_strength", "0.0", extra_options))
LG.y0 = orig_y0 + pseudo_mix_strength * (denoised - orig_y0)
LG.y0_inv = orig_y0_inv + pseudo_mix_strength * (denoised - orig_y0_inv)
if sigmas[-1] == 0 and sigmas[-2] == NS.sigma_min:
eps, denoised = RK(x, NS.sigma_min, x, NS.sigma_min)
x = denoised
eps = eps.to(model_device)
denoised = denoised.to(model_device)
x = x.to(model_device)
if not (UNSAMPLE and sigmas[1] > sigmas[0]) and not extra_options_flag("preview_last_step_always", extra_options):
preview_callback(x, eps, denoised, x_, eps_, data_, step, sigma, sigma_next, callback, extra_options, FINAL_STEP=True)
model.inner_model.inner_model.diffusion_model.raw_x = x.clone()
model.inner_model.inner_model.diffusion_model.last_seed = NS.noise_sampler.last_seed
return x
def preview_callback(x, eps, denoised, x_, eps_, data_, step, sigma, sigma_next, callback, extra_options, FINAL_STEP=False):
if FINAL_STEP:
denoised_callback = denoised
elif extra_options_flag("eps_substep_preview", extra_options):
row_callback = int(get_extra_options_kv("eps_substep_preview", "0", extra_options))
denoised_callback = eps_[row_callback]
elif extra_options_flag("denoised_substep_preview", extra_options):
row_callback = int(get_extra_options_kv("denoised_substep_preview", "0", extra_options))
denoised_callback = data_[row_callback]
elif extra_options_flag("x_substep_preview", extra_options):
row_callback = int(get_extra_options_kv("x_substep_preview", "0", extra_options))
denoised_callback = x_[row_callback]
elif extra_options_flag("eps_preview", extra_options):
denoised_callback = eps
elif extra_options_flag("denoised_preview", extra_options):
denoised_callback = denoised
elif extra_options_flag("x_preview", extra_options):
denoised_callback = x
else:
denoised_callback = data_[0]
callback({'x': x, 'i': step, 'sigma': sigma, 'sigma_next': sigma_next, 'denoised': denoised_callback.to(torch.float32)}) if callback is not None else None
return