File size: 37,750 Bytes
3dabe4a |
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 |
import logging
from os import environ
import modules.scripts as scripts
import gradio as gr
from functools import reduce
from scripts.incant_utils import plot_tools
from einops import rearrange
from scripts.ui_wrapper import UIWrapper
from modules import script_callbacks
from modules import extra_networks
from modules import prompt_parser
from modules import sd_hijack
from modules.script_callbacks import CFGDenoiserParams
from modules.processing import StableDiffusionProcessing
from modules import shared
import math
import torch
from torch.nn import functional as F
from torchvision.transforms import GaussianBlur
from warnings import warn
from typing import Callable, Dict, Optional
from collections import OrderedDict
logger = logging.getLogger(__name__)
logger.setLevel(environ.get("SD_WEBUI_LOG_LEVEL", logging.INFO))
"""
Unofficial implementation of algorithms in "Multi-Concept T2I-Zero: Tweaking Only The Text Embeddings and Nothing Else"
Also implements some "Reduce distortion in generation" algorithms from "Enhancing Semantic Fidelity in Text-to-Image Synthesis: Attention Regulation in Diffusion Models"
@misc{tunanyan2023multiconcept,
title={Multi-Concept T2I-Zero: Tweaking Only The Text Embeddings and Nothing Else},
author={Hazarapet Tunanyan and Dejia Xu and Shant Navasardyan and Zhangyang Wang and Humphrey Shi},
year={2023},
eprint={2310.07419},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@misc{zhang2024enhancing,
title={Enhancing Semantic Fidelity in Text-to-Image Synthesis: Attention Regulation in Diffusion Models},
author={Yang Zhang and Teoh Tze Tzun and Lim Wei Hern and Tiviatis Sim and Kenji Kawaguchi},
year={2024},
eprint={2403.06381},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Author: v0xie
GitHub URL: https://github.com/v0xie/sd-webui-incantations
"""
handles = []
token_indices = [0]
class T2I0StateParams:
def __init__(self):
self.attnreg: bool = False
self.ema_smoothing_factor: float = 2.0
self.step_start : int = 0
self.step_end : int = 25
self.token_count: int = 0
self.tokens: list[int] = [] # [0, 20]
self.window_size_period: int = 10 # [0, 20]
self.ctnms_alpha: float = 0.05 # [0., 1.] if abs value of difference between uncodition and concept-conditioned is less than this, then zero out the concept-conditioned values less than this
self.correction_threshold: float = 0.5 # [0., 1.]
self.correction_strength: float = 0.25 # [0., 1.) # larger bm is less volatile changes in momentum
self.strength = 1.0
self.width = None
self.height = None
self.dims = []
self.cbs_similarities: list = None # we can precompute this?
class T2I0ExtensionScript(UIWrapper):
def __init__(self):
self.cached_c = [None, None]
self.handles = []
# Extension title in menu UI
def title(self) -> str:
return "Multi T2I-Zero"
# Decide to show menu in txt2img or img2img
def show(self, is_img2img):
return scripts.AlwaysVisible
# Setup menu ui detail
def setup_ui(self, is_img2img) -> list:
with gr.Accordion('Multi-Concept T2I-Zero', open=False):
active = gr.Checkbox(value=False, default=False, label="Active", elem_id='t2i0_active')
step_start = gr.Slider(value=1, minimum=0, maximum=150, default=1, step=1, label="Step Start", elem_id='t2i0_step_start', info="Start applying the correction at this step. Set to > 1 if using EMA.")
step_end = gr.Slider(value=25, minimum=0, maximum=150, default=1, step=1, label="Step End", elem_id='t2i0_step_end')
with gr.Row():
tokens = gr.Textbox(visible=True, value="", label="Tokens", elem_id='t2i0_tokens', info="Comma separated list of indices of tokens to condition on. Leave empty to condition on all tokens. Example: For prompt 'A cat and a dog', 'A': 0, 'cat': 1, 'and': 2, 'a': 3, 'dog': 4")
with gr.Row():
window_size = gr.Slider(value = 2, minimum = 0, maximum = 100, step = 1, label="Correction by Similarities Window Size", elem_id = 't2i0_window_size', info="Exclude contribution of tokens with indices += this value from the current token index.")
correction_threshold = gr.Slider(value = 0.5, minimum = 0., maximum = 1.0, step = 0.01, label="CbS Score Threshold", elem_id = 't2i0_correction_threshold', info="Filter dimensions with similarity below this threshold")
correction_strength = gr.Slider(value = 0.0, minimum = 0.0, maximum = 2.0, step = 0.01, label="CbS Correction Strength", elem_id = 't2i0_correction_strength', info="The strength of the correction")
with gr.Row():
attnreg = gr.Checkbox(visible=False, value=False, default=False, label="Use Attention Regulation", elem_id='t2i0_use_attnreg')
ctnms_alpha = gr.Slider(value = 0.1, minimum = 0.0, maximum = 1.0, step = 0.01, label="Alpha for Cross-Token Non-Maximum Suppression", elem_id = 't2i0_ctnms_alpha', info="Contribution of the suppressed attention map, default 0.1")
ema_factor = gr.Slider(value=0.0, minimum=0.0, maximum=4.0, default=2.0, label="EMA Smoothing Factor", elem_id='t2i0_ema_factor', info="Based on method from [arXiv:2403.06381]")
active.do_not_save_to_config = True
attnreg.do_not_save_to_config = True
step_start.do_not_save_to_config = True
step_end.do_not_save_to_config = True
window_size.do_not_save_to_config = True
correction_threshold.do_not_save_to_config = True
correction_strength.do_not_save_to_config = True
attnreg.do_not_save_to_config = True
ctnms_alpha.do_not_save_to_config = True
ema_factor.do_not_save_to_config = True
tokens.do_not_save_to_config = True
self.infotext_fields = [
(active, lambda d: gr.Checkbox.update(value='T2I-0 Active' in d)),
#(attnreg, lambda d: gr.Checkbox.update(value='T2I-0 AttnReg' in d)),
(window_size, 'T2I-0 Window Size'),
(step_start, 'T2I-0 Step Start'),
(step_end, 'T2I-0 Step End'),
(correction_threshold, 'T2I-0 CbS Score Threshold'),
(correction_strength, 'T2I-0 CbS Correction Strength'),
(ctnms_alpha, 'T2I-0 CTNMS Alpha'),
(ema_factor, 'T2I-0 CTNMS EMA Smoothing Factor'),
(tokens, 'T2I-0 Tokens'),
]
self.paste_field_names = [
't2i0_active',
't2i0_attnreg',
't2i0_window_size',
't2i0_ctnms_alpha',
't2i0_correction_threshold',
't2i0_correction_strength'
't2i0_ema_factor',
't2i0_step_start',
't2i0_step_end',
't2i0_tokens'
]
return [active, attnreg, window_size, ctnms_alpha, correction_threshold, correction_strength, tokens, ema_factor, step_end, step_start]
def process_batch(self, p: StableDiffusionProcessing, *args, **kwargs):
self.t2i0_process_batch(p, *args, **kwargs)
def t2i0_process_batch(self, p: StableDiffusionProcessing, active, attnreg, window_size, ctnms_alpha, correction_threshold, correction_strength, tokens, ema_factor, step_end, step_start, *args, **kwargs):
active = getattr(p, "t2i0_active", active)
# use_attnreg = getattr(p, "t2i0_attnreg", attnreg)
ema_factor = getattr(p, "t2i0_ema_factor", ema_factor)
step_start = getattr(p, "t2i0_step_start", step_start)
step_end = getattr(p, "t2i0_step_end", step_end)
if active is False:
return
window_size = getattr(p, "t2i0_window_size", window_size)
ctnms_alpha = getattr(p, "t2i0_ctnms_alpha", ctnms_alpha)
correction_threshold = getattr(p, "t2i0_correction_threshold", correction_threshold)
correction_strength = getattr(p, "t2i0_correction_strength", correction_strength)
tokens = getattr(p, "t2i0_tokens", tokens)
p.extra_generation_params.update({
"T2I-0 Active": active,
#"T2I-0 AttnReg": attnreg,
"T2I-0 Window Size": window_size,
"T2I-0 Step Start": step_start,
"T2I-0 Step End": step_end,
"T2I-0 CbS Score Threshold": correction_threshold,
"T2I-0 CbS Correction Strength": correction_strength,
"T2I-0 CTNMS Alpha": ctnms_alpha,
"T2I-0 CTNMS EMA Smoothing Factor": ema_factor,
"T2I-0 Tokens": tokens,
})
self.create_hook(p, active, attnreg, window_size, ctnms_alpha, correction_threshold, correction_strength, tokens, ema_factor, step_end, step_start, p.width, p.height)
def parse_concept_prompt(self, prompt:str) -> list[str]:
"""
Separate prompt by comma into a list of concepts
TODO: parse prompt into a list of concepts using A1111 functions
>>> g = lambda prompt: self.parse_concept_prompt(prompt)
>>> g("")
[]
>>> g("apples")
['apples']
>>> g("apple, banana, carrot")
['apple', 'banana', 'carrot']
"""
if len(prompt) == 0:
return []
return [x.strip() for x in prompt.split(",")]
def create_hook(self, p, active, attnreg, window_size, ctnms_alpha, correction_threshold, correction_strength, tokens, ema_factor, step_end, step_start, width, height, *args, **kwargs):
# Sanity check
cross_attn_modules = self.get_cross_attn_modules()
if len(cross_attn_modules) == 0:
logger.error("No cross attention modules found, cannot run T2I-0")
return
if len(tokens) > 0:
try:
token_indices = [int(x) for x in tokens.split(",")]
except ValueError:
logger.error("Invalid token indices, must be comma separated integers")
raise
else:
token_indices = []
# Create a list of parameters for each concept
t2i0_params = []
#for _, strength in concept_conds:
params = T2I0StateParams()
params.attnreg = attnreg
params.ema_smoothing_factor = ema_factor
params.step_start = step_start
params.step_end = step_end
params.window_size_period = window_size
params.ctnms_alpha = ctnms_alpha
params.correction_threshold = correction_threshold
params.correction_strength = correction_strength
params.strength = 1.0
params.width = width
params.height = height
params.dims = [width, height]
params.token_count, _ = get_token_count(p.prompt, p.steps, True)
token_indices = [x+1 for x in token_indices if x >= 0 and x < params.token_count]
params.tokens = token_indices
t2i0_params.append(params)
# Use lambda to call the callback function with the parameters to avoid global variables
y = lambda params: self.on_cfg_denoiser_callback(params, t2i0_params)
# un = lambda params: self.unhook_callbacks()
# Hook callbacks
if ctnms_alpha > 0:
self.ready_hijack_forward(ctnms_alpha, width, height, ema_factor, step_start, step_end, token_indices, params.token_count)
logger.debug('Hooked callbacks')
script_callbacks.on_cfg_denoiser(y)
script_callbacks.on_script_unloaded(self.unhook_callbacks)
def postprocess_batch(self, p, *args, **kwargs):
self.t2i0_postprocess_batch(p, *args, **kwargs)
def t2i0_postprocess_batch(self, p, active, *args, **kwargs):
self.unhook_callbacks()
active = getattr(p, "t2i0_active", active)
if active is False:
return
def unhook_callbacks(self):
global handles
logger.debug('Unhooked callbacks')
cross_attn_modules = self.get_cross_attn_modules()
for module in cross_attn_modules:
self.remove_field_cross_attn_modules(module, 't2i0_last_attn_map')
self.remove_field_cross_attn_modules(module, 't2i0_step')
self.remove_field_cross_attn_modules(module, 't2i0_step_start')
self.remove_field_cross_attn_modules(module, 't2i0_step_end')
self.remove_field_cross_attn_modules(module, 't2i0_ema_factor')
self.remove_field_cross_attn_modules(module, 't2i0_ema')
self.remove_field_cross_attn_modules(module, 'plot_num')
self.remove_field_cross_attn_modules(module, 't2i0_tokens')
self.remove_field_cross_attn_modules(module, 't2i0_token_count')
self.remove_field_cross_attn_modules(module, 't2i0_to_v_map')
self.remove_field_cross_attn_modules(module.to_k, 't2i0_parent_module')
self.remove_field_cross_attn_modules(module.to_v, 't2i0_parent_module')
_remove_all_forward_hooks(module, 'cross_token_non_maximum_suppression')
# _remove_all_forward_hooks(module, 'cross_token_non_maximum_suppression_pre')
# _remove_all_forward_hooks(module.to_k, 't2i0_to_k_hook')
_remove_all_forward_hooks(module.to_v, 't2i0_to_v_hook')
script_callbacks.remove_current_script_callbacks()
def apply_attnreg(self, f, C, alpha, B, *args, **kwargs):
"""
Apply attention regulation on an embedding.
Args:
f (Tensor): The embedding tensor of shape (n, d).
C (list): Indices of selected tokens.
alpha (float): Attnreg strength.
B (float): Lagrange multiplier B > 0
gamma (int): Window size for the windowing function.
Returns:
Tensor: The corrected embedding tensor.
"""
n, d = f.shape
f_tilde = f.detach().clone() # Copy the embedding tensor
# for token_idx, c in enumerate(C):
# pass
return f_tilde
def correction_by_similarities(self, f, C, percentile, gamma, alpha, tokens=None, token_count=77):
"""
Apply the Correction by Similarities algorithm on embeddings.
Args:
f (Tensor): The embedding tensor of shape (n, d).
C (list): Indices of selected tokens.
percentile (float): Percentile to use for score threshold.
gamma (int): Window size for the windowing function.
alpha (float): Correction strength.
tokens (list): List of token indices to condition on (default is all tokens if empty list).
Returns:
Tensor: The corrected embedding tensor.
"""
if alpha == 0:
return f
n, d = f.shape
token_indices = tokens
min_idx = 1
max_idx = min(token_count+1, n)
if token_indices is None:
token_indices = list(range(min_idx, max_idx))
if token_indices == []:
token_indices = list(range(min_idx, max_idx))
else:
token_indices = [x+1 for x in token_indices if x >= 0 and x < n]
f_tilde = f.detach().clone() # Copy the embedding tensor
# Define a windowing function
def psi(c, gamma, n, dtype, device, min_idx, max_idx):
window = torch.zeros(n, dtype=dtype, device=device)
start = max(min_idx, c - gamma)
end = min(max_idx, c + gamma + 1)
window[start:end] = 1
return window
def threshold_filter(t, tau):
""" Threshold filter function
Filters product values below a threshold tau and normalizes them to leave only the most similar dimensions.
Arguments:
t: torch.Tensor - The tensor to threshold
tau: float - The threshold value
Returns:
bool: True if the value is above the threshold, False otherwise
"""
pass
for c in token_indices:
if c < 0 or c >= n:
continue
Sc = f[c] * f # Element-wise multiplication
# calculate score threshold to filter out values under score threshold
# often there is a huge difference between the max and min values, so we use a log-like function instead
k = 10
e= 2.718281
pct_max = 1/(1+1e-10)
pct_min = 1e-16
# max of 0.999... to 0.0000...1
pct = min(pct_max, max(pct_min, 1 - e**(-k * percentile)))
tau = torch.quantile(Sc, pct)
Sc_tilde = Sc * (Sc > tau) # Apply threshold and filter
Sc_tilde /= Sc_tilde.max() # Normalize
window = psi(c, gamma, n, Sc_tilde.dtype, Sc_tilde.device, min_idx, max_idx).unsqueeze(1) # Apply windowing function
Sc_tilde *= window
f_c_tilde = torch.sum(Sc_tilde * f, dim=0) # Combine embeddings
f_tilde[c] = (1 - alpha) * f[c] + alpha * f_c_tilde # Blend embeddings
return f_tilde
def ready_hijack_forward(self, alpha, width, height, ema_factor, step_start, step_end, tokens, token_count):
""" Create a hook to modify the output of the forward pass of the cross attention module
Arguments:
alpha: float - The strength of the CTNMS correction, default 0.1
width: int - The width of the final output image map
height: int - The height of the final output image map
ema_factor: float - EMA smoothing factor, default 2.0
step_start: int - Wait to apply CTNMS until this step
step_end: int - The number of steps to apply the CTNMS correction, after which don't
tokens: list[int] - List of token indices to condition on
token_count: int - The number of tokens in the prompt
Only modifies the output of the cross attention modules that get context (i.e. text embedding)
"""
cross_attn_modules = self.get_cross_attn_modules()
if len(cross_attn_modules) == 0:
logger.error("No cross attention modules found, cannot run T2I-0")
return
# add field for last_attn_map
plot_num = 0
for module in cross_attn_modules:
self.add_field_cross_attn_modules(module, 't2i0_last_attn_map', None)
self.add_field_cross_attn_modules(module, 't2i0_step', int(-1))
self.add_field_cross_attn_modules(module, 't2i0_step_start', int(step_start))
self.add_field_cross_attn_modules(module, 't2i0_step_end', int(step_end))
self.add_field_cross_attn_modules(module, 't2i0_ema', None)
self.add_field_cross_attn_modules(module, 't2i0_ema_factor', float(ema_factor))
self.add_field_cross_attn_modules(module, 'plot_num', int(plot_num))
self.add_field_cross_attn_modules(module, 't2i0_to_v_map', None)
self.add_field_cross_attn_modules(module.to_v, 't2i0_parent_module', [module])
self.add_field_cross_attn_modules(module, 't2i0_token_count', int(token_count))
self.add_field_cross_attn_modules(module, 'gaussian_blur', GaussianBlur(kernel_size=3, sigma=1).to(device=shared.device))
if tokens is not None:
self.add_field_cross_attn_modules(module, 't2i0_tokens', torch.tensor(tokens).to(device=shared.device, dtype=torch.int64))
else:
self.add_field_cross_attn_modules(module, 't2i0_tokens', None)
plot_num += 1
# def cross_token_non_maximum_suppression_pre(module, args, kwargs):
# pass
# pass
def cross_token_non_maximum_suppression(module, input, kwargs, output):
module.t2i0_step += 1
context = kwargs.get('context', None)
if context is None:
return
if context.shape[1] % 77 != 0:
logger.error("Context shape is not divisible by 77, cannot run T2I-0")
return
current_step = module.t2i0_step
start_step = module.t2i0_step_start
end_step = module.t2i0_step_end
# Select token indices, default is ALL tokens
token_count = module.t2i0_token_count
token_indices = module.t2i0_tokens
if current_step > end_step and end_step > 0:
return
if current_step < start_step:
return
batch_size, sequence_length, inner_dim = output.shape
max_dims = width*height
factor = math.isqrt(max_dims // sequence_length) # should be a square of 2
downscale_width = width // factor
downscale_height = height // factor
if downscale_width * downscale_height != sequence_length:
print(f"Error: Width: {width}, height: {height}, Downscale width: {downscale_width}, height: {downscale_height}, Factor: {factor}, Max dims: {max_dims}\n")
return
# h = module.heads
# head_dim = inner_dim // h
dtype = output.dtype
device = output.device
# Multiply text embeddings into visual embeddings
to_v_map = module.t2i0_to_v_map.detach().clone()
to_v_inner_dim = to_v_map.size(-2)
to_v_map = (to_v_map @ output.transpose(1, 2)).transpose(1, 2)
to_v_attention_map = to_v_map.view(batch_size, downscale_height, downscale_width, to_v_inner_dim)
# Original attention map
attention_map = output.view(batch_size, downscale_height, downscale_width, inner_dim)
if token_indices is None:
selected_tokens = torch.arange(1, token_count, device=output.device)
elif len(token_indices) == 0:
selected_tokens = torch.arange(1, token_count, device=output.device)
else:
selected_tokens = module.t2i0_tokens
if module.t2i0_ema is None:
module.t2i0_ema = output.detach().clone()
# Extract the attention maps for the selected tokens
AC = to_v_attention_map[:, :, :, selected_tokens] # Extracting relevant attention maps
# Extract and process the selected attention maps
# GaussianBlur expects the input [..., C, H, W]
gaussian_blur = module.gaussian_blur
AC = AC.permute(0, 3, 1, 2)
AC = gaussian_blur(AC) # Applying Gaussian smoothing
AC = AC.permute(0, 2, 3, 1)
# Find the maximum contributing token for each pixel
M = torch.argmax(AC, dim=-1)
one_hot_M = F.one_hot(M, num_classes=to_v_attention_map.size(-1)).to(dtype=dtype, device=device)
# the attention map is of shape [batch_size, height, width, inner_dim]
one_hot_M_z = rearrange(one_hot_M, 'b h w c -> b (h w) c')
one_hot_M_z = one_hot_M_z @ module.t2i0_to_v_map
one_hot_M_z = rearrange(one_hot_M_z, 'b (h w) c -> b h w c', h=downscale_height, w=downscale_width)
suppressed_attention_map = one_hot_M_z * attention_map
# Reshape back to original dimensions
suppressed_attention_map = suppressed_attention_map.view(batch_size, sequence_length, inner_dim)
# Calculate the EMA of the suppressed attention map
if module.t2i0_ema_factor > 0:
ema = module.t2i0_ema
ema_factor = module.t2i0_ema_factor / (1 + current_step)
# Add the suppressed attention map to the EMA
ema = ema_factor * ema + (1 - ema_factor) * suppressed_attention_map
module.t2i0_ema = ema
out_tensor = (1 -alpha) * output + (alpha) * ema
#out_tensor = (1-alpha) * ema + alpha * suppressed_attention_map
else:
out_tensor = (1-alpha) * output + alpha * suppressed_attention_map
return out_tensor
def t2i0_to_k_hook(module, input, kwargs, output):
pass
pass
def t2i0_to_v_hook(module, input, kwargs, output):
module.t2i0_parent_module[0].t2i0_to_v_map = output
# Hook
for module in cross_attn_modules:
# handle = module.to_k.register_forward_hook(t2i0_to_k_hook, with_kwargs=True)
handle = module.to_v.register_forward_hook(t2i0_to_v_hook, with_kwargs=True)
handle = module.register_forward_hook(cross_token_non_maximum_suppression, with_kwargs=True)
# handle = module.register_forward_pre_hook(cross_token_non_maximum_suppression_pre, with_kwargs=True)
def get_cross_attn_modules(self):
""" Get all cross attention modules """
try:
m = shared.sd_model
nlm = m.network_layer_mapping
cross_attn_modules = [m for m in nlm.values() if 'CrossAttention' in m.__class__.__name__ and 'attn2' in m.network_layer_name]
return cross_attn_modules
except AttributeError:
logger.exception("AttributeError while getting cross attention modules")
return []
except Exception:
logger.exception("Error while getting cross attention modules")
return []
def add_field_cross_attn_modules(self, module, field, value):
""" Add a field to a module if it doesn't exist """
if not hasattr(module, field):
setattr(module, field, value)
def remove_field_cross_attn_modules(self, module, field):
""" Remove a field from a module if it exists """
if hasattr(module, field):
delattr(module, field)
def on_cfg_denoiser_callback(self, params: CFGDenoiserParams, t2i0_params: list[T2I0StateParams]):
if isinstance(params.text_cond, dict):
text_cond = params.text_cond['crossattn'] # SD XL
else:
text_cond = params.text_cond # SD 1.5
sp = t2i0_params[0]
window_size = sp.window_size_period
correction_strength = sp.correction_strength
score_threshold = sp.correction_threshold
step = params.sampling_step
step_start = sp.step_start
step_end = sp.step_end
tokens = sp.tokens if sp.tokens is not None else []
if step_start > step:
return
if step > step_end:
return
for batch_idx, batch in enumerate(text_cond):
window = list(range(0, len(batch)))
f_bar = self.correction_by_similarities(batch, window, score_threshold, window_size, correction_strength, tokens)
if isinstance(params.text_cond, dict):
params.text_cond['crossattn'][batch_idx] = f_bar
else:
params.text_cond[batch_idx] = f_bar
return
def get_xyz_axis_options(self) -> dict:
xyz_grid = [x for x in scripts.scripts_data if x.script_class.__module__ in ("xyz_grid.py", "scripts.xyz_grid")][0].module
extra_axis_options = {
xyz_grid.AxisOption("[T2I-0] Active", str, t2i0_apply_override('t2i0_active', boolean=True), choices=xyz_grid.boolean_choice(reverse=True)),
xyz_grid.AxisOption("[T2I-0] Step Start", int, t2i0_apply_field("t2i0_step_start")),
xyz_grid.AxisOption("[T2I-0] Step End", int, t2i0_apply_field("t2i0_step_end")),
xyz_grid.AxisOption("[T2I-0] CbS Window Size", int, t2i0_apply_field("t2i0_window_size")),
xyz_grid.AxisOption("[T2I-0] CbS Score Threshold", float, t2i0_apply_field("t2i0_correction_threshold")),
xyz_grid.AxisOption("[T2I-0] CbS Correction Strength", float, t2i0_apply_field("t2i0_correction_strength")),
xyz_grid.AxisOption("[T2I-0] CTNMS Alpha", float, t2i0_apply_field("t2i0_ctnms_alpha")),
xyz_grid.AxisOption("[T2I-0] CTNMS EMA Smoothing Factor", float, t2i0_apply_field("t2i0_ema_factor")),
}
return extra_axis_options
def plot_attention_map(attention_map: torch.Tensor, title, x_label="X", y_label="Y", save_path=None, plot_type="default"):
""" Plots an attention map using matplotlib.pyplot
Arguments:
attention_map: Tensor - The attention map to plot
title: str - The title of the plot
x_label: str (optional) - The x-axis label
y_label: str (optional) - The y-axis label
save_path: str (optional) - The path to save the plot
Returns:
PIL.Image: The plot as a PIL image
"""
if attention_map.dim() == 3:
attention_map = attention_map.squeeze(0).mean(2)
plot_tools.plot_attention_map(attention_map, title, x_label, y_label, save_path, plot_type)
def debug_plot_attention_map(attention_map):
""" Plots an attention map using matplotlib.pyplot
Arguments:
attention_map: Tensor - The attention map to plot
title: str - The title of the plot
x_label: str (optional) - The x-axis label
y_label: str (optional) - The y-axis label
save_path: str (optional) - The path to save the plot
Returns:
PIL.Image: The plot as a PIL image
"""
plot_attention_map(
attention_map,
"Debug Output",
save_path="F:\\incant\\temp\\AAA_out_temp.png"
)
# XYZ Plot
# Based on @mcmonkey4eva's XYZ Plot implementation here: https://github.com/mcmonkeyprojects/sd-dynamic-thresholding/blob/master/scripts/dynamic_thresholding.py
def t2i0_apply_override(field, boolean: bool = False):
def fun(p, x, xs):
if boolean:
x = True if x.lower() == "true" else False
setattr(p, field, x)
return fun
def t2i0_apply_field(field):
def fun(p, x, xs):
if not hasattr(p, "t2i0_active"):
p.t2i0_active = True
setattr(p, field, x)
return fun
# taken from modules/ui.py
def get_token_count(text, steps, is_positive: bool = True):
try:
text, _ = extra_networks.parse_prompt(text)
if is_positive:
_, prompt_flat_list, _ = prompt_parser.get_multicond_prompt_list([text])
else:
prompt_flat_list = [text]
prompt_schedules = prompt_parser.get_learned_conditioning_prompt_schedules(prompt_flat_list, steps)
except Exception:
# a parsing error can happen here during typing, and we don't want to bother the user with
# messages related to it in console
prompt_schedules = [[[steps, text]]]
flat_prompts = reduce(lambda list1, list2: list1+list2, prompt_schedules)
prompts = [prompt_text for step, prompt_text in flat_prompts]
token_count, max_length = max([sd_hijack.model_hijack.get_prompt_lengths(prompt) for prompt in prompts], key=lambda args: args[0])
return token_count, max_length
# thanks torch; removing hooks DOESN'T WORK
# thank you to @ProGamerGov for this https://github.com/pytorch/pytorch/issues/70455
def _remove_all_forward_hooks(
module: torch.nn.Module, hook_fn_name: Optional[str] = None
) -> None:
"""
This function removes all forward hooks in the specified module, without requiring
any hook handles. This lets us clean up & remove any hooks that weren't property
deleted.
Warning: Various PyTorch modules and systems make use of hooks, and thus extreme
caution should be exercised when removing all hooks. Users are recommended to give
their hook function a unique name that can be used to safely identify and remove
the target forward hooks.
Args:
module (nn.Module): The module instance to remove forward hooks from.
hook_fn_name (str, optional): Optionally only remove specific forward hooks
based on their function's __name__ attribute.
Default: None
"""
if hook_fn_name is None:
warn("Removing all active hooks can break some PyTorch modules & systems.")
def _remove_hooks(m: torch.nn.Module, name: Optional[str] = None) -> None:
if hasattr(module, "_forward_hooks"):
if m._forward_hooks != OrderedDict():
if name is not None:
dict_items = list(m._forward_hooks.items())
m._forward_hooks = OrderedDict(
[(i, fn) for i, fn in dict_items if fn.__name__ != name]
)
else:
m._forward_hooks: Dict[int, Callable] = OrderedDict()
def _remove_child_hooks(
target_module: torch.nn.Module, hook_name: Optional[str] = None
) -> None:
for _, child in target_module._modules.items():
if child is not None:
_remove_hooks(child, hook_name)
_remove_child_hooks(child, hook_name)
# Remove hooks from target submodules
_remove_child_hooks(module, hook_fn_name)
# Remove hooks from the target module
_remove_hooks(module, hook_fn_name)
|