Spaces:
Runtime error
Runtime error
File size: 29,375 Bytes
68324dd |
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 |
import gradio as gr
import os
import PIL
from PIL import Image
from pathlib import Path
import numpy as np
import numpy.random as npr
from contextlib import nullcontext
import torch
import torchvision.transforms as tvtrans
from lib.cfg_helper import model_cfg_bank
from lib.model_zoo import get_model
from lib.model_zoo.ddim_vd import DDIMSampler_VD, DDIMSampler_VD_DualContext
from lib.model_zoo.ddim_dualcontext import DDIMSampler_DualContext
from lib.experiments.sd_default import color_adjust
n_sample_image = 2
n_sample_text = 4
cache_examples = True
class vd_inference(object):
def __init__(self, type='official'):
if type in ['dc', '2-flow']:
cfgm_name = 'vd_dc_noema'
sampler = DDIMSampler_DualContext
pth = 'pretrained/vd-dc.pth'
elif type in ['official', '4-flow']:
cfgm_name = 'vd_noema'
sampler = DDIMSampler_VD
pth = 'pretrained/vd-official.pth'
cfgm = model_cfg_bank()(cfgm_name)
net = get_model()(cfgm)
sd = torch.load(pth, map_location='cpu')
net.load_state_dict(sd, strict=False)
self.use_cuda = torch.cuda.is_available()
if self.use_cuda:
net.to('cuda')
self.model_name = cfgm_name
self.net = net
self.sampler = sampler(net)
def regularize_image(self, x):
BICUBIC = PIL.Image.Resampling.BICUBIC
if isinstance(x, str):
x = Image.open(x).resize([512, 512], resample=BICUBIC)
x = tvtrans.ToTensor()(x)
elif isinstance(x, PIL.Image.Image):
x = x.resize([512, 512], resample=BICUBIC)
x = tvtrans.ToTensor()(x)
elif isinstance(x, np.ndarray):
x = PIL.Image.fromarray(x).resize([512, 512], resample=BICUBIC)
x = tvtrans.ToTensor()(x)
elif isinstance(x, torch.Tensor):
pass
else:
assert False, 'Unknown image type'
assert (x.shape[1]==512) & (x.shape[2]==512), \
'Wrong image size'
if self.use_cuda:
x = x.to('cuda')
return x
def decode(self, z, xtype, ctype, color_adj='None', color_adj_to=None):
net = self.net
if xtype == 'image':
x = net.autokl_decode(z)
color_adj_flag = (color_adj!='None') and (color_adj is not None)
color_adj_simple = color_adj=='Simple'
color_adj_keep_ratio = 0.5
if color_adj_flag and (ctype=='vision'):
x_adj = []
for xi in x:
color_adj_f = color_adjust(ref_from=(xi+1)/2, ref_to=color_adj_to)
xi_adj = color_adj_f((xi+1)/2, keep=color_adj_keep_ratio, simple=color_adj_simple)
x_adj.append(xi_adj)
x = x_adj
else:
x = torch.clamp((x+1.0)/2.0, min=0.0, max=1.0)
x = [tvtrans.ToPILImage()(xi) for xi in x]
return x
elif xtype == 'text':
prompt_temperature = 1.0
prompt_merge_same_adj_word = True
x = net.optimus_decode(z, temperature=prompt_temperature)
if prompt_merge_same_adj_word:
xnew = []
for xi in x:
xi_split = xi.split()
xinew = []
for idxi, wi in enumerate(xi_split):
if idxi!=0 and wi==xi_split[idxi-1]:
continue
xinew.append(wi)
xnew.append(' '.join(xinew))
x = xnew
return x
def inference(self, xtype, cin, ctype, scale=7.5, n_samples=None, color_adj=None,):
net = self.net
sampler = self.sampler
ddim_steps = 50
ddim_eta = 0.0
if xtype == 'image':
n_samples = n_sample_image if n_samples is None else n_samples
elif xtype == 'text':
n_samples = n_sample_text if n_samples is None else n_samples
if ctype in ['prompt', 'text']:
c = net.clip_encode_text(n_samples * [cin])
u = None
if scale != 1.0:
u = net.clip_encode_text(n_samples * [""])
elif ctype in ['vision', 'image']:
cin = self.regularize_image(cin)
ctemp = cin*2 - 1
ctemp = ctemp[None].repeat(n_samples, 1, 1, 1)
c = net.clip_encode_vision(ctemp)
u = None
if scale != 1.0:
dummy = torch.zeros_like(ctemp)
u = net.clip_encode_vision(dummy)
if xtype == 'image':
h, w = [512, 512]
shape = [n_samples, 4, h//8, w//8]
z, _ = sampler.sample(
steps=ddim_steps,
shape=shape,
conditioning=c,
unconditional_guidance_scale=scale,
unconditional_conditioning=u,
xtype=xtype, ctype=ctype,
eta=ddim_eta,
verbose=False,)
x = self.decode(z, xtype, ctype, color_adj=color_adj, color_adj_to=cin)
return x
elif xtype == 'text':
n = 768
shape = [n_samples, n]
z, _ = sampler.sample(
steps=ddim_steps,
shape=shape,
conditioning=c,
unconditional_guidance_scale=scale,
unconditional_conditioning=u,
xtype=xtype, ctype=ctype,
eta=ddim_eta,
verbose=False,)
x = self.decode(z, xtype, ctype)
return x
def application_disensemble(self, cin, n_samples=None, level=0, color_adj=None,):
net = self.net
scale = 7.5
sampler = self.sampler
ddim_steps = 50
ddim_eta = 0.0
n_samples = n_sample_image if n_samples is None else n_samples
cin = self.regularize_image(cin)
ctemp = cin*2 - 1
ctemp = ctemp[None].repeat(n_samples, 1, 1, 1)
c = net.clip_encode_vision(ctemp)
u = None
if scale != 1.0:
dummy = torch.zeros_like(ctemp)
u = net.clip_encode_vision(dummy)
if level == 0:
pass
else:
c_glb = c[:, 0:1]
c_loc = c[:, 1: ]
u_glb = u[:, 0:1]
u_loc = u[:, 1: ]
if level == -1:
c_loc = self.remove_low_rank(c_loc, demean=True, q=50, q_remove=1)
u_loc = self.remove_low_rank(u_loc, demean=True, q=50, q_remove=1)
if level == -2:
c_loc = self.remove_low_rank(c_loc, demean=True, q=50, q_remove=2)
u_loc = self.remove_low_rank(u_loc, demean=True, q=50, q_remove=2)
if level == 1:
c_loc = self.find_low_rank(c_loc, demean=True, q=10)
u_loc = self.find_low_rank(u_loc, demean=True, q=10)
if level == 2:
c_loc = self.find_low_rank(c_loc, demean=True, q=2)
u_loc = self.find_low_rank(u_loc, demean=True, q=2)
c = torch.cat([c_glb, c_loc], dim=1)
u = torch.cat([u_glb, u_loc], dim=1)
h, w = [512, 512]
shape = [n_samples, 4, h//8, w//8]
z, _ = sampler.sample(
steps=ddim_steps,
shape=shape,
conditioning=c,
unconditional_guidance_scale=scale,
unconditional_conditioning=u,
xtype='image', ctype='vision',
eta=ddim_eta,
verbose=False,)
x = self.decode(z, 'image', 'vision', color_adj=color_adj, color_adj_to=cin)
return x
def find_low_rank(self, x, demean=True, q=20, niter=10):
if demean:
x_mean = x.mean(-1, keepdim=True)
x_input = x - x_mean
else:
x_input = x
u, s, v = torch.pca_lowrank(x_input, q=q, center=False, niter=niter)
ss = torch.stack([torch.diag(si) for si in s])
x_lowrank = torch.bmm(torch.bmm(u, ss), torch.permute(v, [0, 2, 1]))
if demean:
x_lowrank += x_mean
return x_lowrank
def remove_low_rank(self, x, demean=True, q=20, niter=10, q_remove=10):
if demean:
x_mean = x.mean(-1, keepdim=True)
x_input = x - x_mean
else:
x_input = x
u, s, v = torch.pca_lowrank(x_input, q=q, center=False, niter=niter)
s[:, 0:q_remove] = 0
ss = torch.stack([torch.diag(si) for si in s])
x_lowrank = torch.bmm(torch.bmm(u, ss), torch.permute(v, [0, 2, 1]))
if demean:
x_lowrank += x_mean
return x_lowrank
def application_dualguided(self, cim, ctx, n_samples=None, mixing=0.5, color_adj=None, ):
net = self.net
scale = 7.5
sampler = DDIMSampler_VD_DualContext(net)
ddim_steps = 50
ddim_eta = 0.0
n_samples = n_sample_image if n_samples is None else n_samples
ctemp0 = self.regularize_image(cim)
ctemp1 = ctemp0*2 - 1
ctemp1 = ctemp1[None].repeat(n_samples, 1, 1, 1)
cim = net.clip_encode_vision(ctemp1)
uim = None
if scale != 1.0:
dummy = torch.zeros_like(ctemp1)
uim = net.clip_encode_vision(dummy)
ctx = net.clip_encode_text(n_samples * [ctx])
utx = None
if scale != 1.0:
utx = net.clip_encode_text(n_samples * [""])
h, w = [512, 512]
shape = [n_samples, 4, h//8, w//8]
z, _ = sampler.sample_dc(
steps=ddim_steps,
shape=shape,
first_conditioning=[uim, cim],
second_conditioning=[utx, ctx],
unconditional_guidance_scale=scale,
xtype='image',
first_ctype='vision',
second_ctype='prompt',
eta=ddim_eta,
verbose=False,
mixed_ratio=(1-mixing), )
x = self.decode(z, 'image', 'vision', color_adj=color_adj, color_adj_to=ctemp0)
return x
def application_i2t2i(self, cim, ctx_n, ctx_p, n_samples=None, color_adj=None,):
net = self.net
scale = 7.5
sampler = DDIMSampler_VD_DualContext(net)
ddim_steps = 50
ddim_eta = 0.0
prompt_temperature = 1.0
n_samples = n_sample_image if n_samples is None else n_samples
ctemp0 = self.regularize_image(cim)
ctemp1 = ctemp0*2 - 1
ctemp1 = ctemp1[None].repeat(n_samples, 1, 1, 1)
cim = net.clip_encode_vision(ctemp1)
uim = None
if scale != 1.0:
dummy = torch.zeros_like(ctemp1)
uim = net.clip_encode_vision(dummy)
n = 768
shape = [n_samples, n]
zt, _ = sampler.sample(
steps=ddim_steps,
shape=shape,
conditioning=cim,
unconditional_guidance_scale=scale,
unconditional_conditioning=uim,
xtype='text', ctype='vision',
eta=ddim_eta,
verbose=False,)
ztn = net.optimus_encode([ctx_n])
ztp = net.optimus_encode([ctx_p])
ztn_norm = ztn / ztn.norm(dim=1)
zt_proj_mag = torch.matmul(zt, ztn_norm[0])
zt_perp = zt - zt_proj_mag[:, None] * ztn_norm
zt_newd = zt_perp + ztp
ctx_new = net.optimus_decode(zt_newd, temperature=prompt_temperature)
ctx_new = net.clip_encode_text(ctx_new)
ctx_p = net.clip_encode_text([ctx_p])
ctx_new = torch.cat([ctx_new, ctx_p.repeat(n_samples, 1, 1)], dim=1)
utx_new = net.clip_encode_text(n_samples * [""])
utx_new = torch.cat([utx_new, utx_new], dim=1)
cim_loc = cim[:, 1: ]
cim_loc_new = self.find_low_rank(cim_loc, demean=True, q=10)
cim_new = cim_loc_new
uim_new = uim[:, 1:]
h, w = [512, 512]
shape = [n_samples, 4, h//8, w//8]
z, _ = sampler.sample_dc(
steps=ddim_steps,
shape=shape,
first_conditioning=[uim_new, cim_new],
second_conditioning=[utx_new, ctx_new],
unconditional_guidance_scale=scale,
xtype='image',
first_ctype='vision',
second_ctype='prompt',
eta=ddim_eta,
verbose=False,
mixed_ratio=0.33, )
x = self.decode(z, 'image', 'vision', color_adj=color_adj, color_adj_to=ctemp0)
return x
vd_inference = vd_inference('official')
def main(mode,
image=None,
prompt=None,
nprompt=None,
pprompt=None,
color_adj=None,
disentanglement_level=None,
dual_guided_mixing=None,
seed=0,):
if seed<0:
seed = 0
np.random.seed(seed)
torch.manual_seed(seed+100)
if mode == 'Text-to-Image':
if (prompt is None) or (prompt == ""):
return None, None
with torch.no_grad():
rv = vd_inference.inference(
xtype = 'image',
cin = prompt,
ctype = 'prompt', )
return rv, None
elif mode == 'Image-Variation':
if image is None:
return None, None
with torch.no_grad():
rv = vd_inference.inference(
xtype = 'image',
cin = image,
ctype = 'vision',
color_adj = color_adj,)
return rv, None
elif mode == 'Image-to-Text':
if image is None:
return None, None
with torch.no_grad():
rv = vd_inference.inference(
xtype = 'text',
cin = image,
ctype = 'vision',)
return None, '\n'.join(rv)
elif mode == 'Text-Variation':
if prompt is None:
return None, None
with torch.no_grad():
rv = vd_inference.inference(
xtype = 'text',
cin = prompt,
ctype = 'prompt',)
return None, '\n'.join(rv)
elif mode == 'Disentanglement':
if image is None:
return None, None
with torch.no_grad():
rv = vd_inference.application_disensemble(
cin = image,
level = disentanglement_level,
color_adj = color_adj,)
return rv, None
elif mode == 'Dual-Guided':
if (image is None) or (prompt is None) or (prompt==""):
return None, None
with torch.no_grad():
rv = vd_inference.application_dualguided(
cim = image,
ctx = prompt,
mixing = dual_guided_mixing,
color_adj = color_adj,)
return rv, None
elif mode == 'Latent-I2T2I':
if (image is None) or (nprompt is None) or (nprompt=="") \
or (pprompt is None) or (pprompt==""):
return None, None
with torch.no_grad():
rv = vd_inference.application_i2t2i(
cim = image,
ctx_n = nprompt,
ctx_p = pprompt,
color_adj = color_adj,)
return rv, None
else:
assert False, "No such mode!"
def get_instruction(mode):
t2i_instruction = ["Generate image from text prompt."]
i2i_instruction = [
"Generate image conditioned on reference image.",
"Color Calibration provide an opinion to adjust image color according to reference image.", ]
i2t_instruction = ["Generate text from reference image."]
t2t_instruction = ["Generate text from reference text prompt. (Model insufficiently trained, thus results are still experimental)"]
dis_instruction = [
"Generate a variation of reference image that disentangled for semantic or style.",
"Color Calibration provide an opinion to adjust image color according to reference image.",
"Disentanglement level controls the level of focus towards semantic (-2, -1) or style (1 2). Level 0 serves as Image-Variation.", ]
dug_instruction = [
"Generate image from dual guidance of reference image and text prompt.",
"Color Calibration provide an opinion to adjust image color according to reference image.",
"Guidance Mixing provides linear balances between image and text context. (0 towards image, 1 towards text)", ]
iti_instruction = [
"Generate image variations via image-to-text, text-latent-editing, and then text-to-image. (Still under exploration)",
"Color Calibration provide an opinion to adjust image color according to reference image.",
"Input prompt that will be substract from text/text latent code.",
"Input prompt that will be added to text/text latent code.", ]
if mode == "Text-to-Image":
return '\n'.join(t2i_instruction)
elif mode == "Image-Variation":
return '\n'.join(i2i_instruction)
elif mode == "Image-to-Text":
return '\n'.join(i2t_instruction)
elif mode == "Text-Variation":
return '\n'.join(t2t_instruction)
elif mode == "Disentanglement":
return '\n'.join(dis_instruction)
elif mode == "Dual-Guided":
return '\n'.join(dug_instruction)
elif mode == "Latent-I2T2I":
return '\n'.join(iti_instruction)
#############
# Interface #
#############
if True:
img_output = gr.Gallery(label="Image Result").style(grid=n_sample_image)
txt_output = gr.Textbox(lines=4, label='Text Result', visible=False)
with gr.Blocks() as demo:
gr.HTML(
"""
<div style="position: relative; float: left; text-align: center; width: 60%; min-width:600px; height: 160px; margin: 20px 0 20px 20%;">
<h1 style="font-weight: 900; font-size: 3rem;">
Versatile Diffusion
</h1>
<br>
<h2 style="font-weight: 450; font-size: 1rem;">
We built <b>Versatile Diffusion (VD), the first unified multi-flow multimodal diffusion framework</b>, as a step towards <b>Universal Generative AI</b>.
VD can natively support image-to-text, image-variation, text-to-image, and text-variation,
and can be further extended to other applications such as
semantic-style disentanglement, image-text dual-guided generation, latent image-to-text-to-image editing, and more.
Future versions will support more modalities such as speech, music, video and 3D.
</h2>
<br>
<h3>Xingqian Xu, Atlas Wang, Eric Zhang, Kai Wang,
and <a href="https://www.humphreyshi.com/home">Humphrey Shi</a>
[<a href="https://arxiv.org/abs/2211.08332" style="color:blue;">arXiv</a>]
[<a href="https://github.com/SHI-Labs/Versatile-Diffusion" style="color:blue;">GitHub</a>]
</h3>
</div>
<div style="position: relative; float: right; width: 19.9%; min-width:200px; margin: 20px auto;">
<img src="https://huggingface.co/spaces/shi-labs/Versatile-Diffusion/resolve/main/assets/figures/share_instruction.png">
</div>
""")
mode_input = gr.Radio([
"Text-to-Image", "Image-Variation", "Image-to-Text", "Text-Variation",
"Disentanglement", "Dual-Guided", "Latent-I2T2I"], value='Text-to-Image', label="VD Flows and Applications")
instruction = gr.Textbox(get_instruction("Text-to-Image"), label='Info')
with gr.Row():
with gr.Column():
img_input = gr.Image(label='Image Input', visible=False)
txt_input = gr.Textbox(lines=4, placeholder="Input prompt...", label='Text Input')
ntxt_input = gr.Textbox(label='Remove Prompt', visible=False)
ptxt_input = gr.Textbox(label='Add Prompt', visible=False)
coladj_input = gr.Radio(["None", "Simple"], value='Simple', label="Color Calibration", visible=False)
dislvl_input = gr.Slider(-2, 2, value=0, step=1, label="Disentanglement level", visible=False)
dguide_input = gr.Slider(0, 1, value=0.5, step=0.01, label="Guidance Mixing", visible=False)
seed_input = gr.Number(100, label="Seed", precision=0)
btn = gr.Button("Run")
btn.click(
main,
inputs=[
mode_input,
img_input,
txt_input,
ntxt_input,
ptxt_input,
coladj_input,
dislvl_input,
dguide_input,
seed_input, ],
outputs=[img_output, txt_output])
with gr.Column():
img_output.render()
txt_output.render()
example_mode = [
"Text-to-Image",
"Image-Variation",
"Image-to-Text",
"Text-Variation",
"Disentanglement",
"Dual-Guided",
"Latent-I2T2I"]
def get_example(mode):
if mode == 'Text-to-Image':
case = [
['a dream of a village in china, by Caspar David Friedrich, matte painting trending on artstation HQ', 23],
['a beautiful grand nebula in the universe', 24],
['heavy arms gundam penguin mech', 25],
]
elif mode == "Image-Variation":
case = [
['assets/space.jpg', 'None', 26],
['assets/train.jpg', 'Simple', 27],
]
elif mode == "Image-to-Text":
case = [
['assets/boy_and_girl.jpg' , 28],
['assets/house_by_lake.jpg', 29],
]
elif mode == "Text-Variation":
case = [
['a dream of a village in china, by Caspar David Friedrich, matte painting trending on artstation HQ' , 32],
['a beautiful grand nebula in the universe' , 33],
['heavy arms gundam penguin mech', 34],
]
elif mode == "Disentanglement":
case = [
['assets/vermeer.jpg', 'Simple', -2, 30],
['assets/matisse.jpg', 'Simple', 2, 31],
]
elif mode == "Dual-Guided":
case = [
['assets/benz.jpg', 'cyberpunk 2077', 'Simple', 0.75, 22],
['assets/vermeer.jpg', 'a girl with a diamond necklace', 'Simple', 0.66, 21],
]
elif mode == "Latent-I2T2I":
case = [
['assets/ghibli.jpg', 'white house', 'tall castle', 'Simple', 20],
['assets/matisse.jpg', 'fruits and bottles on the table', 'flowers on the table', 'Simple', 21],
]
else:
raise ValueError
case = [[mode] + casei for casei in case]
return case
def get_example_iof(mode):
if mode == 'Text-to-Image':
inps = [txt_input, seed_input]
oups = [img_output]
fn = lambda m, x, y: \
main(mode=m, prompt=x, seed=y)[0]
elif mode == "Image-Variation":
inps = [img_input, coladj_input, seed_input]
oups = [img_output]
fn = lambda m, x, y, z: \
main(mode=m, image=x, color_adj=y, seed=z)[0]
elif mode == "Image-to-Text":
inps = [img_input, seed_input]
oups = [txt_output]
fn = lambda m, x, y: \
main(mode=m, image=x, seed=y)[1]
elif mode == "Text-Variation":
inps = [txt_input, seed_input]
oups = [txt_output]
fn = lambda m, x, y: \
main(mode=m, prompt=x, seed=y)[1]
elif mode == "Disentanglement":
inps = [img_input, coladj_input, dislvl_input, seed_input]
oups = [img_output]
fn = lambda m, x, y, z, w: \
main(mode=m, image=x, color_adj=y, disentanglement_level=z, seed=w)[0]
elif mode == "Dual-Guided":
inps = [img_input, txt_input, coladj_input, dguide_input, seed_input]
oups = [img_output]
fn = lambda m, x, y, z, w, u: \
main(mode=m, image=x, prompt=y, color_adj=z, dual_guided_mixing=w, seed=u)[0]
elif mode == "Latent-I2T2I":
inps = [img_input, ntxt_input, ptxt_input, coladj_input, seed_input]
oups = [img_output]
fn = lambda m, x, y, z, w, u: \
main(mode=m, image=x, nprompt=y, pprompt=z, color_adj=w, seed=u)[0]
else:
raise ValueError
return [mode_input]+inps, oups, fn
with gr.Row():
for emode in example_mode[0:4]:
with gr.Column():
gr.Examples(
label=emode+' Examples',
examples=get_example(emode),
inputs=get_example_iof(emode)[0],
outputs=get_example_iof(emode)[1],
fn = get_example_iof(emode)[2],
cache_examples=cache_examples),
with gr.Row():
for emode in example_mode[4:7]:
with gr.Column():
gr.Examples(
label=emode+' Examples',
examples=get_example(emode),
inputs=get_example_iof(emode)[0],
outputs=get_example_iof(emode)[1],
fn = get_example_iof(emode)[2],
cache_examples=cache_examples),
mode_input.change(
fn=lambda x: gr.update(value=get_instruction(x)),
inputs=mode_input,
outputs=instruction,)
mode_input.change(
fn=lambda x: gr.update(visible=(x not in ['Text-to-Image', 'Text-Variation'])),
inputs=mode_input,
outputs=img_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x in ['Text-to-Image', 'Text-Variation', 'Dual-Guided'])),
inputs=mode_input,
outputs=txt_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x in ['Latent-I2T2I'])),
inputs=mode_input,
outputs=ntxt_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x in ['Latent-I2T2I'])),
inputs=mode_input,
outputs=ptxt_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x not in ['Text-to-Image', 'Image-to-Text', 'Text-Variation'])),
inputs=mode_input,
outputs=coladj_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x=='Disentanglement')),
inputs=mode_input,
outputs=dislvl_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x=='Dual-Guided')),
inputs=mode_input,
outputs=dguide_input,)
mode_input.change(
fn=lambda x: gr.update(visible=(x not in ['Image-to-Text', 'Text-Variation'])),
inputs=mode_input,
outputs=img_output,)
mode_input.change(
fn=lambda x: gr.update(visible=(x in ['Image-to-Text', 'Text-Variation'])),
inputs=mode_input,
outputs=txt_output,)
gr.HTML(
"""
<div style="text-align: center; max-width: 1200px; margin: 20px auto;">
<h3>
<b>Caution</b>:
We would like the raise the awareness of users of this demo of its potential issues and concerns.
Like previous large foundation models, Versatile Diffusion could be problematic in some cases, partially due to the imperfect training data and pretrained network (VAEs / context encoders) with limited scope.
In its future research phase, VD may do better on tasks such as text-to-image, image-to-text, etc., with the help of more powerful VAEs, more sophisticated network designs, and more cleaned data.
So far, we keep all features available for research testing both to show the great potential of the VD framework and to collect important feedback to improve the model in the future.
We welcome researchers and users to report issues with the HuggingFace community discussion feature or email the authors.
</h3>
<br>
<h3>
<b>Biases and content acknowledgement</b>:
Beware that VD may output content that reinforces or exacerbates societal biases, as well as realistic faces, pornography, and violence.
VD was trained on the LAION-2B dataset, which scraped non-curated online images and text, and may contained unintended exceptions as we removed illegal content.
VD in this demo is meant only for research purposes.
</h3>
</div>
""")
# demo.launch(share=True)
demo.launch(debug=True)
|