File size: 30,286 Bytes
c25e2cc
3cdacdf
078a9c5
3cdacdf
5cb2cc0
3cdacdf
6255790
7e3c69d
6255790
5e25b83
6255790
c633c03
 
 
c25e2cc
c633c03
 
 
 
 
 
96afb60
c633c03
dbc34e0
 
c633c03
 
 
 
0e70f76
c633c03
 
 
 
 
 
 
 
6255790
 
 
 
 
 
 
 
 
 
 
 
 
e79152d
 
6255790
 
 
09eb804
6255790
 
6a2fd41
6255790
 
 
 
 
e79152d
 
6255790
 
 
 
 
9cd2450
a44c2bb
 
 
 
7856099
a44c2bb
 
f02e93c
7856099
a44c2bb
 
 
 
 
 
 
 
 
 
27ec4ac
6a5a59b
a44c2bb
7593f04
 
a44c2bb
 
 
05f89f0
017df60
05f89f0
a44c2bb
7856099
a44c2bb
9cd2450
add968e
a44c2bb
9df32ef
a44c2bb
7593f04
db41b3f
a217b80
 
 
7593f04
5d818c6
a44c2bb
6255790
c633c03
a44c2bb
017df60
a44c2bb
 
257ea11
de47faa
257ea11
 
 
 
 
cbeefd9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b4d4a0c
c633c03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9156300
88f076f
 
a44c2bb
7e3c69d
48a3bb9
db50056
660a4aa
afa6f27
3489b04
48a3bb9
87c9d81
48a3bb9
 
 
87c9d81
48a3bb9
 
 
 
 
1ff3548
 
48a3bb9
3489b04
7bb8383
a9bc524
 
 
a44c2bb
 
 
a9bc524
a44c2bb
 
a9bc524
a44c2bb
a9bc524
a44c2bb
a9bc524
a44c2bb
 
 
a9bc524
 
a44c2bb
a9bc524
a44c2bb
 
a9bc524
 
a44c2bb
a9bc524
a44c2bb
a9bc524
 
 
 
23c49e0
a44c2bb
c6ff99b
 
8c532f2
c6ff99b
cbeefd9
8c532f2
cbeefd9
 
 
8c532f2
cbeefd9
8c532f2
cbeefd9
 
 
8c532f2
cbeefd9
8c532f2
cbeefd9
 
 
8c532f2
cbeefd9
3718eb4
 
 
 
 
 
 
 
 
8c532f2
c6ff99b
a44c2bb
017df60
 
 
6d75eb4
a44c2bb
 
 
 
8c532f2
 
7e3c69d
 
5d818c6
89ef3d2
3718eb4
 
 
a44c2bb
7bb8383
017df60
 
7856099
a44c2bb
 
65837e4
d76aee0
a44c2bb
3718eb4
77d316c
f948a49
6a5a59b
d58e1aa
e2d7d01
c2870fd
e2d7d01
77d316c
8c532f2
3718eb4
 
 
 
 
 
 
 
8c532f2
 
3718eb4
 
 
8c532f2
 
 
 
5d818c6
a44c2bb
5d818c6
cbeefd9
 
 
330dade
cbeefd9
87c9d81
cbeefd9
 
c6ff99b
cbeefd9
 
 
8575e70
cbeefd9
 
65837e4
c6ff99b
33984db
cbeefd9
33984db
c6ff99b
cbeefd9
8c532f2
 
 
 
 
 
cbeefd9
8c532f2
 
a44c2bb
8575e70
cbeefd9
 
33984db
c6ff99b
7e3c69d
c6ff99b
33984db
c6ff99b
cbeefd9
8c532f2
c6ff99b
8c532f2
 
 
cbeefd9
8c532f2
a44c2bb
8c532f2
cbeefd9
 
8c532f2
33984db
7e3c69d
33984db
 
 
cbeefd9
8c532f2
 
 
 
 
cbeefd9
8c532f2
 
 
3718eb4
a44c2bb
cbeefd9
6a5a59b
7b3a214
a44c2bb
7bb8383
3718eb4
cbeefd9
7bb8383
b885715
a44c2bb
3718eb4
a44c2bb
c6ff99b
cbeefd9
 
 
 
 
 
 
 
 
6d75eb4
a44c2bb
cbeefd9
 
3718eb4
 
a44c2bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbeefd9
 
 
 
 
a44c2bb
cbeefd9
 
 
8c532f2
 
6a5a59b
8c532f2
a44c2bb
8c532f2
6a5a59b
a44c2bb
7593f04
 
a44c2bb
 
 
017df60
 
 
a44c2bb
9cd2450
3522127
8c532f2
7b3a214
a44c2bb
 
 
7b3a214
 
 
257ea11
 
 
 
cbeefd9
9cd2450
7b3a214
cbeefd9
3718eb4
017df60
89ef3d2
6a5a59b
6d75eb4
017df60
 
a44c2bb
8c532f2
a44c2bb
40c1596
a44c2bb
6d75eb4
 
a44c2bb
 
 
6d75eb4
 
 
a44c2bb
6d75eb4
 
5d818c6
8c532f2
cbeefd9
 
 
a44c2bb
 
 
 
 
6207a5b
a44c2bb
 
d8ed7f3
6d75eb4
a44c2bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6d75eb4
a44c2bb
 
 
7e3c69d
a44c2bb
 
 
7e3c69d
a44c2bb
cbeefd9
 
3718eb4
a44c2bb
 
7e3c69d
a44c2bb
3718eb4
cbeefd9
3718eb4
a44c2bb
 
 
 
 
 
 
 
 
 
 
 
 
 
7e3c69d
a44c2bb
 
 
 
 
 
7e3c69d
a44c2bb
 
7e3c69d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7f61c74
3489b04
 
a44c2bb
a93910d
b30a076
3489b04
 
 
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
import gradio as gr
import torch
import numpy as np
import requests
import random
from io import BytesIO
from utils import *
from constants import *
from inversion_utils import *
from modified_pipeline_semantic_stable_diffusion import SemanticStableDiffusionPipeline
from torch import autocast, inference_mode
from diffusers import StableDiffusionPipeline
from diffusers import DDIMScheduler
from transformers import AutoProcessor, BlipForConditionalGeneration

# load pipelines
sd_model_id = "stabilityai/stable-diffusion-2-base"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
sd_pipe = StableDiffusionPipeline.from_pretrained(sd_model_id).to(device)
sd_pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler")
sem_pipe = SemanticStableDiffusionPipeline.from_pretrained(sd_model_id).to(device)
blip_processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(device)



## IMAGE CPATIONING ##
def caption_image(input_image):

  inputs = blip_processor(images=input_image, return_tensors="pt").to(device)
  pixel_values = inputs.pixel_values

  generated_ids = blip_model.generate(pixel_values=pixel_values, max_length=50)
  generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
  return generated_caption


## DDPM INVERSION AND SAMPLING ##
def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):

  #  inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf, 
  #  based on the code in https://github.com/inbarhub/DDPM_inversion
   
  #  returns wt, zs, wts:
  #  wt - inverted latent
  #  wts - intermediate inverted latents
  #  zs - noise maps

  sd_pipe.scheduler.set_timesteps(num_diffusion_steps)

  # vae encode image
  with autocast("cuda"), inference_mode():
      w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215).float()

  # find Zs and wts - forward process
  wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=True, num_inference_steps=num_diffusion_steps)
  return zs, wts


def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):

    # reverse process (via Zs and wT)
    w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=True, zs=zs[skip:])
    
    # vae decode image
    with autocast("cuda"), inference_mode():
        x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
    if x0_dec.dim()<4:
        x0_dec = x0_dec[None,:,:,:]
    img = image_grid(x0_dec)
    return img


def reconstruct(tar_prompt,
                tar_cfg_scale,
                skip,
                wts, zs,
                do_reconstruction,
                reconstruction,
                reconstruct_button
               ):

    if reconstruct_button == "Hide Reconstruction":
      return reconstruction.value, reconstruction, ddpm_edited_image.update(visible=False), do_reconstruction, "Show Reconstruction"

    else:
      if do_reconstruction:
          reconstruction_img = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
          reconstruction = gr.State(value=reconstruction_img)
          do_reconstruction = False
      return reconstruction.value, reconstruction, ddpm_edited_image.update(visible=True), do_reconstruction, "Hide Reconstruction"


def load_and_invert(
                    input_image,
                    do_inversion,
                    seed, randomize_seed,
                    wts, zs,
                    src_prompt ="",
                    tar_prompt="",
                    steps=100,
                    src_cfg_scale = 3.5,
                    skip=36,
                    tar_cfg_scale=15,
                    progress=gr.Progress(track_tqdm=True)

):


    x0 = load_512(input_image, device=device)

    if do_inversion or randomize_seed:
        # invert and retrieve noise maps and latent
        zs_tensor, wts_tensor = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
        wts = gr.State(value=wts_tensor)
        zs = gr.State(value=zs_tensor)
        do_inversion = False

    return wts, zs, do_inversion, inversion_progress.update(visible=False)

## SEGA ##

def edit(input_image,
            wts, zs,
            tar_prompt,
            steps,
            skip,
            tar_cfg_scale,
            edit_concept_1,edit_concept_2,edit_concept_3,
            guidnace_scale_1,guidnace_scale_2,guidnace_scale_3,
            warmup_1, warmup_2, warmup_3,
            neg_guidance_1, neg_guidance_2, neg_guidance_3,
            threshold_1, threshold_2, threshold_3,
         do_reconstruction,
         reconstruction):


    if edit_concept_1 != "" or edit_concept_2 != "" or edit_concept_3 != "":
      editing_args = dict(
      editing_prompt = [edit_concept_1,edit_concept_2,edit_concept_3],
      reverse_editing_direction = [ neg_guidance_1, neg_guidance_2, neg_guidance_3,],
      edit_warmup_steps=[warmup_1, warmup_2, warmup_3,],
      edit_guidance_scale=[guidnace_scale_1,guidnace_scale_2,guidnace_scale_3],
      edit_threshold=[threshold_1, threshold_2, threshold_3],
      edit_momentum_scale=0.3,
      edit_mom_beta=0.6,
      eta=1,)

      latnets = wts.value[skip].expand(1, -1, -1, -1)
      sega_out = sem_pipe(prompt=tar_prompt, latents=latnets, guidance_scale = tar_cfg_scale,
                          num_images_per_prompt=1,
                          num_inference_steps=steps,
                          use_ddpm=True,  wts=wts.value, zs=zs.value[skip:], **editing_args)
      
      return sega_out.images[0], reconstruct_button.update(visible=True), do_reconstruction, reconstruction
    
    else: # if sega concepts were not added, performs regular ddpm sampling
      
      if do_reconstruction: # if ddpm sampling wasn't computed
          pure_ddpm_img = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
          reconstruction = gr.State(value=pure_ddpm_img)
          do_reconstruction = False
          return pure_ddpm_img, reconstruct_button.update(visible=False), do_reconstruction, reconstruction
      
      return reconstruction, reconstruct_button.update(visible=False), do_reconstruction, reconstruction
        

def randomize_seed_fn(seed, randomize_seed):
    if randomize_seed:
        seed = random.randint(0, np.iinfo(np.int32).max)
    torch.manual_seed(seed)
    return seed

    
def get_example():
    case = [
        [
            'examples/source_a_cat_sitting_next_to_a_mirror.jpeg', 
            'a cat sitting next to a mirror',
            'watercolor painting of a cat sitting next to a mirror',
            100,
            36,
            15,
            'Schnauzer dog', 'cat',
            5.5,
            1,
            'examples/ddpm_sega_watercolor_painting_a_cat_sitting_next_to_a_mirror_plus_dog_minus_cat.png'
             ],
        [
            'examples/source_a_man_wearing_a_brown_hoodie_in_a_crowded_street.jpeg', 
            'a man wearing a brown hoodie in a crowded street',
            'a robot wearing a brown hoodie in a crowded street',
            100,
            36,
            15,
            'painting','',
            10,
            1,
            'examples/ddpm_sega_painting_of_a_robot_wearing_a_brown_hoodie_in_a_crowded_street.png'
             ],
    [
            'examples/source_wall_with_framed_photos.jpeg', 
            '',
            '',
            100,
            36,
            15,
            'pink drawings of muffins','',
            10,
            1,
            'examples/ddpm_sega_plus_pink_drawings_of_muffins.png'
             ],
    [
            'examples/source_an_empty_room_with_concrete_walls.jpg', 
            'an empty room with concrete walls',
            'glass walls',
            100,
            36,
            17,
            'giant elephant','',
            10,
            1,
            'examples/ddpm_sega_glass_walls_gian_elephant.png'
             ]]
    return case



########
# demo #
########


intro = """
<h1 style="font-weight: 1400; text-align: center; margin-bottom: 7px;">
   LEDITS
</h1>
<h2 style="font-weight: 800; text-align: center; margin-bottom: 7px;">
   Real Image Latent Editing with Edit Friendly DDPM and Semantic Guidance
</h2>

<h3 style="font-weight: 600; text-align: center; margin-bottom: 7px;">
   This is a demo for LEDITS - a combined approach for real image editing introduced in:
   <a href="https://editing-images-project.hf.space/" style="text-decoration: underline;" target="_blank"> Project Page </a>
</h3>

 

<p style="font-size: 0.9rem; margin: 0rem; line-height: 1.2em; margin-top:1em">
For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
<a href="https://huggingface.co/spaces/editing-images/edit_friendly_ddpm_x_sega?duplicate=true">
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
<p/>"""

help_text = """
- **Getting Started - edit images with DDPM X SEGA:**

    The are 3 general setting options you can play with -

    1. **Pure DDPM Edit -** Describe the desired edited output image in detail
    2. **Pure SEGA Edit -** Keep the target prompt empty ***or*** with a description of the original image and add editing concepts for Semantic Gudiance editing
    3. **Combined -** Describe the desired edited output image in detail and add additional SEGA editing concepts on top
- **Getting Started - Tips**

    While the best approach depends on your editing objective and source image,  we can layout a few guiding tips to use as a starting point -

    1. **DDPM** is usually more suited for scene/style changes and major subject changes (for example ) while **SEGA** allows for more fine grained control, changes are more delicate, more suited for adding details (for example facial expressions and attributes, subtle style modifications, object adding/removing)
    2. The more you describe the scene in the target prompt (both the parts and details you wish to keep the same and those you wish to change), the better the result
    3. **Combining DDPM Edit with SEGA -**
    Try dividing your editing objective to more significant scene/style/subject changes and detail adding/removing and more moderate changes. Then describe the major changes in a detailed target prompt and add the more fine grained details as SEGA concepts.
    4. **Reconstruction:** Using an empty source prompt + target prompt will lead to a perfect reconstruction
- **Fidelity vs creativity**:

    Bigger values → more fidelity, smaller values → more creativity

    1. `Skip Steps`
    2. `Warmup` (SEGA)
    3. `Threshold`  (SEGA)

    Bigger values → more creativity, smaller values → more fidelity

    1. `Guidance Scale`
    2. `Concept Guidance Scale` (SEGA)
"""

with gr.Blocks(css="style.css") as demo:

    def add_concept(sega_concepts_counter):
      if sega_concepts_counter == 1:
        return row2.update(visible=True), row2_advanced.update(visible=True), row3.update(visible=False), row3_advanced.update(visible=False), add_concept_button.update(visible=True), 2
      else:
        return row2.update(visible=True), row2_advanced.update(visible=True), row3.update(visible=True), row3_advanced.update(visible=True), add_concept_button.update(visible=False), 3

    def update_display_concept_1(add_1, edit_concept_1, neg_guidance_1):
      if add_1 == 'Add' and edit_concept_1 != "":
        return box1.update(visible=True), edit_concept_1, concept_1.update(visible=True), edit_concept_1, guidnace_scale_1.update(visible=True), neg_guidance_1, "Clear"
      else: # remove
        return box1.update(visible=False),"", concept_1.update(visible=False), "", guidnace_scale_1.update(visible=False), False, "Add"

    def update_display_concept_2(add_2, edit_concept_2, neg_guidance_2):
      if add_2 == 'Add' and edit_concept_2 != "":
        return box2.update(visible=True), edit_concept_2, concept_2.update(visible=True),edit_concept_2, guidnace_scale_2.update(visible=True), neg_guidance_2, "Clear"
      else: # remove
        return box2.update(visible=False),"", concept_2.update(visible=False), "", guidnace_scale_2.update(visible=False), False, "Add"

    def update_display_concept_3(add_3, edit_concept_3, neg_guidance_3):
      if add_3 == 'Add'and edit_concept_3 != "":
        return box3.update(visible=True), edit_concept_3, concept_3.update(visible=True), edit_concept_3, guidnace_scale_3.update(visible=True), neg_guidance_3, "Clear"
      else: # remove
        return box3.update(visible=False), "", concept_3.update(visible=False), "", guidnace_scale_3.update(visible=False), False, "Add"

    def display_editing_options(run_button, clear_button, sega_tab):
      return run_button.update(visible=True), clear_button.update(visible=True), sega_tab.update(visible=True)

    # def update_gallery_display(prev_output_image, sega_edited_image):
    #   if prev_output_image is None:
    #     return sega_edited_image, gallery.update(visible=True), sega_edited_image
    #   else:
    #     return prev_output_image, gallery.update(visible=True), sega_edited_image



    def reset_do_inversion():
        do_inversion = True
        return do_inversion

    def reset_do_reconstruction():
      do_reconstruction = True
      return  do_reconstruction

    def update_inversion_progress_visibility(input_image, do_inversion):
      if do_inversion and not input_image is None:
          return inversion_progress.update(visible=True)
      else:
        return inversion_progress.update(visible=False)

    def undo():
      return


    gr.HTML(intro)
    wts = gr.State()
    zs = gr.State()
    reconstruction = gr.State()
    do_inversion = gr.State(value=True)
    do_reconstruction = gr.State(value=True)
    sega_concepts_counter = gr.State(1)



    with gr.Row():
        input_image = gr.Image(label="Input Image", interactive=True)
        ddpm_edited_image = gr.Image(label=f"DDPM Reconstructed Image", interactive=False, visible=False)
        sega_edited_image = gr.Image(label=f"DDPM + SEGA Edited Image", interactive=False)
        input_image.style(height=365, width=365)
        ddpm_edited_image.style(height=365, width=365)
        sega_edited_image.style(height=365, width=365)

    with gr.Row():
      with gr.Box(visible=False) as box1:
        concept_1 = gr.Button(visible=False)
        guidnace_scale_1 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
                                                  value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
                                                  step=0.5, interactive=True,visible=False)
      with gr.Box(visible=False) as box2:
       concept_2 = gr.Button(visible=False)
       guidnace_scale_2 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
                                                value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
                                                step=0.5, interactive=True,visible=False)
      with gr.Box(visible=False) as box3:
       concept_3 = gr.Button(visible=False)
       guidnace_scale_3 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
                                                value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
                                                step=0.5, interactive=True,visible=False)


    with gr.Row():
        inversion_progress = gr.Textbox(visible=False, label="Inversion progress")

    # with gr.Tabs() as tabs:
          # with gr.TabItem('1. Describe the desired output', id=0):
    with gr.Row().style(mobile_collapse=False, equal_height=True):
                tar_prompt = gr.Textbox(
                                label="Image Description",
                                # show_label=False,
                                max_lines=1, value="",
                                placeholder="Enter your target prompt", 
                            )
                # caption_button = gr.Button("Caption Image")
          # with gr.TabItem('2. Add SEGA edit concepts', id=1):
    intro_segs = gr.Markdown("Add SEGA Edit Concepts")
              # 1st SEGA concept
    with gr.Row().style(mobile_collapse=False, equal_height=True):
          with gr.Column(scale=3, min_width=100):
                  edit_concept_1 = gr.Textbox(
                                  label="Edit Concept",
                                  show_label=False,
                                  max_lines=1, value="",
                                  placeholder="Enter your 1st edit prompt",
                              )
          with gr.Column(scale=1, min_width=100):
                  neg_guidance_1 = gr.Checkbox(
                      label='Negative Guidance')

                  # guidnace_scale_1 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
                  #                              value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
                  #                              step=0.5, interactive=True)
          with gr.Column(scale=1, min_width=100):

                  add_1 = gr.Button('Add')

              # 2nd SEGA concept
    with gr.Row(visible=False).style(equal_height=True) as row2:
        with gr.Column(scale=3, min_width=100):
                  edit_concept_2 = gr.Textbox(
                                  label="Edit Concept",
                                  show_label=False,
                                  max_lines=1,
                                  placeholder="Enter your 2st edit prompt",
                              )
        with gr.Column(scale=1, min_width=100):
                  neg_guidance_2 = gr.Checkbox(
                      label='Negative Guidance',visible=True)
                  # guidnace_scale_2 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
                  #                              value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
                  #                              step=0.5, interactive=True)
        with gr.Column(scale=1, min_width=100):
                  add_2 = gr.Button('Add')

              # 3rd SEGA concept
    with gr.Row(visible=False).style(equal_height=True) as row3:
      with gr.Column(scale=3, min_width=100):
                 edit_concept_3 = gr.Textbox(
                                  label="Edit Concept",
                                  show_label=False,
                                  max_lines=1,
                                  placeholder="Enter your 3rd edit prompt",
                              )
      with gr.Column(scale=1, min_width=100):
                 neg_guidance_3 = gr.Checkbox(
                      label='Negative Guidance',visible=True)
                #  guidnace_scale_3 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
                #                                value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
                #                                step=0.5, interactive=True)
      with gr.Column(scale=1, min_width=100):
                 add_3 = gr.Button('Add')




    with gr.Row().style(mobile_collapse=False, equal_height=True):
                add_concept_button = gr.Button("+")


    with gr.Row():
        run_button = gr.Button("Edit", visible=True)
        

    with gr.Accordion("Advanced Options", open=False):
      with gr.Tabs() as tabs:

          with gr.TabItem('General options', id=2):
            with gr.Row():
                with gr.Column(min_width=100):
                   clear_button = gr.Button("Clear", visible=True)
                   src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True, placeholder="")
                   steps = gr.Number(value=100, precision=0, label="Num Diffusion Steps", interactive=True)
                   src_cfg_scale = gr.Number(value=3.5, label=f"Source Guidance Scale", interactive=True)
                   

                with gr.Column(min_width=100):
                    reconstruct_button = gr.Button("Show Reconstruction", visible=False)
                    skip = gr.Slider(minimum=0, maximum=60, value=36, label="Skip Steps", interactive=True)
                    tar_cfg_scale = gr.Slider(minimum=7, maximum=30,value=15, label=f"Guidance Scale", interactive=True)
                    seed = gr.Number(value=0, precision=0, label="Seed", interactive=True)
                    randomize_seed = gr.Checkbox(label='Randomize seed', value=False)

          with gr.TabItem('SEGA options', id=3) as sega_advanced_tab:
             # 1st SEGA concept
              with gr.Row().style(mobile_collapse=False, equal_height=True):
                  warmup_1 = gr.Slider(label='Warmup', minimum=0, maximum=50,
                                       value=DEFAULT_WARMUP_STEPS,
                                       step=1, interactive=True)
                  threshold_1 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99,
                                          value=DEFAULT_THRESHOLD, steps=0.01, interactive=True)

              # 2nd SEGA concept
              with gr.Row(visible=False) as row2_advanced:
                  warmup_2 = gr.Slider(label='Warmup', minimum=0, maximum=50,
                                       value=DEFAULT_WARMUP_STEPS,
                                       step=1, interactive=True)
                  threshold_2 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99,
                                          value=DEFAULT_THRESHOLD,
                                          steps=0.01, interactive=True)
              # 3rd SEGA concept
              with gr.Row(visible=False) as row3_advanced:
                  warmup_3 = gr.Slider(label='Warmup', minimum=0, maximum=50,
                                       value=DEFAULT_WARMUP_STEPS, step=1,
                                       interactive=True)
                  threshold_3 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99,
                                          value=DEFAULT_THRESHOLD, steps=0.01,
                                          interactive=True)

    # caption_button.click(
    #     fn = caption_image,
    #     inputs = [input_image],
    #     outputs = [tar_prompt]
    # )

    add_1.click(fn = update_display_concept_1, inputs=[add_1, edit_concept_1, neg_guidance_1],  outputs=[box1, concept_1, concept_1, edit_concept_1, guidnace_scale_1,neg_guidance_1, add_1])
    add_2.click(fn = update_display_concept_2, inputs=[add_2, edit_concept_2, neg_guidance_2],  outputs=[box2, concept_2, concept_2, edit_concept_2, guidnace_scale_2,neg_guidance_2, add_2])
    add_3.click(fn = update_display_concept_3, inputs=[add_3, edit_concept_3, neg_guidance_3],  outputs=[box3, concept_3, concept_3, edit_concept_3, guidnace_scale_3,neg_guidance_3, add_3])


    add_concept_button.click(fn = add_concept, inputs=sega_concepts_counter,
               outputs= [row2, row2_advanced, row3, row3_advanced, add_concept_button, sega_concepts_counter], queue = False)

    run_button.click(fn = update_inversion_progress_visibility, inputs =[input_image,do_inversion], outputs=[inversion_progress],queue=False).then(
        fn=load_and_invert,
        inputs=[input_image,
                do_inversion,
                seed, randomize_seed,
                wts, zs,
                src_prompt,
                tar_prompt,
                steps,
                src_cfg_scale,
                skip,
                tar_cfg_scale
        ],
        outputs=[wts, zs, do_inversion, inversion_progress],
    ).then(fn = update_inversion_progress_visibility, inputs =[input_image,do_inversion], outputs=[inversion_progress],queue=False).success(
        fn=edit,
        inputs=[input_image,
                wts, zs,
                tar_prompt,
                steps,
                skip,
                tar_cfg_scale,
                edit_concept_1,edit_concept_2,edit_concept_3,
                guidnace_scale_1,guidnace_scale_2,guidnace_scale_3,
                warmup_1, warmup_2, warmup_3,
                neg_guidance_1, neg_guidance_2, neg_guidance_3,
                threshold_1, threshold_2, threshold_3, do_reconstruction, reconstruction

        ],
        outputs=[sega_edited_image, reconstruct_button, do_reconstruction, reconstruction])
    # .success(fn=update_gallery_display, inputs= [prev_output_image, sega_edited_image], outputs = [gallery, gallery, prev_output_image])



    # Automatically start inverting upon input_image change
    input_image.change(
        fn = reset_do_inversion,
        outputs = [do_inversion],
        queue = False).then(fn = update_inversion_progress_visibility, inputs =[input_image,do_inversion],
                            outputs=[inversion_progress],queue=False).then(
        fn=load_and_invert,
        inputs=[input_image,
                do_inversion,
                seed, randomize_seed,
                wts, zs,
                src_prompt,
                tar_prompt,
                steps,
                src_cfg_scale,
                skip,
                tar_cfg_scale,
        ],
        # outputs=[ddpm_edited_image, wts, zs, do_inversion],
        outputs=[wts, zs, do_inversion, inversion_progress],
    ).then(fn = update_inversion_progress_visibility, inputs =[input_image,do_inversion],
           outputs=[inversion_progress],queue=False).then(fn = caption_image,
        inputs = [input_image],
        outputs = [tar_prompt]).then(
              lambda: reconstruct_button.update(visible=False),
              outputs=[reconstruct_button]).then(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction],
        queue = False)


    # Repeat inversion (and reconstruction) when these params are changed:
    src_prompt.change(
        fn = reset_do_inversion,
        outputs = [do_inversion], queue = False).then(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction], queue = False)

    steps.change(
        fn = reset_do_inversion,
        outputs = [do_inversion], queue = False).then(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction], queue = False)


    src_cfg_scale.change(
        fn = reset_do_inversion,
        outputs = [do_inversion], queue = False).then(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction], queue = False)

    # Repeat only reconstruction these params are changed:

    tar_prompt.change(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction], queue = False)

    tar_cfg_scale.change(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction], queue = False)

    skip.change(
        fn = reset_do_reconstruction,
        outputs = [do_reconstruction], queue = False)



    clear_components = [input_image,ddpm_edited_image,ddpm_edited_image,sega_edited_image, do_inversion,
                                   src_prompt, steps, src_cfg_scale, seed,
                                  tar_prompt, skip, tar_cfg_scale, reconstruct_button,reconstruct_button,
                                  edit_concept_1, guidnace_scale_1,guidnace_scale_1,warmup_1,  threshold_1, neg_guidance_1, concept_1, concept_1,
                                  edit_concept_2, guidnace_scale_2,guidnace_scale_2,warmup_2,  threshold_2, neg_guidance_2, concept_2, concept_2, row2, row2_advanced,
                                  edit_concept_3, guidnace_scale_3,guidnace_scale_3,warmup_3,  threshold_3, neg_guidance_3, concept_3,concept_3, row3, row3_advanced ]

    clear_components_output_vals = [None, None,ddpm_edited_image.update(visible=False), None, True,
                     "", DEFAULT_DIFFUSION_STEPS, DEFAULT_SOURCE_GUIDANCE_SCALE, DEFAULT_SEED,
                     "", DEFAULT_SKIP_STEPS, DEFAULT_TARGET_GUIDANCE_SCALE, reconstruct_button.update(value="Show Reconstruction"),reconstruct_button.update(visible=False),
                     "", DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,guidnace_scale_1.update(visible=False), DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD, DEFAULT_NEGATIVE_GUIDANCE, "", concept_1.update(visible=False),
                     "", DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,guidnace_scale_2.update(visible=False), DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD, DEFAULT_NEGATIVE_GUIDANCE, "", concept_2.update(visible=False), row2.update(visible=False), row2_advanced.update(visible=False),
                     "", DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,guidnace_scale_3.update(visible=False), DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD, DEFAULT_NEGATIVE_GUIDANCE, "",concept_3.update(visible=False), row3.update(visible=False), row3_advanced.update(visible=False)
                         ]


    clear_button.click(lambda: clear_components_output_vals, outputs =clear_components)

    reconstruct_button.click(lambda: ddpm_edited_image.update(visible=True), outputs=[ddpm_edited_image]).then(fn = reconstruct,
                inputs = [tar_prompt,
                tar_cfg_scale,
                skip,
                wts, zs,
                do_reconstruction,
                reconstruction,
                          reconstruct_button],
                outputs = [ddpm_edited_image,reconstruction, ddpm_edited_image, do_reconstruction, reconstruct_button])

    randomize_seed.change(
        fn = randomize_seed_fn,
        inputs = [seed, randomize_seed],
        outputs = [seed],
        queue = False)
    
    # gr.Examples(
    #     label='Examples',
    #     examples=get_example(),
    #     inputs=[input_image, src_prompt, tar_prompt, steps,
    #                 # src_cfg_scale,
    #                 skip,
    #                 tar_cfg_scale,
    #                 edit_concept_1,
    #                 edit_concept_2,
    #                 guidnace_scale_1,
    #                 warmup_1,
    #                 # neg_guidance,
    #                 sega_edited_image
    #            ],
    #     outputs=[sega_edited_image],
    #     # fn=edit,
    #     # cache_examples=True
    # )





demo.queue()
demo.launch(share=False)