-
Notifications
You must be signed in to change notification settings - Fork 24
/
app.py
1131 lines (1005 loc) · 58.1 KB
/
app.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
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import argparse
import base64
import io
import os
import sys
import re
import cv2
import gradio as gr
import numpy as np
import requests
from functools import partial
from PIL import Image, ImageOps
from app_utils import ImageBoxState, bbox_draw, open_image, mask_to_bbox
import imageio
import tempfile
from omegaconf import OmegaConf
import torch
from diffusers import I2VGenXLPipeline
from diffusers.utils import export_to_gif, export_to_video
from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
from urllib.request import urlopen
from PIL import Image
os.environ["BASE_HOME"] = os.path.dirname(__file__)
sys.path.append(os.path.dirname(__file__))
sys.path.append(os.path.join(os.environ['BASE_HOME'], 'modules/GLIGEN/demo'))
import modules.GLIGEN.demo.app as GLIGEN
import modules.GLIGEN.demo.gligen.task_grounded_generation as GLIGEN_generation
sys.path.append(os.path.join(os.environ['BASE_HOME'], 'modules/SEEM/demo_code'))
sys.path.append(os.path.join(os.environ['BASE_HOME'], 'modules/SEEM/demo_code/tasks'))
import modules.SEEM.demo_code.app as SEEM
import modules.SEEM.demo_code.utils.visualizer as visual
sys.path.append(os.path.join(os.environ['BASE_HOME'], 'modules/StableVideo'))
import modules.StableVideo.app as stablevideo
from vitron.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, OBJS_TOKEN_INDEX, DEFAULT_VIDEO_TOKEN, DEFAULT_OBJS_TOKEN
from vitron.conversation import conv_templates, SeparatorStyle
from vitron.model.builder import load_pretrained_model
from vitron.utils import disable_torch_init
from vitron.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria, tokenizer_image_region_token, preprocess_region, show_image_with_bboxes
def load_model(model_base, model_path, model_name):
disable_torch_init()
cache_dir = 'cache_dir'
device = 'cuda'
load_4bit, load_8bit = False, False
tokenizer, model, processor, _ = load_pretrained_model(model_path, model_base, model_name, load_8bit, load_4bit, device=device, cache_dir=cache_dir)
image_processor = processor['image']
video_processor = processor['video']
conv_mode = "llava_v1"
conv = conv_templates[conv_mode].copy()
return tokenizer, model, image_processor, video_processor, conv
model_path = 'checkpoints/Vitron-lora'
model_base = 'checkpoints/Vitron-base'
model_name = 'vitron-llava-7b-lora'
tokenizer, model, image_processor, video_processor, conv = load_model(model_path=model_path, model_base=model_base, model_name=model_name)
print('load model successfully')
def save_image_to_local(image: Image.Image):
# TODO: Update so the url path is used, to prevent repeat saving.
if not os.path.exists('temp'):
os.mkdir('temp')
filename = os.path.join('temp', next(tempfile._get_candidate_names()) + '.jpg')
image.save(filename)
return filename
def save_video_to_local(video):
if not os.path.exists('temp'):
os.mkdir('temp')
filename = os.path.join('temp', next(tempfile._get_candidate_names()) + '.mp4')
# export_to_video(video, filename)
writer = imageio.get_writer(filename, format='FFMPEG', fps=8)
for frame in video:
if isinstance(frame, Image.Image):
frame = np.array(frame)
writer.append_data(frame)
writer.close()
return filename
def image_generation(prompt="a black swan swimming in a pond surrounded by green plants"):
"""
:param prompt: text
:return:
"""
cache_file = os.path.join('checkpoints/gligen/gligen-generation-text-box', 'diffusion_pytorch_model.bin')
pretrained_ckpt_gligen = torch.load(cache_file, map_location='cpu')
cache_config = os.path.join('checkpoints/gligen/demo_config_legacy', 'gligen-generation-text-box.pth')
config = torch.load(cache_config, map_location='cpu')
config = OmegaConf.create(config["_content"])
config.update(
{'folder': 'create_samples', 'official_ckpt': 'ckpts/sd-v1-4.ckpt', 'guidance_scale': 5, 'alpha_scale': 1})
config.model['params']['is_inpaint'] = False
config.model['params']['is_style'] = True
loaded_model_list = GLIGEN_generation.load_ckpt(config, pretrained_ckpt_gligen)
instruction = dict(prompt=prompt, save_folder_name='gen_res', batch_size=1,
phrases=['placeholder'], has_text_mask=1, has_image_mask=0,
images=[], alpha_type=[0.3, 0, 0.7], guidance_scale=7.5, fix_seed=True,
rand_seed=0, actual_mask=None, inpainting_boxes_nodrop=None, locations=[])
sample_list, over_list = GLIGEN_generation.grounded_generation_box(loaded_model_list, instruction)
image_path = save_image_to_local(sample_list[0])
print(f'Generated image save into {image_path}')
return image_path
def image_segmentation(image_path, track_text, sketch_pad=None):
"""
Based on the input image, we segment the image and return the segmented image.
Args:
image (Image): The input image.
track_text (str): The reference text.
sketch_pad (Dict):
['image']: array
['mask']: array
Returns:
Image: The segmented image.
"""
print('Calling SEEM_app.inference')
if image_path is None:
return None, None
img = open_image(image_path)
width, height = img.size
if len(track_text) == 0 and sketch_pad is None:
# segment all
compose_img = {'image': img, 'mask': img}
task = []
image, _, labels = SEEM.inference(image=compose_img, task=task, reftxt=track_text)
return image[0], _, labels
if sketch_pad is not None:
compose_img = {'image': open_image(sketch_pad['image']), 'mask': sketch_pad['image']}
# print('mask path: ', save_image_to_local(open_image(sketch_pad['image'])))
# print('image segmentation / sketch_pad', sketch_pad) # sketch_pad['image']: array, sketch_pad['mask']: array
width, height = compose_img['image'].width, compose_img['image'].height
task = ['Stroke']
else:
compose_img = {'image': img, 'mask': img}
task = ['Text']
image, masks, labels = SEEM.inference(image=compose_img, task=task, reftxt=track_text)
mask_pred = masks[0].astype("uint8")
mask_pred = cv2.resize(mask_pred, (width, height), interpolation=cv2.INTER_LANCZOS4)
mask_pred = mask_pred.astype("uint8")
print('mask_pred: ', mask_pred)
mask_demo = visual.GenericMask(mask_pred, height, width)
bbox = mask_demo.bbox()
mask = {'mask': mask_pred, 'boxes': bbox}
return image[0], mask, labels
def image_editing(image_path=None, sketch_pad=None,
prompt="Turn the swan's neck into a wooden sail; Turn the swan into a wood boat"):
"""
:param image_path: the path of the image that wu want to edit
:param sketch_pad: the sketchpad input, formact: {"image": Image, "mask": Image}.
:param prompt: text prompt
:return: generate image with 512X512 resolution
"""
if image_path is None:
return None, None
image = open_image(image_path)
width, height = image.size
text = prompt
texts = [x.strip() for x in text.split(';')]
boxes = []
masks = []
if np.where(sketch_pad['mask'].max(0) != 0)[0].shape[0] == 0 or np.where(sketch_pad['mask'].max(1) != 0)[0].shape[0] == 0:
print('sketch pad is none')
# if there is no sketch_pad, i.e., no specification for image editing. Thus, first segmenting the image based on text， then inpainting the image
for t in texts:
_, t_mask, _ = image_segmentation(image_path=image_path, track_text=t)
boxes.append(t_mask['boxes'])
masks.append(t_mask['mask'])
state = {'boxes': boxes}
merged_mask = np.zeros((height, width))
print(merged_mask.shape)
for mask in masks:
merged_mask = np.logical_or(merged_mask, mask)
gen_images, state_list = GLIGEN.generate(task='Grounded Inpainting', language_instruction=prompt,
sketch_pad=None,
grounding_texts=prompt, alpha_sample=1.0, guidance_scale=30,
batch_size=1, fix_seed=False, rand_seed=0, use_actual_mask=False,
append_grounding=False,
style_cond_image=None, state=state, inpainting_mask=merged_mask,
inpainting_image=image
)
else:
boxes = mask_to_bbox(sketch_pad['mask'])
state = {'boxes': [boxes]}
print('sketch pad is not none')
print('state: ', state)
gen_images, state_list = GLIGEN.generate(task='Grounded Inpainting', language_instruction=prompt,
sketch_pad=sketch_pad,
grounding_texts=prompt, alpha_sample=1.0, guidance_scale=30,
batch_size=1, fix_seed=False, rand_seed=0, use_actual_mask=False,
append_grounding=False,
style_cond_image=None, state=state, inpainting_mask=None,
inpainting_image=image
)
# gen_images[0].save(save_path)
image_path = save_image_to_local(gen_images[0])
print(f'Generated image save into {image_path}')
return image_path, state_list
def video_generation(prompt,
num_inference_steps=50,
num_frames=24,
guidance_scale=7.5,
):
"""
Based on the input text prompt, we generate the corresponding video.
Args:
prompt (str): The input text prompt.
num_inference_steps (int): The number of inference steps.
num_frames (int): The number of frames.
guidance_scale (float): The guidance scale.
"""
pipe = DiffusionPipeline.from_pretrained("checkpoints/zeroscope", torch_dtype=torch.float16)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
# prompt = prompt
video_frames = pipe(prompt, num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
height=320, width=576,
num_frames=num_frames).frames[0]
video_frames = [np.array(frame) for frame in video_frames]
video_path = save_video_to_local(video_frames)
print(f'Generated video save into {video_path}')
return video_path
def video_tracking(video_path=None, sketch_pad=None, track_prompt="", text_prompt=""):
"""
Based on the input video, we track the video and return the tracked video.
Args:
video_path (str): The input video path.
track_text (str): The track text.
sketch_pad (dict): The sketchpad input with format {'image': Image, 'mask': Image}.
track_prompt (str): The track prompt.
text_prompt (str): The text prompt.
if no sketchpad, the text prompt is used to segment the image, obtaining the foreground images.
Returns:
str: The tracked video path.
"""
if video_path is None:
return None
if sketch_pad is None:
i_video_path = video_path.split('/')[-2]
img, o = video_editing(video_path=i_video_path, fore_prompt=text_prompt, back_prompt="")
image_path = save_image_to_local(img)
img = Image.open(image_path)
compose_img = {'image': img, 'mask': img}
else:
# compose_img = sketch_pad
compose_img = {'image': open_image(sketch_pad['image']), 'mask': sketch_pad['mask']}
# compose_img = {'image': sketch_pad['ibs'].image, 'mask': sketch_pad['ibs'].masks[-1]}
# print(save_image_to_local(open_image(sketch_pad['image']))) # an image with bbox
# print(save_image_to_local(open_image(sketch_pad['mask']))) # a binary mask with strech
_, output_video_name = SEEM.inference("examples/placeholder.png", task=['Video'],
video_pth=video_path, refimg=compose_img, reftxt=track_prompt)
return output_video_name
def video_editing(video_path=None, fore_prompt=None, back_prompt=None):
"""
:param video_path (str): directory of the video to be modified. The file structure shoud be like this
./data/xxx
-- checkpoint
-- config.json
-- xxx.mp4
-- texture.orig1.png
-- texture.orig2.png
-- xxx
-- 00000.jpg
-- 00001.jpg
:param fore_prompt (str): prompt for modifying foreground, such as "turn the orange into bread"
:param back_prompt (str): prompt for modifying background, such as "change the background into the blue"
"""
if video_path is None:
return None, None
st = stablevideo.StableVideo(base_cfg="checkpoints/stablevideo/cldm_v15.yaml",
canny_model_cfg="checkpoints/stablevideo/control_sd15_canny.pth",
depth_model_cfg="checkpoints/stablevideo/control_sd15_depth.pth",
save_memory=False)
st.load_canny_model(base_cfg='checkpoints/stablevideo/cldm_v15.yaml',
canny_model_cfg='checkpoints/stablevideo/control_sd15_canny.pth')
st.load_depth_model(base_cfg='checkpoints/stablevideo/cldm_v15.yaml',
depth_model_cfg='checkpoints/stablevideo/control_sd15_depth.pth', )
video_save_name, f_atlas_origin, b_altas_origin = st.load_video(video_path=video_path, video_name=os.path.basename(video_path))
print(video_save_name)
f_atlas = st.advanced_edit_foreground(prompt=fore_prompt)
# print(type(f_atlas))
b_altas = st.edit_background(back_prompt)
output_video_name = st.render_without_mask(f_atlas, b_altas)
print(output_video_name)
return f_atlas, output_video_name
def image_to_video(image_path=None,
text_prompt=None,
):
"""
Based on the input image and text prompt, we generate the corresponding video.
"""
if image_path is None or text_prompt is None:
return None, None
pipe = I2VGenXLPipeline.from_pretrained("checkpoints/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
pipe.enable_model_cpu_offload()
image = Image.open(image_path).convert("RGB")
negative_prompt = "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms"
generator = torch.manual_seed(8800)
frames = pipe(
prompt=text_prompt,
image=image,
num_inference_steps=50,
negative_prompt=negative_prompt,
guidance_scale=9.0,
generator=generator
).frames[0]
video_path = save_video_to_local(frames)
print(f'Generated video save into {video_path}')
return video_path
def find_module_content(data):
pattern = r'<module>(.*?)</module>'
match = re.search(pattern, data)
if match:
return match.group(1)
else:
return None
def find_instruction_content(data):
pattern = r'<instruction>(.*?)</instruction>'
match = re.findall(pattern, data)
if match:
res = []
for _res in match:
res.append(_res.split(':')[-1].strip())
return res
else:
return None
def find_region_instrction_content(data):
pattern = r'<region>(.*?)</region>'
match = re.search(pattern, data)
if match:
return match.group(1)
else:
return None
def remove_special_tags(text):
"""
remove the content between the tags and also the tags: <module></module> <instruction></instruction> <region></region> <SP></SP>
"""
pattern = r'<[^>]+>(.*?)<[^>]+>' # match all the tags
return re.sub(pattern, '', text)
def parse_model_output(model_output):
"""
Based on the model output, we parse the model output and return the parsed instructions.
Args:
model_output (str): The model output.
"""
# Parse the model output
module = find_module_content(model_output)
instruction = find_instruction_content(model_output)
region = find_region_instrction_content(model_output)
output = remove_special_tags(model_output)
return output, module, instruction, region
# projection
tasks = {
'A': image_generation,
'B': image_segmentation,
'C': image_editing,
'D': video_generation,
'E': video_tracking,
'F': video_editing,
'G': image_to_video,
}
def get_utterence(query, video_processor, image_processor):
"""
Based on the query, we compose the corresponding utterence.
Args:
query (list): The input query. query[0]-> text, query[1]-> image, query[2]-> video
video_processor (VideoProcessor): The video processor.
image_processor (ImageProcessor): The image processor.
"""
res_utterance = ''
video_tensor = None
image_tensor = None
region = None
if query[1] is not None and query[2] is not None:
# input includes video and image
res_utterance = ' '.join([DEFAULT_IMAGE_TOKEN] * model.get_video_tower().config.num_frames) + ' ' + DEFAULT_IMAGE_TOKEN + '\n' + query[0]
image_tensor = image_processor.preprocess(query[1], return_tensors='pt')['pixel_values'][0]
video_tensor = video_processor(query[2], return_tensors='pt')['pixel_values']
region = query[3]
elif query[1] is not None and query[2] is None:
# input includes image but no video
res_utterance = DEFAULT_IMAGE_TOKEN + '\n' + query[0]
image_tensor = image_processor.preprocess(query[1], return_tensors='pt')['pixel_values'][0]
region = query[3]
elif query[1] is None and query[2] is not None:
# input includes no image but video
res_utterance = ' '.join([DEFAULT_IMAGE_TOKEN] * model.get_video_tower().config.num_frames) + '\n' + query[0]
video_tensor = video_processor(query[2], return_tensors='pt')['pixel_values'][0]
region = query[3]
else:
# input includes no video but image
res_utterance = query[0]
return res_utterance, video_tensor, image_tensor, region
def re_predict(user_input, input_image_state, input_image, out_imagebox,
input_video_state, input_video, video_sketch_pad, history, chatbox, *args):
q, a = history.pop()
chatbox.pop()
# input_image_state = new_state()['ibs'].update_image(open_image(q[1]))
q_utterance, q_video_tensor, q_image_tensor, q_region = get_utterence(q, video_processor, image_processor)
return predict(q_utterance, input_image_state, q[1], out_imagebox,
input_video_state, q[2], video_sketch_pad, history, chatbox,
*args)
def predict(user_input, input_image_state, input_image, out_imagebox,
input_video_state, input_video, video_sketch_pad, history, chatbox,
*args):
"""
Based on the user input and history, we generate the response and update the history.
Args:
tokenizer (Tokenizer):
The tokenizer for process user input instructions.
model (Model): The model.
image_processor (ImageProcessor):
The image processor for process image.
video_processor (VideoProcessor):
The video processor for process video.
conv(Conversation): Conversation class.
history (list):
The history. [[(q1, v1, i1, r1), (a1, v1, i1, r1)], [(q2, v2, i2, r2), (a2, v2, i2, r2)]
chatbox (list): The chatbox.
input_image_state (dict): {'ibs': ImageBoxState}.
Saving the image state including the image, box and mask.
input_image (Numpy.ndarray):
The input image.
out_imagebox (dict):
The output image box. {'image': Numpy.ndarray, 'mask': Numpy.ndarray}.
input_video_state (dict): {'ibs': ImageBoxState}.
Saving the video state including the video frame list, current frame, box and mask.
input_video (str):
The file path of input video.
video_sketch_pad (dict):
The video sketch pad of each frame. {'image': Numpy.ndarray, 'mask': Numpy.ndarray}.
configs (dict): The configurations.
"""
video_tensors = []
image_tensors = []
input_region = []
config = create_cfg(*args)
default_input_region = [0, 0, 224, 224]
if history is not None:
print('history: ', history)
# default_input_region = [0, 0, 224, 224]
for idx, _his in enumerate(history):
print(f'idx: {idx}, history[idx]: {_his}')
q, a = _his
q_utterance, q_video_tensor, q_image_tensor, q_region = get_utterence(q, video_processor, image_processor)
conv.append_message(conv.roles[0], q_utterance)
a_utterance, a_video_tensor, a_image_tensor, a_region = get_utterence(a, video_processor, image_processor)
conv.append_message(conv.roles[1], a_utterance)
if q_video_tensor is not None:
video_tensors.append(q_video_tensor)
input_region.append(q_region)
if q_image_tensor is not None:
image_tensors.append(q_image_tensor)
input_region.append(q_region)
if a_video_tensor is not None:
video_tensors.append(a_video_tensor)
input_region.append(q_region)
if a_image_tensor is not None:
image_tensors.append(a_image_tensor)
input_region.append(a_region)
inp = ''
_user_input = user_input
query_img_path = ''
if input_video is not None:
inp = inp + ' '.join([DEFAULT_IMAGE_TOKEN] * model.get_video_tower().config.num_frames)
video_tensors.append(video_processor(input_video, return_tensors='pt')['pixel_values'][0]) # 'input_video' should be a file_path
_user_input += f'<br><video controls playsinline width="500" style="display: inline-block;" src="./file={input_video}"></video>'
input_region.append(default_input_region)
if input_image is not None:
inp = inp + ' ' + DEFAULT_IMAGE_TOKEN
_image = open_image(input_image['image'])
image_tensors.append(image_processor.preprocess(_image, return_tensors='pt')['pixel_values'][0])
ori_im_size = [_image.width, _image.height]
if input_image_state is not None and len(input_image_state['ibs'].boxes) > 0:
bbox = input_image_state['ibs'].boxes[-1]
input_region.append(bbox)
ori_im_size = [input_image_state['ibs'].width, input_image_state['ibs'].height]
input_region = [preprocess_region(_bbox, ori_im_size, [224, 224]) for _bbox in input_region]
inp = inp + '\n' + DEFAULT_OBJS_TOKEN + ' '
else:
input_region.append(default_input_region)
query_img_path = save_image_to_local(_image)
_user_input += f'<br><img src="./file={query_img_path}" style="display: inline-block;width: 250px;max-height: 400px;">'
inp = inp + '\n' + user_input if inp.endswith('>') else inp + user_input
print('inp: ', inp)
conv.append_message(conv.roles[0], inp)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
conv.clear_message()
print('prompt: ', prompt)
# input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).cuda()
input_ids = tokenizer_image_region_token(prompt, tokenizer, OBJS_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
keywords = [stop_str]
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
if len(image_tensors) == 0 and len(video_tensors) == 0:
# no image or video input
tensor = [torch.zeros(3, image_processor.crop_size['height'], image_processor.crop_size['width']).to(model.device, dtype=torch.float16)]
input_region = [default_input_region]
else:
tensor = [_tensor.to(model.device, dtype=torch.float16) for _tensor in video_tensors+image_tensors]
assert len(input_region) == len(tensor)
with torch.inference_mode():
output_ids = model.generate(
input_ids,
images=tensor,
regions = input_region,
do_sample=True,
temperature=config['temperature'],
top_p = config['top_p'],
max_new_tokens=1024,
use_cache=True,
stopping_criteria=[stopping_criteria])
outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
print('model outputs: ', outputs)
output, module, instruction, region = parse_model_output(outputs)
print('parsed output: ', output)
print('module: ', module)
print('instruction: ', instruction)
print('region: ', region)
# _user_input = None
if input_image_state is not None and input_image_state['ibs'].raw_image is not None:
query_img_path = save_image_to_local(input_image_state['ibs'].raw_image)
print(query_img_path)
else:
query_img_path = None
if module and module in tasks:
if module == 'A':
if instruction is not None and len(instruction) > 0:
ans_image_path = image_generation(prompt=instruction[0])
_response = output + f'<br><img src="./file={ans_image_path}" style="display: inline-block;width: 250px;max-height: 400px;">'
else:
_response = output
ans_image_path = None
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, ans_image_path, None, default_input_region)))
elif module == 'B':
if instruction is not None and len(instruction) > 0:
image_seg, pad, label = image_segmentation(image_path=query_img_path, track_text=instruction[0],
sketch_pad=input_image)
ans_image_path = save_image_to_local(image_seg)
_response = output + f'<br><img src="./file={ans_image_path}" style="display: inline-block;width: 250px;max-height: 400px;">'
else:
_response = output
ans_image_path = None
print(f'image file save into {ans_image_path}')
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, ans_image_path, None, default_input_region)))
elif module == 'C':
if instruction is not None and len(instruction) > 0:
ans_image_path, _ = image_editing(image_path=query_img_path, sketch_pad=input_image, prompt=instruction[0])
_response = output + f'<br><img src="./file={ans_image_path}" style="display: inline-block;width: 250px;max-height: 400px;">'
else:
_response = output
ans_image_path = None
print(f'image file save into {ans_image_path}')
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, ans_image_path, None, default_input_region)))
elif module == 'D':
if instruction is not None and len(instruction) > 0:
ans_video_path = video_generation(prompt=instruction[0], num_inference_steps=config['num_inference_steps_for_vid'], num_frames=config['num_frames'], guidance_scale=config['guidance_scale_for_vid'])
_response = output + f'<br><video controls playsinline width="500" style="display: inline-block;" src="./file={ans_video_path}"></video>'
else:
_response = output
ans_image_path = None
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, None, ans_video_path, default_input_region)))
elif module == 'E':
if input_video_state['ibs'].raw_image is not None:
query_img_path = save_image_to_local(input_video_state['ibs'].raw_image)
else:
query_img_path = None
if instruction is not None and len(instruction) > 0:
ans_video_path = video_tracking(video_path=input_video, sketch_pad=video_sketch_pad, track_prompt=instruction[0])
_response = output + f'<br><video controls playsinline width="500" style="display: inline-block;" src="./file={ans_video_path}"></video>'
else:
_response = output
ans_video_path = None
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, None, ans_video_path, default_input_region)))
elif module == 'F':
if instruction is not None and len(instruction) >= 2:
_, ans_video_path = video_editing(video_path=input_video, fore_prompt=instruction[0], back_prompt=instruction[1])
_response = output + f'<br><video controls playsinline width="500" style="display: inline-block;" src="./file={ans_video_path}"></video>'
else:
_response = output
ans_video_path = None
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, None, ans_video_path, default_input_region)))
elif module == 'G':
if instruction is not None and len(instruction) > 0:
ans_video_path = image_to_video(image_path=query_img_path, text_prompt=instruction[0])
_response = output + f'<br><video controls playsinline width="500" style="display: inline-block;" src="./file={ans_video_path}"></video>'
else:
_response = output
ans_video_path = None
chatbox.append((_user_input, _response))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, None, ans_video_path, default_input_region)))
else:
raise NotImplementedError(f'The module {module} is not implemented.')
else:
chatbox.append((_user_input, output))
history.append(((user_input, query_img_path, input_video, input_region[-1]), (output, None, None, default_input_region)))
return chatbox, history, None, None, None, None
def new_state():
return {"ibs": ImageBoxState()}
def upload_image(sketch_pad: dict, state: dict):
image = sketch_pad['image']
# print('sketch_pad', sketch_pad) # {'image': array unit8, 'mask': array unit8}
image = open_image(image)
ibs = state["ibs"]
ibs.update_image(image)
return image, state
def reset_state(input_image_state, input_video_state):
ibs = input_image_state["ibs"]
ibs.reset_state()
ibs = input_video_state["ibs"]
ibs.reset_state()
return input_image_state, input_video_state, None, None, None, None, None, [], []
def create_cfg(seed, top_p, temperature,
guidance_scale_for_img_edit, num_inference_steps_for_img_edit,
guidance_scale_for_vid, num_inference_steps_for_vid, num_frames,
num_inference_steps_for_vid_edit, guidance_scale_for_vide_edit):
cfg_dict = {
"seed": seed,
"top_p": top_p,
"temperature": temperature,
"guidance_scale_for_img_edit":guidance_scale_for_img_edit,
"num_inference_steps_for_img_edit": num_inference_steps_for_img_edit,
"guidance_scale_for_vid": guidance_scale_for_vid,
"num_inference_steps_for_vid": num_inference_steps_for_vid,
"num_frames": num_frames,
"num_inference_steps_for_vid_edit": num_inference_steps_for_vid_edit,
"guidance_scale_for_vide_edit": guidance_scale_for_vide_edit
}
return cfg_dict
def extract_frames(value, state: dict):
"""
Based on the input video, we extract the video frames, and return the frames as a list of images.
modified the code, as we do not need to extract all frames in the video, we only need to extract some frames.
"""
# Get the video file path
video_path = value
# Open the video file
vidcap = cv2.VideoCapture(video_path)
# Get the frame count
frame_count = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
# Get the frame rate
frame_rate = int(vidcap.get(cv2.CAP_PROP_FPS))
# Get the total duration of the video
duration = frame_count / frame_rate
# Get the number of frames to be extracted
num_frames = 8
# Get the frame interval
frame_interval = int(frame_count / num_frames)
# Get the frames
frames = []
# Loop through the frames
for i in range(0, frame_count, frame_interval):
# Set the frame position
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
# Read the frame
success, frame = vidcap.read()
# Convert the frame to RGB
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
# Convert the frame to PIL image
frame = Image.fromarray(frame)
# Append the frame to the list of frames
frames.append(frame)
ibs = state["ibs"]
ibs.update_image_list(frames)
return frames[0], state
def edit_video_frame(sketch_pad, state):
"""
Based on the input video frame, we extract the video frame, and return the frame as an image.
"""
def binarize(x):
return (x != 0).astype('uint8') * 255
image = sketch_pad['image']
image = open_image(image)
mask = sketch_pad['mask'].sum(-1) if sketch_pad['mask'].ndim == 3 else sketch_pad['mask']
mask = binarize(mask)
ibs = state["ibs"]
ibs.update_image(image)
ibs.update_mask(mask)
return state
def select_next_frame(state):
ibs = state["ibs"]
ibs.cnt += 1
if ibs.cnt >= len(ibs.image_list):
ibs.cnt = 0
ibs.update_image(ibs.image_list[ibs.cnt])
return ibs.image_list[ibs.cnt], state
def clear_video_and_frame(state):
ibs = state["ibs"]
# print(ibs.image_list) # List[Image]
# print(ibs.image) # Image
ibs.reset_state()
return state, None, None
def clear_image_and_sketch_pad(state):
ibs = state["ibs"]
ibs.reset_state()
return state, None, None
def clear_image_and_video(image_state, video_state):
image_state["ibs"].reset_state()
video_state["ibs"].reset_state()
return image_state, video_state, None, None, None, None, [], []
def clear_input(image_state, video_state):
image_state["ibs"].reset_state()
video_state["ibs"].reset_state()
return image_state, video_state, None, None, None, None, None, [], []
# return None
class ImageMask(gr.components.Image):
"""
Sets: source="canvas", tool="sketch"
"""
is_template = True
def __init__(self, **kwargs):
super().__init__(source="upload", tool="sketch", interactive=True, **kwargs)
def preprocess(self, x):
if isinstance(x, str):
x = {'image': x, 'mask': x}
elif isinstance(x, dict):
if (x['mask'] is None and x['image'] is None):
x
elif (x['image'] is None):
x['image'] = str(x['mask'])
elif (x['mask'] is None):
x['mask'] = str(x['image']) #not sure why mask/mask is None sometimes, this prevents preprocess crashing
elif x is not None:
assert False, 'Unexpected type {0} in ImageMask preprocess()'.format(type(x))
return super().preprocess(x)
TITLE = """
<h1 align="center" style="display: flex;flex-direction: row;justify-content: center;font-size: 40pt;align-content: center;align-items: center;"> <img src="./file=vitron.png" width="80" height="80" style="margin-right: 10px;">VITRON</h1>
<div align="center" style="display: flex;"><a href='https://vitron-llm.github.io/'><img src='https://img.shields.io/badge/Project-Page-Green'></a>       <a href='https://github.com/SkyworkAI/Vitron'><img src='https://img.shields.io/badge/Github-Code-blue'></a>       <a href='https://is.gd/aGu0VV'><img src='https://img.shields.io/badge/Paper-PDF-red'></a>       <a href='https://youtu.be/wiGMJzoQVu4'><img src='https://img.shields.io/badge/video-YouTube-FF0000'></a></div>
"""
INTRODUCTION = """
<h2>Introduction</h2>
<p>This is the demo page of Vitron, a universal pixel-level vision LLM, designed for comprehensive understanding (perceiving and reasoning), generating, segmenting (grounding and tracking), editing (inpainting) of both static image and dynamic video content.</p>
<h2>Term of Use</h2>
<p>The service is a research preview intended for non-commercial use only. The current initial version of Vitron, limited by the quantity of fine-tuning data and the quality of the base models, may generate some low-quality or hallucinated content. Please interpret the results with caution. We will continue to update the model to enhance its performance. Thank you for trying the demo! If you have any questions or feedback, feel free to contact us.</p>
"""
def build_demo():
demo = gr.Blocks(title="Vitron", css='style.css')
with demo:
gr.HTML(TITLE)
gr.HTML(INTRODUCTION)
with gr.Row():
with gr.Column(scale=7, min_width=500):
with gr.Row():
chatbot = gr.Chatbot(label='Vitron Chatbot', height=500, elem_id='chatbox', avatar_images=((os.path.join(os.path.dirname(__file__), 'user.png')), (os.path.join(os.path.dirname(__file__), "vitron.png"))))
with gr.Row():
user_input = gr.Textbox(label='User Input', placeholder='Enter your text here', elem_id='user_input', lines=3)
with gr.Row():
with gr.Column(scale=3):
with gr.Tab('🌁 Image'):
with gr.Row():
# input_image = gr.Image(label='Input Image', type='numpy',
# shape=(512, 512),
# # height=200, width=200,
# elem_id='img2img_image',
# interactive=True, tool='sketch',
# brush_radius=20.0, visible=True)
input_image = ImageMask(label="Input Image", type="numpy",
shape=(512, 512),
# height=200, width=200,
elem_id='img2img_image',
# tool='sketch',
brush_radius=20.0, visible=True)
with gr.Row():
clearImageBtn = gr.Button("Clear Image and Sketch Pad", elem_id='clear_image')
with gr.Row():
out_imagebox = gr.Image(label='Parsed Sketch Pad', type='numpy',
shape=(512, 512),
# height=200, width=200,
elem_id='out_imagebox')
input_image_state = gr.State(new_state())
with gr.Column(scale=3):
with gr.Tab('🎬 Video'):
with gr.Row():
input_video = gr.Video(label='Input Video', format='mp4', visible=True) #.style(height=200) # , value=None, interactive=True
with gr.Row():
with gr.Column(scale=0.3):
nextFrameBtn = gr.Button("Next Frame", elem_id='next_frmame', variant="primary")
clearFrameBtn = gr.Button("Clear Video & Frame", elem_id='clear_frmame')
# with gr.Column(scale=0.3):
with gr.Row():
# video_sketch_pad = gr.Image(label='Video Frame', type='numpy',
# shape=(512, 512),
# # height=200, width=200,
# elem_id='video_sketch_pad',
# interactive=True, tool='sketch',
# brush_radius=20.0, visible=True)
video_sketch_pad = ImageMask(label='Video Frame', type='numpy',
shape=(512, 512),
# height=200, width=200,
elem_id='video_sketch_pad',
# interactive=True, tool='sketch',
brush_radius=20.0, visible=True)
input_video_state = gr.State(new_state())
with gr.Column(scale=3, min_width=300):
with gr.Group():
seed = gr.Slider(0, 9999, value=1234, label="SEED", interactive=True)
with gr.Accordion('Text Advanced Options', open=True):
top_p = gr.Slider(0, 1, value=0.01, step=0.01, label="Top P", interactive=True)
temperature = gr.Slider(0, 1, value=1.0, step=0.01, label="Temperature", interactive=True)
with gr.Accordion('Image Editing Advanced Options', open=True):
guidance_scale_for_img_edit = gr.Slider(1, 10, value=7.5, step=0.5, label="Guidance scale",
interactive=True)
num_inference_steps_for_img_edit = gr.Slider(10, 50, value=50, step=1, label="Number of inference steps",
interactive=True)
with gr.Accordion('Video Generation Options', open=False):
guidance_scale_for_vid = gr.Slider(1, 10, value=7.5, step=0.5, label="Guidance scale",
interactive=True)
num_inference_steps_for_vid = gr.Slider(10, 50, value=50, step=1, label="Number of inference steps",
interactive=True)
num_frames = gr.Slider(label='Number of frames', minimum=16, maximum=32, step=8, value=24,
interactive=True,
info='Note that the content of the video also changes when you change the number of frames.')
with gr.Accordion('Video Editing Advanced Options', open=False):
num_inference_steps_for_vid_edit = gr.Slider(10, 50, value=50, step=1, label="Number of inference steps",
interactive=True)
guidance_scale_for_vide_edit = gr.Slider(1, 100, value=50, step=10, label="The audio length in seconds",
interactive=True)
configs = [
seed, top_p, temperature,
guidance_scale_for_img_edit, num_inference_steps_for_img_edit,
guidance_scale_for_vid, num_inference_steps_for_vid, num_frames,
num_inference_steps_for_vid_edit, guidance_scale_for_vide_edit
]
with gr.Tab("🎯 Operation"):
with gr.Row(scale=1):
submitBtn = gr.Button(value="Submit & Run", variant="primary")
with gr.Row(scale=1):
resubmitBtn = gr.Button("Rerun")
with gr.Row(scale=1):
emptyBtn = gr.Button("Clear History")
with gr.Tab("📔 Guidebook"):
with gr.Accordion('📖 Open it for detailed instruction', open=False,):
gr.Markdown("#### 🍀Input Image")
gr.Markdown('You can upload an image in here')
gr.Markdown('#### 🎋Parsed Sketch Pad')
gr.Markdown('When you scribble on the input image, we will parse the sketch and draw the corresponding bounding box on the image')
gr.Markdown('#### 🌵Input Video')
gr.Markdown('You can upload a video in here')
gr.Markdown('#### 🌳Video Frame')
gr.Markdown('The video frames are extracted automatically with the first frame displayed for you. Mark any object you wish to track directly on the frame and use the `Next Frame` button to proceed to subsequent frames.')
gr.Markdown('#### 🎄Examples')
gr.Markdown('We provide examples, you can choose one example and then click `Submit&Run` to obtain the results.')
gr.Markdown('The results may be delayed due to network speeds or model processing efficiency. If you have any questions, feel free to reach out to the authors. Thank you for your patience.🫰🫰')
# input_image.upload(fn=clear_fn2, inputs=emptyBtn, outputs=[output_text, out_imagebox, input_image_state])
# input_image.clear(fn=clear_fn2, inputs=emptyBtn, outputs=[output_text, out_imagebox, input_image_state])
history = gr.State([])
with gr.Tab('Image Understanding'):
img_und = gr.Examples(
[
['What is the role of the net in a tennis game, and what strategies might a player adopt when positioned close to the net?', 'examples/000000015269.jpg'],
['How large is the pizza in relation to the table', 'examples/000000346930.jpg'],
['Can this food item be considered lasagna?', 'examples/000000036260.jpg'],
['CWhat activity might the woman be engaging in, and what type of entertainment could she be enjoying?', 'examples/000000036904.jpg'],
['Write a detailed description of the given image.', 'examples/000000455523.jpg']
],
[user_input, input_image],
examples_per_page=5,
label=''
)
img_und.dataset.click(clear_video_and_frame, inputs=[input_video_state], outputs=[input_video_state, video_sketch_pad, input_video])
with gr.Tab('Image Generation'):
img_gen = gr.Examples(
[
['I\'ve always been fascinated by wolves and their social dynamics. Can you imagine an image where a gray wolf is seen interacting with its pack, displaying their strong social bonds?'],
['I\'m looking for inspiration for a trendy dining room design. Could you show me an image that exemplifies modern style and elegance?'],
['Hey, I\'m looking for an image of a wild mountain goat standing alone against a plain white background. Can you help me with that?'],
['I\'d love to see an image of a tranquil river flowing by a campground.'],
['Could you please provide me with a video that showcases a scenic left-to-right panoramic view of a city as the camera navigates through its winding streets?']
],
[user_input],
examples_per_page=5,
label=''
)
img_gen.dataset.click(clear_image_and_video, inputs=[input_image_state, input_video_state], outputs=[input_image_state, input_video_state, input_image, out_imagebox, input_video, video_sketch_pad, history, chatbot])
with gr.Tab('Image Segmentation'):
img_seg = gr.Examples(
[
['In this photo, can you show me the man who is eating a hotdog?', 'examples/000000326300.jpg'],
['Hey, I need to highlight the laptop in this picture, can you handle it?', 'examples/000000363079.jpg'],
['Could you please segment out the person on the right side in the image?', 'examples/000000418723.jpg'],
['Could you assist me in locating the woman in this picture', 'examples/000000448663.jpg']
],
[user_input, input_image],
examples_per_page=5,