File size: 24,973 Bytes
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
44f4dde
12996e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44f4dde
 
 
 
 
12996e6
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
44f4dde
 
 
12996e6
44f4dde
 
12996e6
 
44f4dde
12996e6
 
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
44f4dde
 
12996e6
 
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
 
 
 
 
44f4dde
 
12996e6
 
 
 
 
 
44f4dde
 
12996e6
 
 
 
 
 
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
44f4dde
 
12996e6
 
 
 
 
44f4dde
 
12996e6
 
44f4dde
 
 
12996e6
44f4dde
 
 
 
12996e6
44f4dde
 
 
 
12996e6
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
44f4dde
 
 
12996e6
44f4dde
 
 
 
 
 
 
 
12996e6
44f4dde
 
12996e6
 
 
 
 
44f4dde
 
12996e6
 
44f4dde
 
 
12996e6
44f4dde
 
 
 
12996e6
44f4dde
 
 
 
12996e6
 
 
44f4dde
 
 
 
 
 
 
12996e6
 
 
 
 
 
 
 
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
 
 
 
44f4dde
 
 
 
 
 
 
 
12996e6
 
 
 
 
 
 
 
44f4dde
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12996e6
 
 
 
44f4dde
 
 
 
 
 
 
 
 
 
 
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
import gradio as gr
from gradio_toggle import Toggle
import torch
from huggingface_hub import snapshot_download
from transformers import pipeline

from xora.models.autoencoders.causal_video_autoencoder import CausalVideoAutoencoder
from xora.models.transformers.transformer3d import Transformer3DModel
from xora.models.transformers.symmetric_patchifier import SymmetricPatchifier
from xora.schedulers.rf import RectifiedFlowScheduler
from xora.pipelines.pipeline_xora_video import XoraVideoPipeline
from transformers import T5EncoderModel, T5Tokenizer
from xora.utils.conditioning_method import ConditioningMethod
from pathlib import Path
import safetensors.torch
import json
import numpy as np
import cv2
from PIL import Image
import tempfile
import os
import gc
from openai import OpenAI
import re

# Load system prompts
system_prompt_t2v = """당신은 λΉ„λ””μ˜€ 생성을 μœ„ν•œ ν”„λ‘¬ν”„νŠΈ μ „λ¬Έκ°€μž…λ‹ˆλ‹€. 
주어진 ν”„λ‘¬ν”„νŠΈλ₯Ό λ‹€μŒ ꡬ쑰에 맞게 κ°œμ„ ν•΄μ£Όμ„Έμš”:
1. μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
2. ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
3. 캐릭터/객체의 μ™Έλͺ¨λ₯Ό μƒμ„Ένžˆ λ¬˜μ‚¬
4. λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
5. 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
6. μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
7. λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함
λͺ¨λ“  μ„€λͺ…은 ν•˜λ‚˜μ˜ μžμ—°μŠ€λŸ¬μš΄ λ¬Έλ‹¨μœΌλ‘œ μž‘μ„±ν•˜κ³ , 
촬영 감독이 촬영 λͺ©λ‘μ„ μ„€λͺ…ν•˜λŠ” κ²ƒμ²˜λŸΌ ꡬ체적이고 μ‹œκ°μ μœΌλ‘œ μž‘μ„±ν•˜μ„Έμš”.
200단어λ₯Ό λ„˜μ§€ μ•Šλ„λ‘ ν•˜λ˜, μ΅œλŒ€ν•œ μƒμ„Έν•˜κ²Œ μž‘μ„±ν•˜μ„Έμš”."""

system_prompt_i2v = """당신은 이미지 기반 λΉ„λ””μ˜€ 생성을 μœ„ν•œ ν”„λ‘¬ν”„νŠΈ μ „λ¬Έκ°€μž…λ‹ˆλ‹€. 
주어진 ν”„λ‘¬ν”„νŠΈλ₯Ό λ‹€μŒ ꡬ쑰에 맞게 κ°œμ„ ν•΄μ£Όμ„Έμš”:
1. μ£Όμš” λ™μž‘μ„ λͺ…ν™•ν•œ ν•œ λ¬Έμž₯으둜 μ‹œμž‘
2. ꡬ체적인 λ™μž‘κ³Ό 제슀처λ₯Ό μ‹œκ°„ μˆœμ„œλŒ€λ‘œ μ„€λͺ…
3. 캐릭터/객체의 μ™Έλͺ¨λ₯Ό μƒμ„Ένžˆ λ¬˜μ‚¬
4. λ°°κ²½κ³Ό ν™˜κ²½ μ„ΈλΆ€ 사항을 ꡬ체적으둜 포함
5. 카메라 각도와 μ›€μ§μž„μ„ λͺ…μ‹œ
6. μ‘°λͺ…κ³Ό 색상을 μžμ„Ένžˆ μ„€λͺ…
7. λ³€ν™”λ‚˜ κ°‘μž‘μŠ€λŸ¬μš΄ 사건을 μžμ—°μŠ€λŸ½κ²Œ 포함
λͺ¨λ“  μ„€λͺ…은 ν•˜λ‚˜μ˜ μžμ—°μŠ€λŸ¬μš΄ λ¬Έλ‹¨μœΌλ‘œ μž‘μ„±ν•˜κ³ , 
촬영 감독이 촬영 λͺ©λ‘μ„ μ„€λͺ…ν•˜λŠ” κ²ƒμ²˜λŸΌ ꡬ체적이고 μ‹œκ°μ μœΌλ‘œ μž‘μ„±ν•˜μ„Έμš”.
200단어λ₯Ό λ„˜μ§€ μ•Šλ„λ‘ ν•˜λ˜, μ΅œλŒ€ν•œ μƒμ„Έν•˜κ²Œ μž‘μ„±ν•˜μ„Έμš”."""

# Load Hugging Face token if needed
hf_token = os.getenv("HF_TOKEN")
openai_api_key = os.getenv("OPENAI_API_KEY")
client = OpenAI(api_key=openai_api_key)

# Initialize translation pipeline
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")

# Korean text detection function
def contains_korean(text):
    korean_pattern = re.compile('[γ„±-γ…Žγ…-γ…£κ°€-힣]')
    return bool(korean_pattern.search(text))

def translate_korean_prompt(prompt):
    """
    Translate Korean prompt to English if Korean text is detected
    """
    if contains_korean(prompt):
        translated = translator(prompt)[0]['translation_text']
        print(f"Original Korean prompt: {prompt}")
        print(f"Translated English prompt: {translated}")
        return translated
    return prompt

def enhance_prompt(prompt, type="t2v"):
    system_prompt = system_prompt_t2v if type == "t2v" else system_prompt_i2v
    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": prompt},
    ]

    try:
        response = client.chat.completions.create(
            model="gpt-4-1106-preview",
            messages=messages,
            max_tokens=2000,
        )
        enhanced_prompt = response.choices[0].message.content.strip()
        
        print("\n=== ν”„λ‘¬ν”„νŠΈ 증강 κ²°κ³Ό ===")
        print("Original Prompt:")
        print(prompt)
        print("\nEnhanced Prompt:")
        print(enhanced_prompt)
        print("========================\n")
        
        return enhanced_prompt
    except Exception as e:
        print(f"Error during prompt enhancement: {e}")
        return prompt

def update_prompt_t2v(prompt, enhance_toggle):
    return update_prompt(prompt, enhance_toggle, "t2v")

def update_prompt_i2v(prompt, enhance_toggle):
    return update_prompt(prompt, enhance_toggle, "i2v")
    
def update_prompt(prompt, enhance_toggle, type="t2v"):
    if enhance_toggle:
        return enhance_prompt(prompt, type)
    return prompt

# Set model download directory within Hugging Face Spaces
model_path = "asset"
if not os.path.exists(model_path):
    snapshot_download(
        "Lightricks/LTX-Video", local_dir=model_path, repo_type="model", token=hf_token
    )

# Global variables to load components
vae_dir = Path(model_path) / "vae"
unet_dir = Path(model_path) / "unet"
scheduler_dir = Path(model_path) / "scheduler"

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def load_vae(vae_dir):
    vae_ckpt_path = vae_dir / "vae_diffusion_pytorch_model.safetensors"
    vae_config_path = vae_dir / "config.json"
    with open(vae_config_path, "r") as f:
        vae_config = json.load(f)
    vae = CausalVideoAutoencoder.from_config(vae_config)
    vae_state_dict = safetensors.torch.load_file(vae_ckpt_path)
    vae.load_state_dict(vae_state_dict)
    return vae.to(device=device, dtype=torch.bfloat16)

def load_unet(unet_dir):
    unet_ckpt_path = unet_dir / "unet_diffusion_pytorch_model.safetensors"
    unet_config_path = unet_dir / "config.json"
    transformer_config = Transformer3DModel.load_config(unet_config_path)
    transformer = Transformer3DModel.from_config(transformer_config)
    unet_state_dict = safetensors.torch.load_file(unet_ckpt_path)
    transformer.load_state_dict(unet_state_dict, strict=True)
    return transformer.to(device=device, dtype=torch.bfloat16)

def load_scheduler(scheduler_dir):
    scheduler_config_path = scheduler_dir / "scheduler_config.json"
    scheduler_config = RectifiedFlowScheduler.load_config(scheduler_config_path)
    return RectifiedFlowScheduler.from_config(scheduler_config)

# Helper function for image processing
def center_crop_and_resize(frame, target_height, target_width):
    h, w, _ = frame.shape
    aspect_ratio_target = target_width / target_height
    aspect_ratio_frame = w / h
    if aspect_ratio_frame > aspect_ratio_target:
        new_width = int(h * aspect_ratio_target)
        x_start = (w - new_width) // 2
        frame_cropped = frame[:, x_start : x_start + new_width]
    else:
        new_height = int(w / aspect_ratio_target)
        y_start = (h - new_height) // 2
        frame_cropped = frame[y_start : y_start + new_height, :]
    frame_resized = cv2.resize(frame_cropped, (target_width, target_height))
    return frame_resized

def load_image_to_tensor_with_resize(image_path, target_height=512, target_width=768):
    image = Image.open(image_path).convert("RGB")
    image_np = np.array(image)
    frame_resized = center_crop_and_resize(image_np, target_height, target_width)
    frame_tensor = torch.tensor(frame_resized).permute(2, 0, 1).float()
    frame_tensor = (frame_tensor / 127.5) - 1.0
    return frame_tensor.unsqueeze(0).unsqueeze(2)

# Load models
vae = load_vae(vae_dir)
unet = load_unet(unet_dir)
scheduler = load_scheduler(scheduler_dir)
patchifier = SymmetricPatchifier(patch_size=1)
text_encoder = T5EncoderModel.from_pretrained(
    "PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="text_encoder"
).to(device)
tokenizer = T5Tokenizer.from_pretrained(
    "PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="tokenizer"
)

pipeline = XoraVideoPipeline(
    transformer=unet,
    patchifier=patchifier,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    scheduler=scheduler,
    vae=vae,
).to(device)

# Preset options for resolution and frame configuration
# Convert frames to seconds assuming 25 FPS
preset_options = [
    {"label": "1216x704, 1.6초", "width": 1216, "height": 704, "num_frames": 41},
    {"label": "1088x704, 2.0초", "width": 1088, "height": 704, "num_frames": 49},
    {"label": "1056x640, 2.3초", "width": 1056, "height": 640, "num_frames": 57},
    {"label": "992x608, 2.6초", "width": 992, "height": 608, "num_frames": 65},
    {"label": "896x608, 2.9초", "width": 896, "height": 608, "num_frames": 73},
    {"label": "896x544, 3.2초", "width": 896, "height": 544, "num_frames": 81},
    {"label": "832x544, 3.6초", "width": 832, "height": 544, "num_frames": 89},
    {"label": "800x512, 3.9초", "width": 800, "height": 512, "num_frames": 97},
    {"label": "768x512, 3.9초", "width": 768, "height": 512, "num_frames": 97},
    {"label": "800x480, 4.2초", "width": 800, "height": 480, "num_frames": 105},
    {"label": "736x480, 4.5초", "width": 736, "height": 480, "num_frames": 113},
    {"label": "704x480, 4.8초", "width": 704, "height": 480, "num_frames": 121},
    {"label": "704x448, 5.2초", "width": 704, "height": 448, "num_frames": 129},
    {"label": "672x448, 5.5초", "width": 672, "height": 448, "num_frames": 137},
    {"label": "640x416, 6.1초", "width": 640, "height": 416, "num_frames": 153},
    {"label": "672x384, 6.4초", "width": 672, "height": 384, "num_frames": 161},
    {"label": "640x384, 6.8초", "width": 640, "height": 384, "num_frames": 169},
    {"label": "608x384, 7.1초", "width": 608, "height": 384, "num_frames": 177},
    {"label": "576x384, 7.4초", "width": 576, "height": 384, "num_frames": 185},
    {"label": "608x352, 7.7초", "width": 608, "height": 352, "num_frames": 193},
    {"label": "576x352, 8.0초", "width": 576, "height": 352, "num_frames": 201},
    {"label": "544x352, 8.4초", "width": 544, "height": 352, "num_frames": 209},
    {"label": "512x352, 9.3초", "width": 512, "height": 352, "num_frames": 233},
    {"label": "544x320, 9.6초", "width": 544, "height": 320, "num_frames": 241},
    {"label": "512x320, 10.3초", "width": 512, "height": 320, "num_frames": 257},
]

def preset_changed(preset):
    if preset != "Custom":
        selected = next(item for item in preset_options if item["label"] == preset)
        # height, width, num_frames 값을 global λ³€μˆ˜λ‘œ μ—…λ°μ΄νŠΈ
        return (
            selected["height"],
            selected["width"],
            selected["num_frames"],
            gr.update(visible=False),
            gr.update(visible=False),
            gr.update(visible=False),
        )
    else:
        return (
            None,
            None,
            None,
            gr.update(visible=True),
            gr.update(visible=True),
            gr.update(visible=True),
        )


def generate_video_from_text(
    prompt="",
    enhance_prompt_toggle=False,
    negative_prompt="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
    frame_rate=25,
    seed=171198,
    num_inference_steps=41,
    guidance_scale=4,
    height=512,
    width=320,
    num_frames=257,
    progress=gr.Progress(),
):
    if len(prompt.strip()) < 50:
        raise gr.Error(
            "ν”„λ‘¬ν”„νŠΈλŠ” μ΅œμ†Œ 50자 이상이어야 ν•©λ‹ˆλ‹€. 더 μžμ„Έν•œ μ„€λͺ…을 μ œκ³΅ν•΄μ£Όμ„Έμš”.",
            duration=5,
        )

    # Translate Korean prompts to English
    prompt = translate_korean_prompt(prompt)
    negative_prompt = translate_korean_prompt(negative_prompt)

    sample = {
        "prompt": prompt,
        "prompt_attention_mask": None,
        "negative_prompt": negative_prompt,
        "negative_prompt_attention_mask": None,
        "media_items": None,
    }

    generator = torch.Generator(device="cpu").manual_seed(seed)

    def gradio_progress_callback(self, step, timestep, kwargs):
        progress((step + 1) / num_inference_steps)

    try:
        with torch.no_grad():
            images = pipeline(
                num_inference_steps=num_inference_steps,
                num_images_per_prompt=1,
                guidance_scale=guidance_scale,
                generator=generator,
                output_type="pt",
                height=height,
                width=width,
                num_frames=num_frames,
                frame_rate=frame_rate,
                **sample,
                is_video=True,
                vae_per_channel_normalize=True,
                conditioning_method=ConditioningMethod.UNCONDITIONAL,
                mixed_precision=True,
                callback_on_step_end=gradio_progress_callback,
            ).images
    except Exception as e:
        raise gr.Error(
            f"λΉ„λ””μ˜€ 생성 쀑 였λ₯˜κ°€ λ°œμƒν–ˆμŠ΅λ‹ˆλ‹€. λ‹€μ‹œ μ‹œλ„ν•΄μ£Όμ„Έμš”. 였λ₯˜: {e}",
            duration=5,
        )
    finally:
        torch.cuda.empty_cache()
        gc.collect()

    output_path = tempfile.mktemp(suffix=".mp4")
    print(images.shape)
    video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
    video_np = (video_np * 255).astype(np.uint8)
    height, width = video_np.shape[1:3]
    out = cv2.VideoWriter(
        output_path, cv2.VideoWriter_fourcc(*"mp4v"), frame_rate, (width, height)
    )
    for frame in video_np[..., ::-1]:
        out.write(frame)
    out.release()
    del images
    del video_np
    torch.cuda.empty_cache()
    return output_path

def generate_video_from_image(
    image_path,
    prompt="",
    enhance_prompt_toggle=False,
    negative_prompt="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
    frame_rate=25,
    seed=171198,
    num_inference_steps=50,
    guidance_scale=4,
    height=512,
    width=768,
    num_frames=121,
    progress=gr.Progress(),
):
    print("Height: ", height)
    print("Width: ", width)
    print("Num Frames: ", num_frames)

    if len(prompt.strip()) < 50:
        raise gr.Error(
            "ν”„λ‘¬ν”„νŠΈλŠ” μ΅œμ†Œ 50자 이상이어야 ν•©λ‹ˆλ‹€. 더 μžμ„Έν•œ μ„€λͺ…을 μ œκ³΅ν•΄μ£Όμ„Έμš”.",
            duration=5,
        )

    if not image_path:
        raise gr.Error("μž…λ ₯ 이미지λ₯Ό μ œκ³΅ν•΄μ£Όμ„Έμš”.", duration=5)

    # Translate Korean prompts to English
    prompt = translate_korean_prompt(prompt)
    negative_prompt = translate_korean_prompt(negative_prompt)

    media_items = (
        load_image_to_tensor_with_resize(image_path, height, width).to(device).detach()
    )

    sample = {
        "prompt": prompt,
        "prompt_attention_mask": None,
        "negative_prompt": negative_prompt,
        "negative_prompt_attention_mask": None,
        "media_items": media_items,
    }

    generator = torch.Generator(device="cpu").manual_seed(seed)

    def gradio_progress_callback(self, step, timestep, kwargs):
        progress((step + 1) / num_inference_steps)

    try:
        with torch.no_grad():
            images = pipeline(
                num_inference_steps=num_inference_steps,
                num_images_per_prompt=1,
                guidance_scale=guidance_scale,
                generator=generator,
                output_type="pt",
                height=height,
                width=width,
                num_frames=num_frames,
                frame_rate=frame_rate,
                **sample,
                is_video=True,
                vae_per_channel_normalize=True,
                conditioning_method=ConditioningMethod.FIRST_FRAME,
                mixed_precision=True,
                callback_on_step_end=gradio_progress_callback,
            ).images

        output_path = tempfile.mktemp(suffix=".mp4")
        video_np = images.squeeze(0).permute(1, 2, 3, 0).cpu().float().numpy()
        video_np = (video_np * 255).astype(np.uint8)
        height, width = video_np.shape[1:3]
        out = cv2.VideoWriter(
            output_path, cv2.VideoWriter_fourcc(*"mp4v"), frame_rate, (width, height)
        )
        for frame in video_np[..., ::-1]:
            out.write(frame)
        out.release()
    except Exception as e:
        raise gr.Error(
            f"λΉ„λ””μ˜€ 생성 쀑 였λ₯˜κ°€ λ°œμƒν–ˆμŠ΅λ‹ˆλ‹€. λ‹€μ‹œ μ‹œλ„ν•΄μ£Όμ„Έμš”. 였λ₯˜: {e}",
            duration=5,
        )

    finally:
        torch.cuda.empty_cache()
        gc.collect()

    return output_path

def create_advanced_options():
    with gr.Accordion("Step 4: Advanced Options (Optional)", open=False):
        seed = gr.Slider(
            label="Seed",
            minimum=0,
            maximum=1000000,
            step=1,
            value=171198
        )
        inference_steps = gr.Slider(
            label="4.2 Inference Steps",
            minimum=1,
            maximum=50,
            step=1,
            value=50,
            visible=False
        )
        guidance_scale = gr.Slider(
            label="4.3 Guidance Scale",
            minimum=1.0,
            maximum=5.0,
            step=0.1,
            value=4.0,
            visible=False
        )
        height_slider = gr.Slider(
            label="4.4 Height",
            minimum=256,
            maximum=1024,
            step=64,
            value=512,
            visible=False,
        )
        width_slider = gr.Slider(
            label="4.5 Width",
            minimum=256,
            maximum=1024,
            step=64,
            value=768,
            visible=False,
        )
        num_frames_slider = gr.Slider(
            label="4.5 Number of Frames",
            minimum=1,
            maximum=200,
            step=1,
            value=121,
            visible=False,
        )

        return [
            seed,
            inference_steps,
            guidance_scale,
            height_slider,
            width_slider,
            num_frames_slider,
        ]

# Gradio Interface Definition
with gr.Blocks(theme=gr.themes.Soft()) as iface:
    with gr.Tabs():
        # Text to Video Tab
        with gr.TabItem("ν…μŠ€νŠΈλ‘œ λΉ„λ””μ˜€ λ§Œλ“€κΈ°"):
            with gr.Row():
                with gr.Column():
                    txt2vid_prompt = gr.Textbox(
                        label="Step 1: ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
                        placeholder="μƒμ„±ν•˜κ³  싢은 λΉ„λ””μ˜€λ₯Ό μ„€λͺ…ν•˜μ„Έμš” (μ΅œμ†Œ 50자)...",
                        value="κ·€μ—¬μš΄ 고양이",
                        lines=5,
                    )
                    txt2vid_enhance_toggle = Toggle(
                        label="ν”„λ‘¬ν”„νŠΈ κ°œμ„ ",
                        value=False,
                        interactive=True,
                    )

                    txt2vid_negative_prompt = gr.Textbox(
                        label="Step 2: λ„€κ±°ν‹°λΈŒ ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
                        placeholder="λΉ„λ””μ˜€μ—μ„œ μ›ν•˜μ§€ μ•ŠλŠ” μš”μ†Œλ₯Ό μ„€λͺ…ν•˜μ„Έμš”...",
                        value="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
                        lines=2,
                        visible=False
                    )

                    # ν˜„μž¬ μ„ νƒλœ 값듀을 μ €μž₯ν•  μƒνƒœ λ³€μˆ˜λ“€
                    txt2vid_current_height = gr.State(value=512)
                    txt2vid_current_width = gr.State(value=320)
                    txt2vid_current_num_frames = gr.State(value=257)

                    txt2vid_preset = gr.Dropdown(
                        choices=[p["label"] for p in preset_options],
                        value="512x320, 10.3초",
                        label="Step 2: 해상도 프리셋 선택",
                    )

                    txt2vid_frame_rate = gr.Slider(
                        label="Step 3: ν”„λ ˆμž„ 레이트",
                        minimum=21,
                        maximum=30,
                        step=1,
                        value=25,
                        visible=False
                    )

                    txt2vid_advanced = create_advanced_options()
                    txt2vid_generate = gr.Button(
                        "Step 3: λΉ„λ””μ˜€ 생성",
                        variant="primary",
                        size="lg",
                    )

                with gr.Column():
                    txt2vid_output = gr.Video(label="μƒμ„±λœ λΉ„λ””μ˜€")

        # Image to Video Tab
        with gr.TabItem("μ΄λ―Έμ§€λ‘œ λΉ„λ””μ˜€ λ§Œλ“€κΈ°"):
            with gr.Row():
                with gr.Column():
                    img2vid_image = gr.Image(
                        type="filepath",
                        label="Step 1: μž…λ ₯ 이미지 μ—…λ‘œλ“œ",
                        elem_id="image_upload",
                    )
                    img2vid_prompt = gr.Textbox(
                        label="Step 2: ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
                        placeholder="이미지λ₯Ό μ–΄λ–»κ²Œ μ• λ‹ˆλ©”μ΄μ…˜ν™”ν• μ§€ μ„€λͺ…ν•˜μ„Έμš” (μ΅œμ†Œ 50자)...",
                        value="κ·€μ—¬μš΄ 고양이",
                        lines=5,
                    )
                    img2vid_enhance_toggle = Toggle(
                        label="ν”„λ‘¬ν”„νŠΈ 증강",
                        value=False,
                        interactive=True,
                    )
                    img2vid_negative_prompt = gr.Textbox(
                        label="Step 3: λ„€κ±°ν‹°λΈŒ ν”„λ‘¬ν”„νŠΈ μž…λ ₯",
                        placeholder="λΉ„λ””μ˜€μ—μ„œ μ›ν•˜μ§€ μ•ŠλŠ” μš”μ†Œλ₯Ό μ„€λͺ…ν•˜μ„Έμš”...",
                        value="low quality, worst quality, deformed, distorted, warped, motion smear, motion artifacts, fused fingers, incorrect anatomy, strange hands, unattractive",
                        lines=2,
                        visible=False
                    )

                    # ν˜„μž¬ μ„ νƒλœ 값듀을 μ €μž₯ν•  μƒνƒœ λ³€μˆ˜λ“€
                    img2vid_current_height = gr.State(value=512)
                    img2vid_current_width = gr.State(value=768)
                    img2vid_current_num_frames = gr.State(value=97)

                    img2vid_preset = gr.Dropdown(
                        choices=[p["label"] for p in preset_options],
                        value="512x320, 10.3초",
                        label="Step 3: 해상도 프리셋 선택",
                    )

                    img2vid_frame_rate = gr.Slider(
                        label="Step 4: ν”„λ ˆμž„ 레이트",
                        minimum=21,
                        maximum=30,
                        step=1,
                        value=25,
                        visible=False
                    )

                    img2vid_advanced = create_advanced_options()
                    img2vid_generate = gr.Button(
                        "Step 4: λΉ„λ””μ˜€ 생성",
                        variant="primary",
                        size="lg",
                    )

                with gr.Column():
                    img2vid_output = gr.Video(label="μƒμ„±λœ λΉ„λ””μ˜€")

    # Event handlers
    txt2vid_preset.change(
        fn=preset_changed,
        inputs=[txt2vid_preset],
        outputs=[
            txt2vid_current_height,
            txt2vid_current_width,
            txt2vid_current_num_frames,
            *txt2vid_advanced[3:]
        ]
    )

    txt2vid_enhance_toggle.change(
        fn=update_prompt_t2v,
        inputs=[txt2vid_prompt, txt2vid_enhance_toggle],
        outputs=txt2vid_prompt
    )

    txt2vid_generate.click(
        fn=generate_video_from_text,
        inputs=[
            txt2vid_prompt,
            txt2vid_enhance_toggle,
            txt2vid_negative_prompt,
            txt2vid_frame_rate,
            *txt2vid_advanced[:3],  # seed, inference_steps, guidance_scale
            txt2vid_current_height,
            txt2vid_current_width,
            txt2vid_current_num_frames,
        ],
        outputs=txt2vid_output,
        concurrency_limit=1,
        concurrency_id="generate_video",
        queue=True,
    )

    img2vid_preset.change(
        fn=preset_changed,
        inputs=[img2vid_preset],
        outputs=[
            img2vid_current_height,
            img2vid_current_width,
            img2vid_current_num_frames,
            *img2vid_advanced[3:]
        ]
    )

    img2vid_enhance_toggle.change(
        fn=update_prompt_i2v,
        inputs=[img2vid_prompt, img2vid_enhance_toggle],
        outputs=img2vid_prompt
    )

    img2vid_generate.click(
        fn=generate_video_from_image,
        inputs=[
            img2vid_image,
            img2vid_prompt,
            img2vid_enhance_toggle,
            img2vid_negative_prompt,
            img2vid_frame_rate,
            *img2vid_advanced[:3],  # seed, inference_steps, guidance_scale
            img2vid_current_height,
            img2vid_current_width,
            img2vid_current_num_frames,
        ],
        outputs=img2vid_output,
        concurrency_limit=1,
        concurrency_id="generate_video",
        queue=True,
    )

if __name__ == "__main__":
    iface.queue(max_size=64, default_concurrency_limit=1, api_open=False).launch(
        share=True, show_api=False
    )