DawnC commited on
Commit
7070961
·
verified ·
1 Parent(s): 17ce59d

Upload 2 files

Browse files
Files changed (2) hide show
  1. app.py +543 -0
  2. style_transfer.py +752 -0
app.py ADDED
@@ -0,0 +1,543 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import numpy as np
3
+ import torch
4
+ import torch.nn as nn
5
+ import gradio as gr
6
+ import time
7
+ import spaces
8
+ import timm
9
+ from torchvision.ops import nms, box_iou
10
+ import torch.nn.functional as F
11
+ from torchvision import transforms
12
+ from PIL import Image, ImageDraw, ImageFont, ImageFilter
13
+ from breed_health_info import breed_health_info
14
+ from breed_noise_info import breed_noise_info
15
+ from dog_database import get_dog_description
16
+ from scoring_calculation_system import UserPreferences, calculate_compatibility_score
17
+ from recommendation_html_format import format_recommendation_html, get_breed_recommendations
18
+ from history_manager import UserHistoryManager
19
+ from search_history import create_history_tab, create_history_component
20
+ from styles import get_css_styles
21
+ from breed_detection import create_detection_tab
22
+ from breed_comparison import create_comparison_tab
23
+ from breed_recommendation import create_recommendation_tab
24
+ from breed_visualization import create_visualization_tab
25
+ from style_transfer import DogStyleTransfer, create_style_transfer_tab
26
+ from html_templates import (
27
+ format_description_html,
28
+ format_single_dog_result,
29
+ format_multiple_breeds_result,
30
+ format_unknown_breed_message,
31
+ format_not_dog_message,
32
+ format_hint_html,
33
+ format_multi_dog_container,
34
+ format_breed_details_html,
35
+ get_color_scheme,
36
+ get_akc_breeds_link
37
+ )
38
+ from model_architecture import BaseModel, dog_breeds
39
+ from urllib.parse import quote
40
+ from ultralytics import YOLO
41
+ import asyncio
42
+ import traceback
43
+
44
+ history_manager = UserHistoryManager()
45
+
46
+ class ModelManager:
47
+ """
48
+ Singleton class for managing model instances and device allocation
49
+ specifically designed for Hugging Face Spaces deployment.
50
+ """
51
+ _instance = None
52
+ _initialized = False
53
+ _yolo_model = None
54
+ _breed_model = None
55
+ _device = None
56
+
57
+ def __new__(cls):
58
+ if cls._instance is None:
59
+ cls._instance = super().__new__(cls)
60
+ return cls._instance
61
+
62
+ def __init__(self):
63
+ if not ModelManager._initialized:
64
+ self._device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
65
+ ModelManager._initialized = True
66
+
67
+ @property
68
+ def device(self):
69
+ if self._device is None:
70
+ self._device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
71
+ return self._device
72
+
73
+ @property
74
+ def yolo_model(self):
75
+ if self._yolo_model is None:
76
+ self._yolo_model = YOLO('yolov8x.pt')
77
+ return self._yolo_model
78
+
79
+ @property
80
+ def breed_model(self):
81
+ if self._breed_model is None:
82
+ self._breed_model = BaseModel(
83
+ num_classes=len(dog_breeds),
84
+ device=self.device
85
+ ).to(self.device)
86
+
87
+ checkpoint = torch.load(
88
+ 'ConvNextV2Base_best_model.pth',
89
+ map_location=self.device
90
+ )
91
+ self._breed_model.load_state_dict(checkpoint['base_model'], strict=False)
92
+ self._breed_model.eval()
93
+ return self._breed_model
94
+
95
+ # Initialize model manager
96
+ model_manager = ModelManager()
97
+
98
+ def preprocess_image(image):
99
+ """Preprocesses images for model input"""
100
+ if isinstance(image, np.ndarray):
101
+ image = Image.fromarray(image)
102
+
103
+ transform = transforms.Compose([
104
+ transforms.Resize((224, 224)),
105
+ transforms.ToTensor(),
106
+ transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
107
+ ])
108
+
109
+ return transform(image).unsqueeze(0)
110
+
111
+ @spaces.GPU
112
+ def predict_single_dog(image):
113
+ """Predicts dog breed for a single image"""
114
+ image_tensor = preprocess_image(image).to(model_manager.device)
115
+
116
+ with torch.no_grad():
117
+ logits = model_manager.breed_model(image_tensor)[0]
118
+ probs = F.softmax(logits, dim=1)
119
+
120
+ top5_prob, top5_idx = torch.topk(probs, k=5)
121
+ breeds = [dog_breeds[idx.item()] for idx in top5_idx[0]]
122
+ probabilities = [prob.item() for prob in top5_prob[0]]
123
+
124
+ sum_probs = sum(probabilities[:3])
125
+ relative_probs = [f"{(prob/sum_probs * 100):.2f}%" for prob in probabilities[:3]]
126
+
127
+ return probabilities[0], breeds[:3], relative_probs
128
+
129
+ def enhanced_preprocess(image, is_standing=False, has_overlap=False):
130
+ """
131
+ Enhanced image preprocessing function with special handling for different poses
132
+ and overlapping cases.
133
+ """
134
+ target_size = 224
135
+ w, h = image.size
136
+
137
+ if is_standing:
138
+ if h > w * 1.5:
139
+ new_h = target_size
140
+ new_w = min(target_size, int(w * (target_size / h)))
141
+ new_w = max(new_w, int(target_size * 0.6))
142
+ elif has_overlap:
143
+ scale = min(target_size/w, target_size/h) * 0.95
144
+ new_w = int(w * scale)
145
+ new_h = int(h * scale)
146
+ else:
147
+ scale = min(target_size/w, target_size/h)
148
+ new_w = int(w * scale)
149
+ new_h = int(h * scale)
150
+
151
+ resized = image.resize((new_w, new_h), Image.Resampling.LANCZOS)
152
+ final_image = Image.new('RGB', (target_size, target_size), (240, 240, 240))
153
+ paste_x = (target_size - new_w) // 2
154
+ paste_y = (target_size - new_h) // 2
155
+ final_image.paste(resized, (paste_x, paste_y))
156
+
157
+ return final_image
158
+
159
+ @spaces.GPU
160
+ def detect_multiple_dogs(image, conf_threshold=0.3, iou_threshold=0.3):
161
+ """
162
+ Enhanced multiple dog detection with improved bounding box handling and
163
+ intelligent boundary adjustments.
164
+ """
165
+ results = model_manager.yolo_model(image, conf=conf_threshold, iou=iou_threshold)[0]
166
+ img_width, img_height = image.size
167
+ detected_boxes = []
168
+
169
+ # Phase 1: Initial detection and processing
170
+ for box in results.boxes:
171
+ if box.cls.item() == 16: # Dog class
172
+ xyxy = box.xyxy[0].tolist()
173
+ confidence = box.conf.item()
174
+ x1, y1, x2, y2 = map(int, xyxy)
175
+ w = x2 - x1
176
+ h = y2 - y1
177
+
178
+ detected_boxes.append({
179
+ 'coords': [x1, y1, x2, y2],
180
+ 'width': w,
181
+ 'height': h,
182
+ 'center_x': (x1 + x2) / 2,
183
+ 'center_y': (y1 + y2) / 2,
184
+ 'area': w * h,
185
+ 'confidence': confidence,
186
+ 'aspect_ratio': w / h if h != 0 else 1
187
+ })
188
+
189
+ if not detected_boxes:
190
+ return [(image, 1.0, [0, 0, img_width, img_height], False)]
191
+
192
+ # Phase 2: Analysis of detection relationships
193
+ avg_height = sum(box['height'] for box in detected_boxes) / len(detected_boxes)
194
+ avg_width = sum(box['width'] for box in detected_boxes) / len(detected_boxes)
195
+ avg_area = sum(box['area'] for box in detected_boxes) / len(detected_boxes)
196
+
197
+ def calculate_iou(box1, box2):
198
+ x1 = max(box1['coords'][0], box2['coords'][0])
199
+ y1 = max(box1['coords'][1], box2['coords'][1])
200
+ x2 = min(box1['coords'][2], box2['coords'][2])
201
+ y2 = min(box1['coords'][3], box2['coords'][3])
202
+
203
+ if x2 <= x1 or y2 <= y1:
204
+ return 0.0
205
+
206
+ intersection = (x2 - x1) * (y2 - y1)
207
+ area1 = box1['area']
208
+ area2 = box2['area']
209
+ return intersection / (area1 + area2 - intersection)
210
+
211
+ # Phase 3: Processing each detection
212
+ processed_boxes = []
213
+ overlap_threshold = 0.2
214
+
215
+ for i, box_info in enumerate(detected_boxes):
216
+ x1, y1, x2, y2 = box_info['coords']
217
+ w = box_info['width']
218
+ h = box_info['height']
219
+ center_x = box_info['center_x']
220
+ center_y = box_info['center_y']
221
+
222
+ # Check for overlaps
223
+ has_overlap = False
224
+ for j, other_box in enumerate(detected_boxes):
225
+ if i != j and calculate_iou(box_info, other_box) > overlap_threshold:
226
+ has_overlap = True
227
+ break
228
+
229
+ # Adjust expansion strategy
230
+ base_expansion = 0.03
231
+ max_expansion = 0.05
232
+
233
+ is_standing = h > 1.5 * w
234
+ is_sitting = 0.8 <= h/w <= 1.2
235
+ is_abnormal_size = (h * w) > (avg_area * 1.5) or (h * w) < (avg_area * 0.5)
236
+
237
+ if has_overlap:
238
+ h_expansion = w_expansion = base_expansion * 0.8
239
+ else:
240
+ if is_standing:
241
+ h_expansion = min(base_expansion * 1.2, max_expansion)
242
+ w_expansion = base_expansion
243
+ elif is_sitting:
244
+ h_expansion = w_expansion = base_expansion
245
+ else:
246
+ h_expansion = w_expansion = base_expansion * 0.9
247
+
248
+ # Position compensation
249
+ if center_x < img_width * 0.2 or center_x > img_width * 0.8:
250
+ w_expansion *= 0.9
251
+
252
+ if is_abnormal_size:
253
+ h_expansion *= 0.8
254
+ w_expansion *= 0.8
255
+
256
+ # Calculate final bounding box
257
+ expansion_w = w * w_expansion
258
+ expansion_h = h * h_expansion
259
+
260
+ new_x1 = max(0, center_x - (w + expansion_w)/2)
261
+ new_y1 = max(0, center_y - (h + expansion_h)/2)
262
+ new_x2 = min(img_width, center_x + (w + expansion_w)/2)
263
+ new_y2 = min(img_height, center_y + (h + expansion_h)/2)
264
+
265
+ # Crop and process image
266
+ cropped_image = image.crop((int(new_x1), int(new_y1),
267
+ int(new_x2), int(new_y2)))
268
+
269
+ processed_image = enhanced_preprocess(
270
+ cropped_image,
271
+ is_standing=is_standing,
272
+ has_overlap=has_overlap
273
+ )
274
+
275
+ processed_boxes.append((
276
+ processed_image,
277
+ box_info['confidence'],
278
+ [new_x1, new_y1, new_x2, new_y2],
279
+ True
280
+ ))
281
+
282
+ return processed_boxes
283
+
284
+ @spaces.GPU
285
+ def predict(image):
286
+ """
287
+ Main prediction function that handles both single and multiple dog detection.
288
+ Args:
289
+ image: PIL Image or numpy array
290
+ Returns:
291
+ tuple: (html_output, annotated_image, initial_state)
292
+ """
293
+ if image is None:
294
+ return format_hint_html("Please upload an image to start."), None, None
295
+
296
+ try:
297
+ if isinstance(image, np.ndarray):
298
+ image = Image.fromarray(image)
299
+
300
+ # 檢測圖片中的物體
301
+ dogs = detect_multiple_dogs(image)
302
+ color_scheme = get_color_scheme(len(dogs) == 1)
303
+
304
+ # 準備標註
305
+ annotated_image = image.copy()
306
+ draw = ImageDraw.Draw(annotated_image)
307
+
308
+ try:
309
+ font = ImageFont.truetype("arial.ttf", 24)
310
+ except:
311
+ font = ImageFont.load_default()
312
+
313
+ dogs_info = ""
314
+
315
+ # 處理每個檢測到的物體
316
+ for i, (cropped_image, detection_confidence, box, is_dog) in enumerate(dogs):
317
+ print(f"Predict processing - Object {i+1}:")
318
+ print(f" Is dog: {is_dog}")
319
+ print(f" Detection confidence: {detection_confidence:.4f}")
320
+
321
+ # 如果是狗且進行品種預測,在這裡也加入打印語句
322
+ if is_dog:
323
+ top1_prob, topk_breeds, relative_probs = predict_single_dog(cropped_image)
324
+ print(f" Breed prediction - Top probability: {top1_prob:.4f}")
325
+ print(f" Top breeds: {topk_breeds[:3]}")
326
+ color = color_scheme if len(dogs) == 1 else color_scheme[i % len(color_scheme)]
327
+
328
+ # 繪製框和標籤
329
+ draw.rectangle(box, outline=color, width=4)
330
+ label = f"Dog {i+1}" if is_dog else f"Object {i+1}"
331
+ label_bbox = draw.textbbox((0, 0), label, font=font)
332
+ label_width = label_bbox[2] - label_bbox[0]
333
+ label_height = label_bbox[3] - label_bbox[1]
334
+
335
+ # 繪製標籤背景和文字
336
+ label_x = box[0] + 5
337
+ label_y = box[1] + 5
338
+ draw.rectangle(
339
+ [label_x - 2, label_y - 2, label_x + label_width + 4, label_y + label_height + 4],
340
+ fill='white',
341
+ outline=color,
342
+ width=2
343
+ )
344
+ draw.text((label_x, label_y), label, fill=color, font=font)
345
+
346
+ try:
347
+ # 首先檢查是否為狗
348
+ if not is_dog:
349
+ dogs_info += format_not_dog_message(color, i+1)
350
+ continue
351
+
352
+ # 如果是狗,進行品種預測
353
+ top1_prob, topk_breeds, relative_probs = predict_single_dog(cropped_image)
354
+ combined_confidence = detection_confidence * top1_prob
355
+
356
+ # 根據信心度決定輸出格式
357
+ if combined_confidence < 0.15:
358
+ dogs_info += format_unknown_breed_message(color, i+1)
359
+ elif top1_prob >= 0.4:
360
+ breed = topk_breeds[0]
361
+ description = get_dog_description(breed)
362
+ if description is None:
363
+ description = {
364
+ "Name": breed,
365
+ "Size": "Unknown",
366
+ "Exercise Needs": "Unknown",
367
+ "Grooming Needs": "Unknown",
368
+ "Care Level": "Unknown",
369
+ "Good with Children": "Unknown",
370
+ "Description": f"Identified as {breed.replace('_', ' ')}"
371
+ }
372
+ dogs_info += format_single_dog_result(breed, description, color)
373
+ else:
374
+ dogs_info += format_multiple_breeds_result(
375
+ topk_breeds,
376
+ relative_probs,
377
+ color,
378
+ i+1,
379
+ lambda breed: get_dog_description(breed) or {
380
+ "Name": breed,
381
+ "Size": "Unknown",
382
+ "Exercise Needs": "Unknown",
383
+ "Grooming Needs": "Unknown",
384
+ "Care Level": "Unknown",
385
+ "Good with Children": "Unknown",
386
+ "Description": f"Identified as {breed.replace('_', ' ')}"
387
+ }
388
+ )
389
+ except Exception as e:
390
+ print(f"Error formatting results for dog {i+1}: {str(e)}")
391
+ dogs_info += format_unknown_breed_message(color, i+1)
392
+
393
+ # 包裝最終的HTML輸出
394
+ html_output = format_multi_dog_container(dogs_info)
395
+
396
+ # 準備初始狀態
397
+ initial_state = {
398
+ "dogs_info": dogs_info,
399
+ "image": annotated_image,
400
+ "is_multi_dog": len(dogs) > 1,
401
+ "html_output": html_output
402
+ }
403
+
404
+ return html_output, annotated_image, initial_state
405
+
406
+ except Exception as e:
407
+ error_msg = f"An error occurred: {str(e)}\n\nTraceback:\n{traceback.format_exc()}"
408
+ print(error_msg)
409
+ return format_hint_html(error_msg), None, None
410
+
411
+
412
+ def show_details_html(choice, previous_output, initial_state):
413
+ """
414
+ Generate detailed HTML view for a selected breed.
415
+
416
+ Args:
417
+ choice: str, Selected breed option
418
+ previous_output: str, Previous HTML output
419
+ initial_state: dict, Current state information
420
+
421
+ Returns:
422
+ tuple: (html_output, gradio_update, updated_state)
423
+ """
424
+ if not choice:
425
+ return previous_output, gr.update(visible=True), initial_state
426
+
427
+ try:
428
+ breed = choice.split("More about ")[-1]
429
+ description = get_dog_description(breed)
430
+ html_output = format_breed_details_html(description, breed)
431
+
432
+ # Update state
433
+ initial_state["current_description"] = html_output
434
+ initial_state["original_buttons"] = initial_state.get("buttons", [])
435
+
436
+ return html_output, gr.update(visible=True), initial_state
437
+
438
+ except Exception as e:
439
+ error_msg = f"An error occurred while showing details: {e}"
440
+ print(error_msg)
441
+ return format_hint_html(error_msg), gr.update(visible=True), initial_state
442
+
443
+ def main():
444
+ with gr.Blocks(css=get_css_styles()) as iface:
445
+ # Header HTML
446
+
447
+ gr.HTML("""
448
+ <header style='text-align: center; padding: 20px; margin-bottom: 20px;'>
449
+ <h1 style='font-size: 2.5em; margin-bottom: 10px; color: #2D3748;'>
450
+ 🐾 PawMatch AI
451
+ </h1>
452
+ <h2 style='font-size: 1.2em; font-weight: normal; color: #4A5568; margin-top: 5px;'>
453
+ Your Smart Dog Breed Guide
454
+ </h2>
455
+ <div style='width: 50px; height: 3px; background: linear-gradient(90deg, #4299e1, #48bb78); margin: 15px auto;'></div>
456
+ <p style='color: #718096; font-size: 0.9em;'>
457
+ Powered by AI • Breed Recognition • Smart Matching • Companion Guide
458
+ </p>
459
+ </header>
460
+ """)
461
+
462
+ # 先創建歷史組件實例(但不創建標籤頁)
463
+ history_component = create_history_component()
464
+
465
+ # Initialize style transfor
466
+ dog_style_transfer = DogStyleTransfer()
467
+
468
+ with gr.Tabs():
469
+ # 1. breed detection
470
+ example_images = [
471
+ 'Border_Collie.jpg',
472
+ 'Golden_Retriever.jpeg',
473
+ 'Saint_Bernard.jpeg',
474
+ 'Samoyed.jpeg',
475
+ 'French_Bulldog.jpeg'
476
+ ]
477
+ detection_components = create_detection_tab(predict, example_images)
478
+
479
+ # 2. breed comparison
480
+ comparison_components = create_comparison_tab(
481
+ dog_breeds=dog_breeds,
482
+ get_dog_description=get_dog_description,
483
+ breed_health_info=breed_health_info,
484
+ breed_noise_info=breed_noise_info
485
+ )
486
+
487
+ # 3. breed recommendation
488
+ recommendation_components = create_recommendation_tab(
489
+ UserPreferences=UserPreferences,
490
+ get_breed_recommendations=get_breed_recommendations,
491
+ format_recommendation_html=format_recommendation_html,
492
+ history_component=history_component
493
+ )
494
+
495
+ # 4. Visualization Analysis
496
+ with gr.Tab("Visualization Analysis"):
497
+ create_visualization_tab(
498
+ dog_breeds=dog_breeds,
499
+ get_dog_description=get_dog_description,
500
+ calculate_compatibility_score=calculate_compatibility_score,
501
+ UserPreferences=UserPreferences
502
+ )
503
+
504
+ # 5. Style Transfer tab
505
+ with gr.Tab("Style Transfer"):
506
+ style_transfer_components = create_style_transfer_tab(dog_style_transfer)
507
+
508
+
509
+ # 6. History Search
510
+ create_history_tab(history_component)
511
+
512
+ # Footer
513
+ gr.HTML('''
514
+ <div style="
515
+ display: flex;
516
+ align-items: center;
517
+ justify-content: center;
518
+ gap: 20px;
519
+ padding: 20px 0;
520
+ ">
521
+ <p style="
522
+ font-family: 'Arial', sans-serif;
523
+ font-size: 14px;
524
+ font-weight: 500;
525
+ letter-spacing: 2px;
526
+ background: linear-gradient(90deg, #555, #007ACC);
527
+ -webkit-background-clip: text;
528
+ -webkit-text-fill-color: transparent;
529
+ margin: 0;
530
+ text-transform: uppercase;
531
+ display: inline-block;
532
+ ">EXPLORE THE CODE →</p>
533
+ <a href="https://github.com/Eric-Chung-0511/Learning-Record/tree/main/Data%20Science%20Projects/PawMatchAI" style="text-decoration: none;">
534
+ <img src="https://img.shields.io/badge/GitHub-PawMatch_AI-007ACC?logo=github&style=for-the-badge">
535
+ </a>
536
+ </div>
537
+ ''')
538
+
539
+ return iface
540
+
541
+ if __name__ == "__main__":
542
+ iface = main()
543
+ iface.launch()
style_transfer.py ADDED
@@ -0,0 +1,752 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from PIL import Image, ImageEnhance
3
+ import numpy as np
4
+ import gradio as gr
5
+ from diffusers import StableDiffusionImg2ImgPipeline
6
+ import time
7
+ import os
8
+ import base64
9
+ from io import BytesIO
10
+
11
+ class DogStyleTransfer:
12
+ """
13
+ Class for handling dog image style transfer using Stable Diffusion.
14
+ This class manages model loading, image preprocessing, and style transfer operations.
15
+ """
16
+ def __init__(self):
17
+ self.models = {}
18
+ self.device = "cuda" if torch.cuda.is_available() else "cpu"
19
+
20
+ # Check xformers availability
21
+ self.xformers_available = False
22
+ try:
23
+ import xformers
24
+ self.xformers_available = True
25
+ print(f"xformers {xformers.__version__} is available and will be used for memory-efficient attention")
26
+ except ImportError:
27
+ print("xformers not found - will use default attention mechanism")
28
+ except Exception as e:
29
+ print(f"Error checking xformers: {str(e)} - will use default attention mechanism")
30
+
31
+ # Define style to model mapping based on availability
32
+ if self.device == "cuda":
33
+ self.style_model_mapping = {
34
+ "Japanese Anime Style": "Linaqruf/anything-v3.0", # Specialized for anime style
35
+ "Classic Cartoon": "nitrosocke/mo-di-diffusion", # Specialized for Disney style
36
+ "Oil Painting": "runwayml/stable-diffusion-v1-5", # Original model good for painting styles
37
+ "Watercolor": "dreamlike-art/dreamlike-photoreal-2.0", # Photorealistic art style model
38
+ "Cyberpunk": "dreamlike-art/dreamlike-diffusion-1.0" # Dreamlike style model
39
+ }
40
+ else:
41
+ # Lightweight models for CPU mode
42
+ self.style_model_mapping = {
43
+ "Japanese Anime Style": "runwayml/stable-diffusion-v1-5",
44
+ "Classic Cartoon": "runwayml/stable-diffusion-v1-5",
45
+ "Oil Painting": "runwayml/stable-diffusion-v1-5",
46
+ "Watercolor": "runwayml/stable-diffusion-v1-5",
47
+ "Cyberpunk": "runwayml/stable-diffusion-v1-5"
48
+ }
49
+
50
+ # style prompts with each feature
51
+ self.style_prompts = {
52
+ "Japanese Anime Style": "masterpiece, highest quality, genuine anime style illustration of a (dog:1.5), (bold anime aesthetics:1.5), (vibrant saturated colors:1.3), clean distinct lineart, stylized simplified features, expressive anime eyes, (preserve exact animal species:1.8), (maintain original animal breed:1.7), distinctive animal characteristics, (iconic anime art style:1.4), dramatic shading, flat color areas with highlight accents, simplified background elements, characteristic anime proportions, retain animal identity while stylizing, professional anime production quality, no watermarks, no signatures, (do not change animal species:1.8)",
53
+ "Classic Cartoon": "masterpiece, highest quality classic cartoon illustration of a dog, (golden age animation style:1.3), hand-drawn cel animation quality, bold clean outlines, (vibrant solid color fills:1.2), exaggerated expressive features, playful animated poses, classic Disney/Pixar influenced design, professional animation studio quality, simplified but expressive details, perfect smooth linework, rounded stylized forms, cheerful color palette, dynamic motion lines, classic cartoon physics, expressive oversized eyes, joyful personality captured, squash and stretch principles applied, classic cartoon proportions, professional character design, perfect animation keyframe quality, appealing character expression, masterful use of simple shapes, iconic cartoon aesthetic, no watermarks, no signatures",
54
+ "Oil Painting": "masterpiece, museum quality oil painting of a dog, (impasto technique:1.3), visible textured brushstrokes, layered oil pigments, rich depth of color, classical composition, (dramatic chiaroscuro lighting:1.2), Renaissance painting technique, glazing layers, sophisticated color harmony, warm and cool tones balance, expert painterly details, canvas texture visible, traditional realistic portrait style, fine art quality, gallery exhibition standard, rich shadows and highlights, volumetric form definition, atmospheric perspective, professional oil painting techniques, traditional varnished finish, color complexity with subtle undertones, expertly captured fur textures, strong compositional focus, emotional depth, timeless artistic quality, no watermarks, no signatures",
55
+ "Watercolor": "masterpiece, highest quality watercolor painting of a dog, (wet-on-wet technique:1.3), flowing color blends, translucent paint layers, visible paper texture, (controlled paint blooms:1.2), delicate color washes, spontaneous paint flow, preserved white spaces, soft color bleeding effects, subtle granulation textures, feathered edges, luminous transparency, loose expressive brushwork, artistic color pooling, gradient color transitions, minimalist background, playful splatter accents, artistic negative space usage, light-filled composition, watercolor paper grain visible, atmospheric color diffusion, professional traditional watercolor techniques, delicate brush details combined with flowing textures, no watermarks, no signatures",
56
+ "Cyberpunk": "masterpiece, highest quality, hyper-detailed cyberpunk digital art of a dog, (advanced technological integration:1.4), holographic collar interface, bionic limb enhancements, neural implant visuals, data visualization overlay, augmented reality HUD elements, (neon light reflections:1.3), wet street reflections, volumetric fog effects, urban dystopian background, megacity skyline, glowing circuitry details, optical fiber accents, synthetic materials, dramatic neon-lit contrast, cybernetic enhancements, high tech visors, digital distortion effects, information flow visualization, glitchy textures, metallic surfaces with advanced patina, dark atmospheric tone with vibrant neon accents, electrical energy effects, retro-futuristic design elements, near-future technology aesthetic, no watermarks, no signatures"
57
+ }
58
+
59
+ # Feature preservation prompts with weighted emphasis
60
+ self.feature_preservation = {
61
+ "common": "faithful representation of original animal species:(1.6), preserve original animal face structure:(1.5), maintain exact species characteristics:(1.4), accurate distinctive features:(1.3), consistent anatomical structure:(1.2), recognizable animal identity",
62
+ "Japanese Anime Style": "anime style dog with preserved realistic proportions, distinctive dog breed characteristics maintained, dog facial features clearly recognizable",
63
+ "Classic Cartoon": "cartoon style with accurate dog proportions, characteristic breed features preserved, recognizable dog expressions",
64
+ "Oil Painting": "oil painting technique while maintaining anatomical accuracy, realistic dog proportions, distinctive breed characteristics",
65
+ "Watercolor": "watercolor aesthetic with precise breed representation, accurate dog anatomy, distinctive dog features preserved",
66
+ "Cyberpunk": "cyberpunk elements while maintaining accurate dog proportions, recognizable breed features, true-to-life dog expression"
67
+ }
68
+
69
+ # Negative prompts
70
+ self.negative_prompts = {
71
+ "common": "deformed, distorted, disfigured, poorly drawn, bad anatomy, wrong anatomy, extra limbs, missing limbs, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, watermark, signature, text, change of species, wrong animal species, incorrect animal type, different animal, human features",
72
+ "dog_specific": "human face, human features, anthropomorphic, humanoid, human-like features, cartoon eyes, unrealistic eyes",
73
+ "Japanese Anime Style": "photorealistic, 3d render, western cartoon style, pixar style, realistic textured skin",
74
+ "Classic Cartoon": "anime style, manga, realistic, detailed skin texture, painterly, sketch, watercolor style",
75
+ "Oil Painting": "flat colors, digital art, cartoon, cell shaded, smooth texture, anime style",
76
+ "Watercolor": "digital art, 3d render, vector art, perfect linework, hard edges, bold lines",
77
+ "Cyberpunk": "watercolor paint, oil painting, natural scene, traditional art, vintage style, soft colors",
78
+ "species_preservation": "species transformation, change of animal type, incorrect animal features, wrong animal proportions, mixed animal characteristics"
79
+ }
80
+
81
+ # Style descriptions for UI display
82
+ self.style_descriptions = {
83
+ "Japanese Anime Style": "Characterized by vibrant colors, large expressive eyes, and stylized features common in Japanese animation.",
84
+ "Classic Cartoon": "Friendly, rounded features with bold outlines and bright colors typical of classic animated films.",
85
+ "Oil Painting": "Rich textures and depth created through visible brushstrokes and layered color application.",
86
+ "Watercolor": "Soft, transparent washes of color with flowing transitions and subtle color blending.",
87
+ "Cyberpunk": "Futuristic sci-fi aesthetic with neon colors, high contrast, and technological elements."
88
+ }
89
+
90
+ # Set model cache path
91
+ self.model_cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "dog_style_transfer")
92
+ os.makedirs(self.model_cache_dir, exist_ok=True)
93
+
94
+ # Display system info for debugging
95
+ self._print_system_info()
96
+
97
+ def _print_system_info(self):
98
+ """Print system information for debugging purposes"""
99
+ print("\n===== System Information =====")
100
+ print(f"Device: {self.device}")
101
+ print(f"PyTorch version: {torch.__version__}")
102
+
103
+ if self.device == "cuda":
104
+ print(f"CUDA available: {torch.cuda.is_available()}")
105
+ print(f"CUDA version: {torch.version.cuda if hasattr(torch.version, 'cuda') else 'Unknown'}")
106
+ print(f"GPU: {torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'Not available'}")
107
+ print(f"GPU memory: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB" if torch.cuda.is_available() else "Not available")
108
+
109
+ print(f"xformers available: {self.xformers_available}")
110
+ print("============================\n")
111
+
112
+ def load_model(self, style_name):
113
+ """Load the appropriate model based on style, handling xformers compatibility"""
114
+ # Get model ID for the style
115
+ model_id = self.style_model_mapping.get(style_name, "runwayml/stable-diffusion-v1-5")
116
+
117
+ # Check if model is already loaded
118
+ if model_id not in self.models:
119
+ print(f"Loading model {model_id} for {style_name} style...")
120
+
121
+ try:
122
+ # Load model with cache directory
123
+ model = StableDiffusionImg2ImgPipeline.from_pretrained(
124
+ model_id,
125
+ cache_dir=self.model_cache_dir,
126
+ torch_dtype=torch.float16 if self.device == "cuda" else torch.float32,
127
+ safety_checker=None # Remove safety checker to improve speed
128
+ )
129
+
130
+ if self.device == "cuda":
131
+ model = model.to("cuda")
132
+ # Enable memory optimization
133
+ model.enable_attention_slicing()
134
+
135
+ # Try to enable xformers
136
+ try:
137
+ if hasattr(model, 'enable_xformers_memory_efficient_attention'):
138
+ print("Attempting to enable xformers memory efficient attention...")
139
+ model.enable_xformers_memory_efficient_attention()
140
+ print("xformers memory efficient attention enabled successfully!")
141
+ except Exception as e:
142
+ print(f"Warning: Could not enable xformers memory efficient attention: {e}")
143
+ print("Proceeding without xformers optimization - this may use more memory but should still work.")
144
+
145
+ # Store model
146
+ self.models[model_id] = model
147
+ print(f"Model {model_id} loaded successfully!")
148
+ except Exception as e:
149
+ print(f"Error loading model {model_id}: {str(e)}")
150
+ # Fall back to basic model if specific model fails
151
+ if model_id != "runwayml/stable-diffusion-v1-5":
152
+ print("Falling back to default model...")
153
+ return self.load_model("Oil Painting") # Use generic model as fallback
154
+ raise
155
+
156
+ return self.models[model_id]
157
+
158
+ def preprocess_image(self, image, animal_type='dog'):
159
+ """Enhanced preprocessing for dog images before style transfer"""
160
+ # Convert to PIL image if needed
161
+ if isinstance(image, np.ndarray):
162
+ # Handle RGBA images by converting to RGB
163
+ if image.shape[2] == 4:
164
+ image = image[:, :, :3]
165
+ image = Image.fromarray(np.uint8(image))
166
+
167
+ # Resize while maintaining aspect ratio
168
+ width, height = image.size
169
+ max_size = 512 # SD models typically use 512x512 input
170
+ scaling_factor = min(max_size / width, max_size / height)
171
+ new_width = int(width * scaling_factor)
172
+ new_height = int(height * scaling_factor)
173
+ image = image.resize((new_width, new_height), Image.LANCZOS)
174
+
175
+ # Enhance contrast to emphasize dog features
176
+ enhancer = ImageEnhance.Contrast(image)
177
+ image = enhancer.enhance(1.2) # Slightly enhance contrast
178
+
179
+ # Sharpen to improve detail
180
+ enhancer = ImageEnhance.Sharpness(image)
181
+ image = enhancer.enhance(1.3) # Enhance sharpness
182
+
183
+ # Pad if not 512x512, instead of cropping
184
+ if new_width != 512 or new_height != 512:
185
+ new_img = Image.new("RGB", (512, 512), (255, 255, 255))
186
+ # Center the resized image
187
+ offset = ((512 - new_width) // 2, (512 - new_height) // 2)
188
+ new_img.paste(image, offset)
189
+ image = new_img
190
+
191
+ if animal_type != 'dog':
192
+ self.feature_preservation['common'] = 'strict preservation of original animal species:(1.8),' + self.feature_preservation["common"]
193
+
194
+ return image
195
+
196
+ def transform_style(self, image, style_name, strength=0.75, guidance_scale=7.5):
197
+ """
198
+ Transform image to selected style with improved prompts and parameters
199
+
200
+ Args:
201
+ image: Input image
202
+ style_name: Name of the style to apply
203
+ strength: Style transformation strength (0-1)
204
+ guidance_scale: Guidance scale for stable diffusion
205
+
206
+ Returns:
207
+ tuple: (transformed_image, error_message)
208
+ """
209
+ try:
210
+ if image is None:
211
+ return None, "Please upload a dog image first!"
212
+
213
+ start_time = time.time()
214
+ print(f"Starting style transfer: {style_name}")
215
+
216
+ # Adjust parameters based on style
217
+ if style_name == "Japanese Anime Style":
218
+ guidance_scale = 9.0 # Higher guidance for anime style
219
+ strength = 0.8
220
+ num_steps = 50
221
+ elif style_name == "Classic Cartoon":
222
+ guidance_scale = 8.0
223
+ strength = 0.75
224
+ num_steps = 40
225
+ elif style_name == "Oil Painting" or style_name == "Watercolor":
226
+ guidance_scale = 8.0 # Medium guidance for art styles
227
+ strength = 0.85
228
+ num_steps = 50
229
+ elif style_name == "Cyberpunk":
230
+ guidance_scale = 10.0 # Very high guidance for cyberpunk
231
+ strength = 0.85
232
+ num_steps = 50
233
+ else:
234
+ num_steps = 40
235
+
236
+ # Load model for style
237
+ try:
238
+ pipe = self.load_model(style_name)
239
+ except Exception as e:
240
+ print(f"Failed to load specific model for {style_name}: {str(e)}")
241
+ # Fall back to default model
242
+ pipe = self.load_model("Oil Painting")
243
+
244
+ # Enhanced image preprocessing
245
+ pil_image = self.preprocess_image(image)
246
+
247
+ # Get style prompt and add feature preservation
248
+ base_prompt = self.style_prompts.get(style_name, "digital art style, a dog")
249
+
250
+ # Feature preservation prompts - combining common and style-specific
251
+ feature_preservation = f"{self.feature_preservation['common']}, {self.feature_preservation.get(style_name, '')}"
252
+
253
+ # Enhanced positive prompt with feature preservation
254
+ prompt = f"{base_prompt}, {feature_preservation}, (high quality, detailed, sharp focus, professional photography):(1.2)"
255
+
256
+ # Use negative prompt - combining common and style-specific
257
+ negative_prompt = f"{self.negative_prompts['common']}, {self.negative_prompts['dog_specific']}, {self.negative_prompts.get(style_name, '')}"
258
+
259
+ print(f"Using prompt: {prompt}")
260
+ print(f"Using negative prompt: {negative_prompt}")
261
+ print(f"Transformation parameters - Strength: {strength}, Guidance Scale: {guidance_scale}, Steps: {num_steps}")
262
+
263
+ # Limit steps if too large to avoid memory issues
264
+ if num_steps > 60 and self.device == "cuda":
265
+ print("Reducing inference steps to save memory")
266
+ num_steps = 60
267
+
268
+ try:
269
+ # Generate transformed image
270
+ result = pipe(
271
+ prompt=prompt,
272
+ negative_prompt=negative_prompt,
273
+ image=pil_image,
274
+ strength=strength,
275
+ guidance_scale=guidance_scale,
276
+ num_inference_steps=num_steps
277
+ ).images[0]
278
+
279
+ except RuntimeError as e:
280
+ # Handle CUDA out of memory errors
281
+ if "CUDA out of memory" in str(e):
282
+ print("CUDA out of memory error, trying with reduced parameters")
283
+ # Retry with lower settings
284
+ return self._retry_with_lower_settings(pipe, prompt, negative_prompt, pil_image, strength, guidance_scale)
285
+ else:
286
+ # Try without negative prompt
287
+ print(f"Error with negative prompt, retrying without it: {str(e)}")
288
+ try:
289
+ result = pipe(
290
+ prompt=prompt,
291
+ image=pil_image,
292
+ strength=strength,
293
+ guidance_scale=guidance_scale,
294
+ num_inference_steps=30 # Reduce steps
295
+ ).images[0]
296
+ except Exception as retry_error:
297
+ print(f"Retry also failed: {str(retry_error)}")
298
+ raise
299
+
300
+ proc_time = time.time() - start_time
301
+ print(f"Style transfer completed in {proc_time:.2f} seconds")
302
+
303
+ return np.array(result), None
304
+
305
+ except Exception as e:
306
+ error_message = str(e)
307
+ # Provide user-friendly error messages
308
+ if "xformers" in error_message.lower():
309
+ print(f"xformers related error: {error_message}")
310
+ return None, "Style transfer error: xformers optimization unavailable, but functionality not affected. Please click 'Transform Style' button again to continue."
311
+ elif "CUDA out of memory" in error_message:
312
+ print(f"CUDA memory error: {error_message}")
313
+ return None, "GPU memory insufficient. Try reducing parameters or using a smaller image."
314
+ else:
315
+ print(f"Error during style transfer: {error_message}")
316
+ return None, f"Style transfer error: {error_message}"
317
+
318
+
319
+ def _retry_with_lower_settings(self, pipe, prompt, negative_prompt, image, strength, guidance_scale):
320
+ """Retry with lower settings when memory is insufficient"""
321
+ try:
322
+ # First attempt: Reduce inference steps
323
+ print("Attempting with lower settings (steps=20)...")
324
+ result = pipe(
325
+ prompt=prompt,
326
+ negative_prompt=negative_prompt,
327
+ image=image,
328
+ strength=strength,
329
+ guidance_scale=guidance_scale,
330
+ num_inference_steps=20 # Significantly reduce steps
331
+ ).images[0]
332
+ return np.array(result), None
333
+
334
+ except Exception as first_error:
335
+ # Log first failure
336
+ print(f"First retry attempt failed: {str(first_error)}")
337
+
338
+ # Second attempt: Minimum settings
339
+ try:
340
+ print("Attempting with minimum settings (steps=15, strength=0.6)...")
341
+ result = pipe(
342
+ prompt=prompt,
343
+ image=image,
344
+ strength=0.6, # Lower strength
345
+ guidance_scale=7.0, # Use standard setting
346
+ num_inference_steps=15 # Minimum steps
347
+ ).images[0]
348
+ return np.array(result), None
349
+
350
+ except Exception as second_error:
351
+ # Log all failures
352
+ print(f"Second retry attempt also failed: {str(second_error)}")
353
+ print("All retry attempts failed")
354
+
355
+ # Return clear error message
356
+ error_msg = f"Unable to complete style transfer, even with minimal settings: {str(second_error)}"
357
+ return None, error_msg
358
+
359
+ def get_available_styles(self):
360
+ """Get all available style options"""
361
+ return list(self.style_prompts.keys())
362
+
363
+ def get_style_description(self, style_name):
364
+ """Get description for a specific style"""
365
+ return self.style_descriptions.get(style_name, "")
366
+
367
+ def get_model_info(self, style_name):
368
+ """Get the model information for a specific style"""
369
+ model_id = self.style_model_mapping.get(style_name, "runwayml/stable-diffusion-v1-5")
370
+ return f"Powered by: {model_id}"
371
+
372
+ def get_image_download_link(self, image):
373
+ """
374
+ Generate a data URL for downloading the image
375
+
376
+ Args:
377
+ image: PIL Image or numpy array
378
+
379
+ Returns:
380
+ str: Base64 encoded data URL
381
+ """
382
+ if image is None:
383
+ return None
384
+
385
+ # Convert numpy array to PIL Image if needed
386
+ if isinstance(image, np.ndarray):
387
+ image = Image.fromarray(np.uint8(image))
388
+
389
+ # Save image to bytes buffer
390
+ buffer = BytesIO()
391
+ image.save(buffer, format="PNG")
392
+ img_str = base64.b64encode(buffer.getvalue()).decode()
393
+
394
+ return f"data:image/png;base64,{img_str}"
395
+
396
+ def create_style_transfer_tab(dog_style_transfer):
397
+ """Create style transfer tab with UI components"""
398
+
399
+ with gr.Column():
400
+ gr.Markdown("""
401
+ # 🎨 Dog Style Transformation
402
+
403
+ Transform your dog photos into different artistic styles! Upload a dog picture, choose your preferred style, and create unique artwork.
404
+ """)
405
+
406
+ # Add model info and style description display
407
+ gr.Markdown("""
408
+ <div style="background-color: #f0f8ff; padding: 16px; border-radius: 8px; margin: 16px 0; border: 1px solid #cce5ff; box-shadow: 0 3px 10px rgba(0,0,123,0.1);">
409
+ <h3>🐶 Upload a dog photo and select an artistic style</h3>
410
+ <p>After uploading your dog photo, the system will transform it into your chosen artistic style. Try different styles to create stunning effects!</p>
411
+ <p>Our system uses specialized models for each style to ensure the best results.</p>
412
+ <p style="margin-top: 10px; padding: 8px; background-color: #fff9e6; border-left: 4px solid #ffd966; border-radius: 4px;"><b>⏱️ Patience is a virtue!</b> While AI is working its magic, your dog might have time to learn a new trick or two. The transformation can take up to 30 seconds, depending on how photogenic your furry friend is! 🐾</p>
413
+ <p style="margin-top: 10px; padding: 8px; background-color: #e6f9ff; border-left: 4px solid #66c2ff; border-radius: 4px;"><b>🤫 A Little Secret:</b> Although we designed this tool for dogs, it can actually transform any photo! Portraits, landscapes, even your favorite teddy bear — feel free to try them all! Just don’t tell the other dogs… they might get jealous! 😉</p>
414
+ <p style="margin-top: 10px; padding: 8px; background-color: #e6f9e6; border-left: 4px solid #66cc77; border-radius: 4px;"><b>✨ Unlimited Creativity!</b> Sometimes, AI might surprise you with unexpected creative interpretations, adding unique colors or features to your image. ✨</p>
415
+ </div>
416
+ """)
417
+
418
+ with gr.Row():
419
+ with gr.Column(scale=1):
420
+ # Upload image component
421
+ input_image = gr.Image(
422
+ label="Upload Dog Photo",
423
+ type="numpy"
424
+ )
425
+
426
+ style_dropdown = gr.Dropdown(
427
+ choices=dog_style_transfer.get_available_styles(),
428
+ value=dog_style_transfer.get_available_styles()[0],
429
+ label="Select Artistic Style"
430
+ )
431
+
432
+ # Display style description
433
+ style_description = gr.Markdown(
434
+ dog_style_transfer.get_style_description(dog_style_transfer.get_available_styles()[0])
435
+ )
436
+
437
+ # Display model info
438
+ model_info = gr.Markdown(
439
+ dog_style_transfer.get_model_info(dog_style_transfer.get_available_styles()[0])
440
+ )
441
+
442
+ with gr.Row():
443
+ strength_slider = gr.Slider(
444
+ minimum=0.3,
445
+ maximum=0.9,
446
+ value=0.75,
447
+ step=0.05,
448
+ label="Style Intensity (lower values preserve more original details)"
449
+ )
450
+
451
+ # customize Transform style buttom
452
+ style_button = gr.Button("Transform Style", variant="primary")
453
+
454
+ gr.Markdown("""
455
+ <style>
456
+ button.primary {
457
+ background: linear-gradient(90deg, #ff6b6b, #ffa36b, #ffd56b) !important;
458
+ color: white !important;
459
+ font-weight: 600 !important;
460
+ text-shadow: 0 1px 1px rgba(0,0,0,0.2) !important;
461
+ border: none !important;
462
+ }
463
+
464
+ button.primary:hover {
465
+ background: linear-gradient(90deg, #ff5b5b, #ff936b, #ffcf6b) !important;
466
+ box-shadow: 0 4px 8px rgba(0,0,0,0.3) !important;
467
+ }
468
+ </style>
469
+ """)
470
+
471
+ # Progress indicator
472
+ status_indicator = gr.Textbox(
473
+ label="Status",
474
+ value="Upload an image and press 'Transform Style' to begin",
475
+ interactive=False
476
+ )
477
+
478
+ error_output = gr.Textbox(
479
+ visible=False,
480
+ label="Error Message"
481
+ )
482
+
483
+ with gr.Column(scale=1):
484
+ # Output image component
485
+ output_image = gr.Image(
486
+ label="Style Transformation Result"
487
+ )
488
+
489
+ # Hidden component to store the download link
490
+ download_link = gr.HTML(visible=False)
491
+
492
+ # HTML component for actual download
493
+ download_html = gr.HTML(visible=False)
494
+
495
+ gr.Markdown("""
496
+ ### Tips for Best Results
497
+ - Use images with clear dog features and good lighting
498
+ - For best results, use images where the dog is the main subject
499
+ - Different styles work better with different dog breeds
500
+ - Lower the style intensity to preserve more original details
501
+ """)
502
+
503
+ gr.HTML("""
504
+ <style>
505
+ .style-box {
506
+ background: linear-gradient(145deg, #ffffff, #f5f7fa);
507
+ border-radius: 12px;
508
+ box-shadow: 0 4px 20px rgba(0,0,0,0.08);
509
+ padding: 25px 30px;
510
+ margin: 30px 0;
511
+ border: 1px solid rgba(0,0,0,0.05);
512
+ position: relative;
513
+ }
514
+
515
+ .style-box::before {
516
+ content: '';
517
+ position: absolute;
518
+ top: 0;
519
+ left: 0;
520
+ width: 6px;
521
+ height: 100%;
522
+ background: linear-gradient(to bottom, #ff6b6b, #ffa36b, #ffd56b);
523
+ border-radius: 6px 0 0 6px;
524
+ }
525
+
526
+ .style-box h2 {
527
+ color: #333;
528
+ font-size: 24px;
529
+ margin-bottom: 20px;
530
+ padding-bottom: 10px;
531
+ border-bottom: 2px solid #f0f0f0;
532
+ }
533
+
534
+ .style-name {
535
+ font-weight: bold;
536
+ color: #333;
537
+ }
538
+
539
+ .style-desc {
540
+ margin-bottom: 15px;
541
+ padding-bottom: 15px;
542
+ border-bottom: 1px solid #f0f0f0;
543
+ }
544
+
545
+ .style-desc:last-child {
546
+ margin-bottom: 0;
547
+ padding-bottom: 0;
548
+ border-bottom: none;
549
+ }
550
+ </style>
551
+
552
+ <div class="style-box">
553
+ <h2>Style Effect Descriptions</h2>
554
+ <p>Each style transforms your dog photo in a unique way:</p>
555
+
556
+ <div class="style-desc">
557
+ <p><span class="style-name">Japanese Anime Style:</span> Vibrant artwork with fluid animation qualities, expressive features, and dramatic lighting effects. Features soft color gradients, detailed line work, and emotional depth.</p>
558
+ </div>
559
+
560
+ <div class="style-desc">
561
+ <p><span class="style-name">Classic Cartoon:</span> Traditional animation style with bold outlines, solid color fills, and playful character design. Displays exaggerated expressions, simplified forms, and dynamic poses.</p>
562
+ </div>
563
+
564
+ <div class="style-desc">
565
+ <p><span class="style-name">Oil Painting:</span> Classical art technique with visible textured brushstrokes and layered color application. Shows rich depth, dramatic lighting contrast, and sophisticated color harmony.</p>
566
+ </div>
567
+
568
+ <div class="style-desc">
569
+ <p><span class="style-name">Watercolor:</span> Delicate painting style with flowing color blends and translucent layers. Features soft edges, color bleeding effects, and visible paper texture elements.</p>
570
+ </div>
571
+
572
+ <div class="style-desc">
573
+ <p><span class="style-name">Cyberpunk:</span> High-tech futuristic aesthetic with advanced technological elements and neon accents. Incorporates holographic interfaces, digital effects, and urban dystopian elements.</p>
574
+ </div>
575
+ </div>
576
+ """)
577
+
578
+ # Setup event triggers
579
+ def update_progress(value, desc):
580
+ """Update progress bar and description"""
581
+ return gr.update(value=value), gr.update(value=desc)
582
+
583
+ def process_style_transfer(image, style, strength):
584
+ """Process style transfer and prepare download options"""
585
+ if image is None:
586
+ return (
587
+ None,
588
+ gr.update(visible=True, value="Please upload a dog image first!"),
589
+ gr.update(visible=False),
590
+ gr.update(visible=False),
591
+ gr.update(visible=False),
592
+ gr.update(value="Upload an image and press 'Transform Style' to begin")
593
+ )
594
+
595
+ # Display processing status
596
+ status_message = "Processing your image... This may take a moment."
597
+
598
+ # Perform style transfer
599
+ result, error = dog_style_transfer.transform_style(
600
+ image,
601
+ style,
602
+ strength
603
+ )
604
+
605
+ if error:
606
+ return (
607
+ None,
608
+ gr.update(visible=True, value=error),
609
+ gr.update(visible=False),
610
+ gr.update(visible=False),
611
+ gr.update(visible=False),
612
+ gr.update(value="Error occurred. Please try again.")
613
+ )
614
+
615
+ # Generate download link for the image
616
+ if result is not None:
617
+ pil_image = Image.fromarray(result)
618
+ download_data = dog_style_transfer.get_image_download_link(pil_image)
619
+ download_html_content = f"""
620
+ <style>
621
+ .download-btn {{
622
+ display: inline-block;
623
+ background: linear-gradient(90deg, #3498db, #2ecc71);
624
+ color: white !important; /* 確保文字為白色 */
625
+ text-shadow: 0 1px 2px rgba(0,0,0,0.3) !important; /* 增強文字陰影使白色更突出 */
626
+ font-weight: 700 !important; /* 加粗字體 */
627
+ padding: 12px 24px;
628
+ text-align: center;
629
+ text-decoration: none;
630
+ font-size: 16px;
631
+ border-radius: 25px;
632
+ cursor: pointer;
633
+ transition: all 0.3s ease;
634
+ border: none;
635
+ box-shadow: 0 3px 6px rgba(0,0,0,0.16);
636
+ letter-spacing: 0.5px; /* 字母間距,提高可讀性 */
637
+ }}
638
+
639
+ .download-btn:hover {{
640
+ transform: translateY(-2px);
641
+ box-shadow: 0 5px 10px rgba(0,0,0,0.25);
642
+ background: linear-gradient(90deg, #2980b9, #27ae60);
643
+ }}
644
+
645
+ .download-btn:active {{
646
+ transform: translateY(0);
647
+ box-shadow: 0 2px 4px rgba(0,0,0,0.1);
648
+ }}
649
+ </style>
650
+
651
+ <a href="{download_data}" download="dog_{style.replace(' ', '_')}.png" class="download-btn">
652
+ Download Transformed Image
653
+ </a>
654
+ """
655
+
656
+ # Store download data in a hidden element
657
+ # We'll make this invisible to avoid showing base64 encoded data
658
+ hidden_download_data = download_data
659
+
660
+ return (
661
+ result,
662
+ gr.update(visible=False),
663
+ gr.update(visible=False, value=hidden_download_data), # Keep the data hidden
664
+ gr.update(visible=True, value=download_html_content), # Show the HTML button
665
+ gr.update(value="Transform Completed! You can download the image")
666
+ )
667
+ else:
668
+ # Handle the case where result is None but no error was returned
669
+ return (
670
+ None,
671
+ gr.update(visible=True, value="Style transfer failed with no specific error. Please try again."),
672
+ gr.update(visible=False),
673
+ gr.update(visible=False),
674
+ gr.update(visible=False),
675
+ gr.update(value="Something went wrong. Please try again.")
676
+ )
677
+
678
+ # Update style description and model info
679
+ def update_style_info(style):
680
+ return dog_style_transfer.get_style_description(style), dog_style_transfer.get_model_info(style)
681
+
682
+ style_button.click(
683
+ fn=process_style_transfer,
684
+ inputs=[input_image, style_dropdown, strength_slider],
685
+ outputs=[
686
+ output_image,
687
+ error_output,
688
+ download_link,
689
+ download_html,
690
+ ]
691
+ )
692
+
693
+ style_dropdown.change(
694
+ fn=update_style_info,
695
+ inputs=[style_dropdown],
696
+ outputs=[style_description, model_info]
697
+ )
698
+
699
+
700
+ # Add example images
701
+ example_dogs = [
702
+ ["Border_Collie.jpg", "Japanese Anime Style"],
703
+ ["Golden_Retriever.jpeg", "Classic Cartoon"],
704
+ ["Saint_Bernard.jpeg", "Oil Painting"],
705
+ ["Samoyed.jpeg", "Watercolor"],
706
+ ["French_Bulldog.jpeg", "Cyberpunk"]
707
+ ]
708
+
709
+ # Check if Examples feature is available
710
+ try:
711
+ gr.Examples(
712
+ examples=example_dogs,
713
+ inputs=[input_image, style_dropdown]
714
+ )
715
+ except Exception as e:
716
+ print(f"Note: Examples feature not available in your Gradio version: {e}")
717
+
718
+ gr.HTML("""
719
+ <style>
720
+ .attribution-box {
721
+ font-size: 0.85em;
722
+ color: #666;
723
+ margin-top: 20px;
724
+ padding: 18px;
725
+ border-radius: 8px;
726
+ background-color: #f8f9fa;
727
+ border: 1px solid #e9ecef;
728
+ font-style: italic;
729
+ }
730
+
731
+ .attribution-box h4 {
732
+ margin-top: 0;
733
+ color: #495057;
734
+ font-style: normal;
735
+ font-weight: 600;
736
+ margin-bottom: 12px;
737
+ }
738
+
739
+ .attribution-box p {
740
+ margin: 8px 0;
741
+ line-height: 1.5;
742
+ }
743
+ </style>
744
+
745
+ <div class="attribution-box">
746
+ <h4>Attribution</h4>
747
+ <p>This application uses pre-trained diffusion models from Hugging Face for image style transfer. All models are used according to their respective open source licenses for educational and non-commercial purposes.</p>
748
+ <p>Powered by the open source Diffusers library from Hugging Face.</p>
749
+ </div>
750
+ """)
751
+
752
+ return input_image, style_dropdown, style_button, output_image