utkarsh134 commited on
Commit
887da11
·
verified ·
1 Parent(s): 6c4dfb2

2 variation for normal banner

Browse files
Files changed (1) hide show
  1. app/services/banner_service.py +255 -254
app/services/banner_service.py CHANGED
@@ -1,255 +1,256 @@
1
- from io import BytesIO
2
- from typing import Dict
3
-
4
- from google import genai
5
- from google.genai import types
6
- from PIL import Image, ImageEnhance, ImageOps, ImageFilter
7
- import logging
8
- from app.config.settings import settings
9
- from app.utils.reach_upload import upload_to_reach
10
- from app.utils.name_generator import generate_brand_name
11
- import math
12
-
13
- logger = logging.getLogger(__name__)
14
- logging.basicConfig(level=logging.INFO)
15
-
16
- def get_banner_prompts(brand_details: str, primary_color: str) -> Dict[str, str]:
17
- device_prompts = {
18
- 'desktop': f'''Generate a realistic image which includes real people and
19
- The image must show essence of the brand:
20
-
21
- {brand_details}
22
-
23
- Exclude the use of color {primary_color} in the palette. Do not Include any text, typographic or Alphanumeric Elements.''',
24
- 'ipad': f'''Generate a realistic image which includes real people and
25
- The image must show essence of the brand:
26
-
27
- {brand_details}
28
-
29
- Exclude the use of color {primary_color} in the palette. Do not Include any text, typographic or Alphanumeric Elements.''',
30
- 'mobile': f'''Generate a realistic image which includes real people and
31
- The image must show essence of the brand:
32
-
33
- {brand_details}
34
-
35
- Exclude the use of color {primary_color} in the palette. Do not Include any text, typographic or Alphanumeric Elements.''',
36
- }
37
- return device_prompts
38
-
39
- def process_image(image_data: bytes, target_width: int, target_height: int) -> bytes:
40
- """
41
- Resize, sharpen, enhance and compress an image to fit a given resolution and file size.
42
- Optimized for 1024x1024 source images with better quality preservation.
43
- """
44
- img = Image.open(BytesIO(image_data)).convert('RGB')
45
- original_width, original_height = img.size
46
-
47
- # Calculate aspect ratios
48
- original_ratio = original_width / original_height
49
- target_ratio = target_width / target_height
50
-
51
- # Step 1: Smart upscaling based on target size relative to source
52
- # Only upscale if target is significantly larger than source
53
- upscale_threshold = max(target_width, target_height) / max(original_width, original_height)
54
-
55
- if upscale_threshold > 1.2: # Only upscale if target is 20% larger
56
- scale_factor = min(3, math.ceil(upscale_threshold)) # Cap at 3x to avoid excessive processing
57
- intermediate_width = original_width * scale_factor
58
- intermediate_height = original_height * scale_factor
59
-
60
- # Use bicubic for upscaling (better than Lanczos for enlargement)
61
- working_img = img.resize((intermediate_width, intermediate_height), Image.BICUBIC)
62
-
63
- # Apply subtle unsharp mask for upscaled images
64
- working_img = working_img.filter(ImageFilter.UnsharpMask(radius=1.0, percent=120, threshold=3))
65
- else:
66
- working_img = img
67
-
68
- # Step 2: Smart cropping/fitting based on aspect ratio difference
69
- if abs(original_ratio - target_ratio) > 0.1: # Significant aspect ratio difference
70
- # Crop to target ratio first, then resize (preserves more detail)
71
- if original_ratio > target_ratio:
72
- # Source is wider - crop width
73
- new_width = int(working_img.height * target_ratio)
74
- left = (working_img.width - new_width) // 2
75
- working_img = working_img.crop((left, 0, left + new_width, working_img.height))
76
- else:
77
- # Source is taller - crop height
78
- new_height = int(working_img.width / target_ratio)
79
- top = (working_img.height - new_height) // 2
80
- working_img = working_img.crop((0, top, working_img.width, top + new_height))
81
-
82
- # Step 3: Final resize to target dimensions
83
- # Use Lanczos for downscaling, bicubic for upscaling
84
- current_size = max(working_img.size)
85
- target_size = max(target_width, target_height)
86
-
87
- if current_size > target_size:
88
- # Downscaling - use Lanczos
89
- final_img = working_img.resize((target_width, target_height), Image.LANCZOS)
90
- else:
91
- # Upscaling - use bicubic
92
- final_img = working_img.resize((target_width, target_height), Image.BICUBIC)
93
-
94
- # Step 4: Conditional enhancement based on target size
95
- target_pixels = target_width * target_height
96
-
97
- if target_pixels < 500000: # Small images need more sharpening
98
- final_img = ImageEnhance.Sharpness(final_img).enhance(1.4)
99
- final_img = ImageEnhance.Contrast(final_img).enhance(1.15)
100
- final_img = ImageEnhance.Color(final_img).enhance(1.05)
101
- elif target_pixels < 1500000: # Medium images
102
- final_img = ImageEnhance.Sharpness(final_img).enhance(1.2)
103
- final_img = ImageEnhance.Contrast(final_img).enhance(1.1)
104
- final_img = ImageEnhance.Color(final_img).enhance(1.08)
105
- else: # Large images need subtle enhancement
106
- final_img = ImageEnhance.Sharpness(final_img).enhance(1.1)
107
- final_img = ImageEnhance.Contrast(final_img).enhance(1.05)
108
- final_img = ImageEnhance.Color(final_img).enhance(1.03)
109
-
110
- # Step 5: Intelligent compression with quality optimization
111
- buffer = BytesIO()
112
-
113
- # Start with higher quality for better results
114
- quality = 98
115
- max_file_size = 500 * 1024 # 500KB
116
-
117
- while quality >= 75: # Don't go below 75 quality
118
- buffer.seek(0)
119
- buffer.truncate()
120
-
121
- # Adjust subsampling based on quality
122
- if quality >= 95:
123
- subsampling = 0 # No chroma subsampling for highest quality
124
- elif quality >= 85:
125
- subsampling = 1 # Moderate chroma subsampling
126
- else:
127
- subsampling = 2 # Standard chroma subsampling
128
-
129
- final_img.save(
130
- buffer,
131
- format='JPEG',
132
- quality=quality,
133
- optimize=True,
134
- progressive=True,
135
- subsampling=subsampling
136
- )
137
-
138
- current_size = buffer.tell()
139
-
140
- # If we're within size limit or quality is getting too low, stop
141
- if current_size <= max_file_size or quality <= 80:
142
- break
143
-
144
- # Adjust quality reduction based on how far over we are
145
- overage_ratio = current_size / max_file_size
146
- if overage_ratio > 2:
147
- quality -= 8 # Aggressive reduction if way over
148
- elif overage_ratio > 1.5:
149
- quality -= 5 # Moderate reduction
150
- else:
151
- quality -= 3 # Conservative reduction
152
-
153
- return buffer.getvalue()
154
-
155
- def process_images_for_devices(image_data: bytes) -> Dict[str, bytes]:
156
- """
157
- Resize banner image for desktop, tablet (iPad), and mobile.
158
- """
159
- return {
160
- 'desktop': process_image(image_data, 1440, 500),
161
- 'ipad': process_image(image_data, 768, 500),
162
- 'mobile': process_image(image_data, 375, 500)
163
- }
164
-
165
- def generate_banner(brand_details: str, primary_color: str) -> Dict:
166
- """
167
- Generate multiple variations of banner images for each device using Gemini and upload to Reach platform.
168
- Returns a dictionary with URLs for different device sizes and prompt variations.
169
- """
170
- api_key = settings.GEMINI_API_KEY
171
- client = genai.Client(api_key=api_key)
172
- brand_name = generate_brand_name()
173
-
174
- # For now, we just create one prompt set — extendable later
175
- base_prompt = get_banner_prompts(brand_details, primary_color)
176
- all_prompt_sets = [base_prompt] # Add more prompt sets for more variations
177
-
178
- all_banner_results = []
179
-
180
- logger.info(f"Starting banner generation with ID: {brand_name}")
181
-
182
- # Device-specific dimensions
183
- device_dimensions = {
184
- 'desktop': (1440, 500),
185
- 'ipad': (768, 500),
186
- 'mobile': (375, 500)
187
- }
188
-
189
- # File types used for uploading
190
- device_file_types = {
191
- "desktop": "DESKTOP_IMAGE",
192
- "ipad": "IPAD_IMAGE",
193
- "mobile": "WEB_IMAGE"
194
- }
195
-
196
- # Loop through all variations of prompt sets
197
- for variation_index, device_prompts in enumerate(all_prompt_sets, start=1):
198
- banner_results = {}
199
-
200
- for device, prompt in device_prompts.items():
201
- logger.info(f"Generating variation {variation_index}, device: {device}")
202
-
203
- try:
204
- response = client.models.generate_content(
205
- model="gemini-2.0-flash-exp-image-generation",
206
- contents=prompt,
207
- config=types.GenerateContentConfig(response_modalities=['TEXT', 'IMAGE'])
208
- )
209
-
210
- for part in response.candidates[0].content.parts:
211
- if part.inline_data is not None:
212
- image_data = part.inline_data.data
213
- mime_type = part.inline_data.mime_type
214
-
215
- # Determine image format
216
- if "jpeg" in mime_type or "jpg" in mime_type:
217
- image_format = "jpeg"
218
- elif "png" in mime_type:
219
- image_format = "png"
220
- elif "webp" in mime_type:
221
- image_format = "webp"
222
- else:
223
- image_format = "jpeg"
224
-
225
- logger.info(f"Processing image for {device} in variation {variation_index}")
226
-
227
- # Resize and enhance
228
- width, height = device_dimensions[device]
229
- processed_image = process_image(image_data, width, height) # Upload to Reach with UUID-based name
230
- result = upload_to_reach(
231
- processed_image,
232
- f"banner_{brand_name}_{device}_v{variation_index}",
233
- image_format,
234
- 1,
235
- device_file_types[device]
236
- )
237
- banner_results[device] = result['url']
238
-
239
- logger.info(f"{device} banner uploaded successfully for variation {variation_index}")
240
-
241
- except Exception as error:
242
- logger.error(f"Error processing {device} banner in variation {variation_index}: {str(error)}")
243
- raise
244
-
245
- if not banner_results:
246
- logger.error(f"No images were generated for variation {variation_index}")
247
- raise Exception(f"No images were generated for variation {variation_index}")
248
-
249
- all_banner_results.append({
250
- "variation": variation_index,
251
- "banners": banner_results
252
- })
253
-
254
- logger.info("All banner variations generated successfully.")
 
255
  return {"variations": all_banner_results}
 
1
+ from io import BytesIO
2
+ from typing import Dict
3
+
4
+ from google import genai
5
+ from google.genai import types
6
+ from PIL import Image, ImageEnhance, ImageOps, ImageFilter
7
+ import logging
8
+ from app.config.settings import settings
9
+ from app.utils.reach_upload import upload_to_reach
10
+ from app.utils.name_generator import generate_brand_name
11
+ import math
12
+
13
+ logger = logging.getLogger(__name__)
14
+ logging.basicConfig(level=logging.INFO)
15
+
16
+ def get_banner_prompts(brand_details: str, primary_color: str) -> Dict[str, str]:
17
+ device_prompts = {
18
+ 'desktop': f'''Generate a realistic image which includes real people and
19
+ The image must show essence of the brand:
20
+
21
+ {brand_details}
22
+
23
+ Exclude the use of color {primary_color} in the palette. Do not Include any text, typographic or Alphanumeric Elements.''',
24
+ 'ipad': f'''Generate a realistic image which includes real people and
25
+ The image must show essence of the brand:
26
+
27
+ {brand_details}
28
+
29
+ Exclude the use of color {primary_color} in the palette. Do not Include any text, typographic or Alphanumeric Elements.''',
30
+ 'mobile': f'''Generate a realistic image which includes real people and
31
+ The image must show essence of the brand:
32
+
33
+ {brand_details}
34
+
35
+ Exclude the use of color {primary_color} in the palette. Do not Include any text, typographic or Alphanumeric Elements.''',
36
+ }
37
+ return device_prompts
38
+
39
+ def process_image(image_data: bytes, target_width: int, target_height: int) -> bytes:
40
+ """
41
+ Resize, sharpen, enhance and compress an image to fit a given resolution and file size.
42
+ Optimized for 1024x1024 source images with better quality preservation.
43
+ """
44
+ img = Image.open(BytesIO(image_data)).convert('RGB')
45
+ original_width, original_height = img.size
46
+
47
+ # Calculate aspect ratios
48
+ original_ratio = original_width / original_height
49
+ target_ratio = target_width / target_height
50
+
51
+ # Step 1: Smart upscaling based on target size relative to source
52
+ # Only upscale if target is significantly larger than source
53
+ upscale_threshold = max(target_width, target_height) / max(original_width, original_height)
54
+
55
+ if upscale_threshold > 1.2: # Only upscale if target is 20% larger
56
+ scale_factor = min(3, math.ceil(upscale_threshold)) # Cap at 3x to avoid excessive processing
57
+ intermediate_width = original_width * scale_factor
58
+ intermediate_height = original_height * scale_factor
59
+
60
+ # Use bicubic for upscaling (better than Lanczos for enlargement)
61
+ working_img = img.resize((intermediate_width, intermediate_height), Image.BICUBIC)
62
+
63
+ # Apply subtle unsharp mask for upscaled images
64
+ working_img = working_img.filter(ImageFilter.UnsharpMask(radius=1.0, percent=120, threshold=3))
65
+ else:
66
+ working_img = img
67
+
68
+ # Step 2: Smart cropping/fitting based on aspect ratio difference
69
+ if abs(original_ratio - target_ratio) > 0.1: # Significant aspect ratio difference
70
+ # Crop to target ratio first, then resize (preserves more detail)
71
+ if original_ratio > target_ratio:
72
+ # Source is wider - crop width
73
+ new_width = int(working_img.height * target_ratio)
74
+ left = (working_img.width - new_width) // 2
75
+ working_img = working_img.crop((left, 0, left + new_width, working_img.height))
76
+ else:
77
+ # Source is taller - crop height
78
+ new_height = int(working_img.width / target_ratio)
79
+ top = (working_img.height - new_height) // 2
80
+ working_img = working_img.crop((0, top, working_img.width, top + new_height))
81
+
82
+ # Step 3: Final resize to target dimensions
83
+ # Use Lanczos for downscaling, bicubic for upscaling
84
+ current_size = max(working_img.size)
85
+ target_size = max(target_width, target_height)
86
+
87
+ if current_size > target_size:
88
+ # Downscaling - use Lanczos
89
+ final_img = working_img.resize((target_width, target_height), Image.LANCZOS)
90
+ else:
91
+ # Upscaling - use bicubic
92
+ final_img = working_img.resize((target_width, target_height), Image.BICUBIC)
93
+
94
+ # Step 4: Conditional enhancement based on target size
95
+ target_pixels = target_width * target_height
96
+
97
+ if target_pixels < 500000: # Small images need more sharpening
98
+ final_img = ImageEnhance.Sharpness(final_img).enhance(1.4)
99
+ final_img = ImageEnhance.Contrast(final_img).enhance(1.15)
100
+ final_img = ImageEnhance.Color(final_img).enhance(1.05)
101
+ elif target_pixels < 1500000: # Medium images
102
+ final_img = ImageEnhance.Sharpness(final_img).enhance(1.2)
103
+ final_img = ImageEnhance.Contrast(final_img).enhance(1.1)
104
+ final_img = ImageEnhance.Color(final_img).enhance(1.08)
105
+ else: # Large images need subtle enhancement
106
+ final_img = ImageEnhance.Sharpness(final_img).enhance(1.1)
107
+ final_img = ImageEnhance.Contrast(final_img).enhance(1.05)
108
+ final_img = ImageEnhance.Color(final_img).enhance(1.03)
109
+
110
+ # Step 5: Intelligent compression with quality optimization
111
+ buffer = BytesIO()
112
+
113
+ # Start with higher quality for better results
114
+ quality = 98
115
+ max_file_size = 500 * 1024 # 500KB
116
+
117
+ while quality >= 75: # Don't go below 75 quality
118
+ buffer.seek(0)
119
+ buffer.truncate()
120
+
121
+ # Adjust subsampling based on quality
122
+ if quality >= 95:
123
+ subsampling = 0 # No chroma subsampling for highest quality
124
+ elif quality >= 85:
125
+ subsampling = 1 # Moderate chroma subsampling
126
+ else:
127
+ subsampling = 2 # Standard chroma subsampling
128
+
129
+ final_img.save(
130
+ buffer,
131
+ format='JPEG',
132
+ quality=quality,
133
+ optimize=True,
134
+ progressive=True,
135
+ subsampling=subsampling
136
+ )
137
+
138
+ current_size = buffer.tell()
139
+
140
+ # If we're within size limit or quality is getting too low, stop
141
+ if current_size <= max_file_size or quality <= 80:
142
+ break
143
+
144
+ # Adjust quality reduction based on how far over we are
145
+ overage_ratio = current_size / max_file_size
146
+ if overage_ratio > 2:
147
+ quality -= 8 # Aggressive reduction if way over
148
+ elif overage_ratio > 1.5:
149
+ quality -= 5 # Moderate reduction
150
+ else:
151
+ quality -= 3 # Conservative reduction
152
+
153
+ return buffer.getvalue()
154
+
155
+ def process_images_for_devices(image_data: bytes) -> Dict[str, bytes]:
156
+ """
157
+ Resize banner image for desktop, tablet (iPad), and mobile.
158
+ """
159
+ return {
160
+ 'desktop': process_image(image_data, 1440, 500),
161
+ 'ipad': process_image(image_data, 768, 500),
162
+ 'mobile': process_image(image_data, 375, 500)
163
+ }
164
+
165
+ def generate_banner(brand_details: str, primary_color: str) -> Dict:
166
+ """
167
+ Generate multiple variations of banner images for each device using Gemini and upload to Reach platform.
168
+ Returns a dictionary with URLs for different device sizes and prompt variations.
169
+ """
170
+ api_key = settings.GEMINI_API_KEY
171
+ client = genai.Client(api_key=api_key)
172
+ brand_name = generate_brand_name()
173
+
174
+ # Create two identical prompt sets for two variations
175
+ base_prompt = get_banner_prompts(brand_details, primary_color)
176
+ all_prompt_sets = [base_prompt, base_prompt] # Two variations, same prompts
177
+
178
+ all_banner_results = []
179
+
180
+ logger.info(f"Starting banner generation with ID: {brand_name}")
181
+
182
+ # Device-specific dimensions
183
+ device_dimensions = {
184
+ 'desktop': (1440, 500),
185
+ 'ipad': (768, 500),
186
+ 'mobile': (375, 500)
187
+ }
188
+
189
+ # File types used for uploading
190
+ device_file_types = {
191
+ "desktop": "DESKTOP_IMAGE",
192
+ "ipad": "IPAD_IMAGE",
193
+ "mobile": "WEB_IMAGE"
194
+ }
195
+
196
+ # Loop through all variations of prompt sets
197
+ for variation_index, device_prompts in enumerate(all_prompt_sets, start=1):
198
+ banner_results = {}
199
+
200
+ for device, prompt in device_prompts.items():
201
+ logger.info(f"Generating variation {variation_index}, device: {device}")
202
+
203
+ try:
204
+ response = client.models.generate_content(
205
+ model="gemini-2.0-flash-exp-image-generation",
206
+ contents=prompt,
207
+ config=types.GenerateContentConfig(response_modalities=['TEXT', 'IMAGE'])
208
+ )
209
+
210
+ for part in response.candidates[0].content.parts:
211
+ if part.inline_data is not None:
212
+ image_data = part.inline_data.data
213
+ mime_type = part.inline_data.mime_type
214
+
215
+ # Determine image format
216
+ if "jpeg" in mime_type or "jpg" in mime_type:
217
+ image_format = "jpeg"
218
+ elif "png" in mime_type:
219
+ image_format = "png"
220
+ elif "webp" in mime_type:
221
+ image_format = "webp"
222
+ else:
223
+ image_format = "jpeg"
224
+
225
+ logger.info(f"Processing image for {device} in variation {variation_index}")
226
+
227
+ # Resize and enhance
228
+ width, height = device_dimensions[device]
229
+ processed_image = process_image(image_data, width, height)
230
+ # Upload to Reach with UUID-based name
231
+ result = upload_to_reach(
232
+ processed_image,
233
+ f"banner_{brand_name}_{device}_v{variation_index}",
234
+ image_format,
235
+ 1,
236
+ device_file_types[device]
237
+ )
238
+ banner_results[device] = result['url']
239
+
240
+ logger.info(f"{device} banner uploaded successfully for variation {variation_index}")
241
+
242
+ except Exception as error:
243
+ logger.error(f"Error processing {device} banner in variation {variation_index}: {str(error)}")
244
+ raise
245
+
246
+ if not banner_results:
247
+ logger.error(f"No images were generated for variation {variation_index}")
248
+ raise Exception(f"No images were generated for variation {variation_index}")
249
+
250
+ all_banner_results.append({
251
+ "variation": variation_index,
252
+ "banners": banner_results
253
+ })
254
+
255
+ logger.info("All banner variations generated successfully.")
256
  return {"variations": all_banner_results}