Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -2,10 +2,11 @@ import gradio as gr
|
|
2 |
import torch
|
3 |
import numpy as np
|
4 |
import cv2
|
5 |
-
from PIL import Image
|
6 |
from transformers import SamModel, SamProcessor
|
7 |
from diffusers import StableDiffusionInpaintPipeline
|
8 |
-
import
|
|
|
9 |
|
10 |
# Set up device
|
11 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
@@ -23,35 +24,27 @@ inpaint_model = StableDiffusionInpaintPipeline.from_pretrained(
|
|
23 |
torch_dtype=torch.float16 if device == "cuda" else torch.float32
|
24 |
).to(device)
|
25 |
|
26 |
-
def
|
27 |
-
"""Get
|
28 |
-
|
29 |
-
|
30 |
-
|
|
|
|
|
|
|
|
|
|
|
31 |
else:
|
32 |
-
|
33 |
|
34 |
-
|
35 |
-
|
36 |
-
# If no points provided, use grid sampling to identify important areas
|
37 |
-
if points is None:
|
38 |
-
# Create a grid of points to sample the image
|
39 |
-
x_points = np.linspace(w//4, 3*w//4, 5, dtype=int)
|
40 |
-
y_points = np.linspace(h//4, 3*h//4, 5, dtype=int)
|
41 |
-
grid_points = []
|
42 |
-
for y in y_points:
|
43 |
-
for x in x_points:
|
44 |
-
grid_points.append([x, y])
|
45 |
-
points = [grid_points]
|
46 |
-
|
47 |
-
# Process image through SAM
|
48 |
inputs = sam_processor(
|
49 |
-
images=
|
50 |
input_points=points,
|
51 |
return_tensors="pt"
|
52 |
).to(device)
|
53 |
|
54 |
-
# Generate
|
55 |
with torch.no_grad():
|
56 |
outputs = sam_model(**inputs)
|
57 |
masks = sam_processor.image_processor.post_process_masks(
|
@@ -60,123 +53,86 @@ def get_importance_map(image, points=None):
|
|
60 |
inputs["reshaped_input_sizes"].cpu()
|
61 |
)
|
62 |
|
63 |
-
#
|
64 |
-
|
65 |
-
|
66 |
-
importance_map += masks[0][i].numpy().astype(np.float32)
|
67 |
-
|
68 |
-
# Normalize to 0-1
|
69 |
-
if importance_map.max() > 0:
|
70 |
-
importance_map = importance_map / importance_map.max()
|
71 |
-
|
72 |
-
return importance_map
|
73 |
|
74 |
-
def
|
75 |
-
"""Find the optimal placement for the original image within the new canvas based on importance"""
|
76 |
-
oh, ow = original_size
|
77 |
-
nh, nw = new_size
|
78 |
-
|
79 |
-
# If the new size is smaller in any dimension, then just center it
|
80 |
-
if nh <= oh or nw <= ow:
|
81 |
-
x_offset = max(0, (nw - ow) // 2)
|
82 |
-
y_offset = max(0, (nh - oh) // 2)
|
83 |
-
return x_offset, y_offset
|
84 |
-
|
85 |
-
# Calculate all possible positions
|
86 |
-
possible_x = nw - ow + 1
|
87 |
-
possible_y = nh - oh + 1
|
88 |
-
|
89 |
-
best_score = -np.inf
|
90 |
-
best_x = 0
|
91 |
-
best_y = 0
|
92 |
-
|
93 |
-
# Create a border-weighted importance map (gives extra weight to content near borders)
|
94 |
-
y_coords, x_coords = np.ogrid[:oh, :ow]
|
95 |
-
border_weight = np.minimum(np.minimum(x_coords, ow-1-x_coords), np.minimum(y_coords, oh-1-y_coords))
|
96 |
-
border_weight = 1.0 - border_weight / border_weight.max()
|
97 |
-
weighted_importance = importance_map * (1.0 + 0.5 * border_weight)
|
98 |
-
|
99 |
-
# Optimize for 9 positions (corners, center of edges, and center)
|
100 |
-
positions = [
|
101 |
-
(0, 0), # Top-left
|
102 |
-
(0, (possible_y-1)//2), # Middle-left
|
103 |
-
(0, possible_y-1), # Bottom-left
|
104 |
-
((possible_x-1)//2, 0), # Top-center
|
105 |
-
((possible_x-1)//2, (possible_y-1)//2), # Center
|
106 |
-
((possible_x-1)//2, possible_y-1), # Bottom-center
|
107 |
-
(possible_x-1, 0), # Top-right
|
108 |
-
(possible_x-1, (possible_y-1)//2), # Middle-right
|
109 |
-
(possible_x-1, possible_y-1) # Bottom-right
|
110 |
-
]
|
111 |
-
|
112 |
-
# Find position with highest importance score
|
113 |
-
for x, y in positions:
|
114 |
-
# Calculate importance score for this position
|
115 |
-
score = weighted_importance.sum()
|
116 |
-
if score > best_score:
|
117 |
-
best_score = score
|
118 |
-
best_x = x
|
119 |
-
best_y = y
|
120 |
-
|
121 |
-
return best_x, best_y
|
122 |
-
|
123 |
-
def adjust_aspect_ratio(image, target_ratio, prompt=""):
|
124 |
"""Adjust image to target aspect ratio while preserving important content"""
|
125 |
# Convert PIL to numpy if needed
|
126 |
if isinstance(image, Image.Image):
|
127 |
-
image_pil = image
|
128 |
image_np = np.array(image)
|
129 |
else:
|
130 |
image_np = image
|
131 |
-
image_pil = Image.fromarray(image_np)
|
132 |
|
133 |
-
# Get dimensions
|
134 |
h, w = image_np.shape[:2]
|
135 |
current_ratio = w / h
|
136 |
target_ratio_value = eval(target_ratio.replace(':', '/'))
|
137 |
|
138 |
-
#
|
139 |
-
importance_map = get_importance_map(image_np)
|
140 |
-
|
141 |
-
# Calculate new dimensions
|
142 |
if current_ratio < target_ratio_value:
|
143 |
# Need to add width (outpaint left/right)
|
144 |
new_width = int(h * target_ratio_value)
|
145 |
new_height = h
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
146 |
else:
|
147 |
# Need to add height (outpaint top/bottom)
|
148 |
new_width = w
|
149 |
new_height = int(w / target_ratio_value)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
150 |
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
# Place original image at calculated position
|
159 |
-
result[y_offset:y_offset+h, x_offset:x_offset+w] = image_np
|
160 |
-
mask[y_offset:y_offset+h, x_offset:x_offset+w] = 0
|
161 |
-
|
162 |
-
# Convert to PIL for inpainting
|
163 |
-
result_pil = Image.fromarray(result)
|
164 |
mask_pil = Image.fromarray(mask)
|
165 |
|
166 |
-
#
|
167 |
if not prompt or prompt.strip() == "":
|
168 |
-
|
169 |
-
prompt = "seamless extension of the image, same style and content"
|
170 |
-
else:
|
171 |
-
prompt = "seamless extension of the image, same style, same scene, consistent lighting"
|
172 |
|
173 |
-
#
|
174 |
output = inpaint_model(
|
175 |
prompt=prompt,
|
176 |
-
image=
|
177 |
mask_image=mask_pil,
|
178 |
guidance_scale=7.5,
|
179 |
-
num_inference_steps=
|
180 |
).images[0]
|
181 |
|
182 |
return np.array(output)
|
@@ -184,7 +140,7 @@ def adjust_aspect_ratio(image, target_ratio, prompt=""):
|
|
184 |
def process_image(input_image, target_ratio="16:9", prompt=""):
|
185 |
"""Main processing function for the Gradio interface"""
|
186 |
try:
|
187 |
-
# Convert from Gradio format
|
188 |
if isinstance(input_image, dict) and 'image' in input_image:
|
189 |
image = input_image['image']
|
190 |
else:
|
@@ -196,8 +152,11 @@ def process_image(input_image, target_ratio="16:9", prompt=""):
|
|
196 |
else:
|
197 |
image_np = image
|
198 |
|
|
|
|
|
|
|
199 |
# Adjust aspect ratio while preserving content
|
200 |
-
result = adjust_aspect_ratio(image_np, target_ratio, prompt)
|
201 |
|
202 |
# Convert result to PIL for visualization
|
203 |
result_pil = Image.fromarray(result)
|
@@ -209,9 +168,9 @@ def process_image(input_image, target_ratio="16:9", prompt=""):
|
|
209 |
return None
|
210 |
|
211 |
# Create the Gradio interface
|
212 |
-
with gr.Blocks(title="
|
213 |
-
gr.Markdown("#
|
214 |
-
gr.Markdown("Upload an image, choose your target aspect ratio, and the AI
|
215 |
|
216 |
with gr.Row():
|
217 |
with gr.Column():
|
@@ -219,7 +178,7 @@ with gr.Blocks(title="Smart Aspect Ratio Adjuster") as demo:
|
|
219 |
|
220 |
with gr.Row():
|
221 |
aspect_ratio = gr.Dropdown(
|
222 |
-
choices=["16:9", "4:3", "1:1", "9:16", "3:4"
|
223 |
value="16:9",
|
224 |
label="Target Aspect Ratio"
|
225 |
)
|
@@ -242,9 +201,9 @@ with gr.Blocks(title="Smart Aspect Ratio Adjuster") as demo:
|
|
242 |
|
243 |
gr.Markdown("""
|
244 |
## How it works
|
245 |
-
1.
|
246 |
-
2.
|
247 |
-
3.
|
248 |
|
249 |
## Tips
|
250 |
- For best results, provide a descriptive prompt that matches the scene
|
|
|
2 |
import torch
|
3 |
import numpy as np
|
4 |
import cv2
|
5 |
+
from PIL import Image
|
6 |
from transformers import SamModel, SamProcessor
|
7 |
from diffusers import StableDiffusionInpaintPipeline
|
8 |
+
import requests
|
9 |
+
from io import BytesIO
|
10 |
|
11 |
# Set up device
|
12 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
24 |
torch_dtype=torch.float16 if device == "cuda" else torch.float32
|
25 |
).to(device)
|
26 |
|
27 |
+
def get_sam_mask(image, points=None):
|
28 |
+
"""Get segmentation mask using SAM model"""
|
29 |
+
if points is None:
|
30 |
+
# If no points provided, use center point
|
31 |
+
height, width = image.shape[:2]
|
32 |
+
points = [[[width // 2, height // 2]]]
|
33 |
+
|
34 |
+
# Convert to PIL if needed
|
35 |
+
if not isinstance(image, Image.Image):
|
36 |
+
image_pil = Image.fromarray(image)
|
37 |
else:
|
38 |
+
image_pil = image
|
39 |
|
40 |
+
# Process the image and point prompts
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
inputs = sam_processor(
|
42 |
+
images=image_pil,
|
43 |
input_points=points,
|
44 |
return_tensors="pt"
|
45 |
).to(device)
|
46 |
|
47 |
+
# Generate mask
|
48 |
with torch.no_grad():
|
49 |
outputs = sam_model(**inputs)
|
50 |
masks = sam_processor.image_processor.post_process_masks(
|
|
|
53 |
inputs["reshaped_input_sizes"].cpu()
|
54 |
)
|
55 |
|
56 |
+
# Get the mask
|
57 |
+
mask = masks[0][0].numpy()
|
58 |
+
return mask
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
+
def adjust_aspect_ratio(image, mask, target_ratio, prompt=""):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
"""Adjust image to target aspect ratio while preserving important content"""
|
62 |
# Convert PIL to numpy if needed
|
63 |
if isinstance(image, Image.Image):
|
|
|
64 |
image_np = np.array(image)
|
65 |
else:
|
66 |
image_np = image
|
|
|
67 |
|
|
|
68 |
h, w = image_np.shape[:2]
|
69 |
current_ratio = w / h
|
70 |
target_ratio_value = eval(target_ratio.replace(':', '/'))
|
71 |
|
72 |
+
# Determine if we need to add width or height
|
|
|
|
|
|
|
73 |
if current_ratio < target_ratio_value:
|
74 |
# Need to add width (outpaint left/right)
|
75 |
new_width = int(h * target_ratio_value)
|
76 |
new_height = h
|
77 |
+
|
78 |
+
# Calculate padding
|
79 |
+
pad_width = new_width - w
|
80 |
+
pad_left = pad_width // 2
|
81 |
+
pad_right = pad_width - pad_left
|
82 |
+
|
83 |
+
# Create canvas with padding
|
84 |
+
result = np.zeros((new_height, new_width, 3), dtype=np.uint8)
|
85 |
+
# Place original image in the center
|
86 |
+
result[:, pad_left:pad_left+w, :] = image_np
|
87 |
+
|
88 |
+
# Create mask for inpainting
|
89 |
+
inpaint_mask = np.ones((new_height, new_width), dtype=np.uint8) * 255
|
90 |
+
inpaint_mask[:, pad_left:pad_left+w] = 0
|
91 |
+
|
92 |
+
# Perform outpainting using Stable Diffusion
|
93 |
+
result = outpaint_regions(result, inpaint_mask, prompt)
|
94 |
+
|
95 |
else:
|
96 |
# Need to add height (outpaint top/bottom)
|
97 |
new_width = w
|
98 |
new_height = int(w / target_ratio_value)
|
99 |
+
|
100 |
+
# Calculate padding
|
101 |
+
pad_height = new_height - h
|
102 |
+
pad_top = pad_height // 2
|
103 |
+
pad_bottom = pad_height - pad_top
|
104 |
+
|
105 |
+
# Create canvas with padding
|
106 |
+
result = np.zeros((new_height, new_width, 3), dtype=np.uint8)
|
107 |
+
# Place original image in the center
|
108 |
+
result[pad_top:pad_top+h, :, :] = image_np
|
109 |
+
|
110 |
+
# Create mask for inpainting
|
111 |
+
inpaint_mask = np.ones((new_height, new_width), dtype=np.uint8) * 255
|
112 |
+
inpaint_mask[pad_top:pad_top+h, :] = 0
|
113 |
+
|
114 |
+
# Perform outpainting using Stable Diffusion
|
115 |
+
result = outpaint_regions(result, inpaint_mask, prompt)
|
116 |
|
117 |
+
return result
|
118 |
+
|
119 |
+
def outpaint_regions(image, mask, prompt):
|
120 |
+
"""Use Stable Diffusion to outpaint masked regions"""
|
121 |
+
# Convert to PIL images
|
122 |
+
image_pil = Image.fromarray(image)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
123 |
mask_pil = Image.fromarray(mask)
|
124 |
|
125 |
+
# If prompt is empty, use a generic one
|
126 |
if not prompt or prompt.strip() == "":
|
127 |
+
prompt = "seamless extension of the image, same style, same scene"
|
|
|
|
|
|
|
128 |
|
129 |
+
# Generate the outpainting
|
130 |
output = inpaint_model(
|
131 |
prompt=prompt,
|
132 |
+
image=image_pil,
|
133 |
mask_image=mask_pil,
|
134 |
guidance_scale=7.5,
|
135 |
+
num_inference_steps=25
|
136 |
).images[0]
|
137 |
|
138 |
return np.array(output)
|
|
|
140 |
def process_image(input_image, target_ratio="16:9", prompt=""):
|
141 |
"""Main processing function for the Gradio interface"""
|
142 |
try:
|
143 |
+
# Convert from Gradio format
|
144 |
if isinstance(input_image, dict) and 'image' in input_image:
|
145 |
image = input_image['image']
|
146 |
else:
|
|
|
152 |
else:
|
153 |
image_np = image
|
154 |
|
155 |
+
# Get SAM mask to identify important regions
|
156 |
+
mask = get_sam_mask(image_np)
|
157 |
+
|
158 |
# Adjust aspect ratio while preserving content
|
159 |
+
result = adjust_aspect_ratio(image_np, mask, target_ratio, prompt)
|
160 |
|
161 |
# Convert result to PIL for visualization
|
162 |
result_pil = Image.fromarray(result)
|
|
|
168 |
return None
|
169 |
|
170 |
# Create the Gradio interface
|
171 |
+
with gr.Blocks(title="Automatic Aspect Ratio Adjuster") as demo:
|
172 |
+
gr.Markdown("# Automatic Aspect Ratio Adjuster")
|
173 |
+
gr.Markdown("Upload an image, choose your target aspect ratio, and let the AI adjust it while preserving important content.")
|
174 |
|
175 |
with gr.Row():
|
176 |
with gr.Column():
|
|
|
178 |
|
179 |
with gr.Row():
|
180 |
aspect_ratio = gr.Dropdown(
|
181 |
+
choices=["16:9", "4:3", "1:1", "9:16", "3:4"],
|
182 |
value="16:9",
|
183 |
label="Target Aspect Ratio"
|
184 |
)
|
|
|
201 |
|
202 |
gr.Markdown("""
|
203 |
## How it works
|
204 |
+
1. SAM (Segment Anything Model) identifies important content in your image
|
205 |
+
2. The algorithm calculates how to adjust the aspect ratio while preserving this content
|
206 |
+
3. Stable Diffusion fills in the new areas with AI-generated content that matches the original image
|
207 |
|
208 |
## Tips
|
209 |
- For best results, provide a descriptive prompt that matches the scene
|