wip outpainting
Browse files- app.py +194 -5
- requirements.txt +1 -0
app.py
CHANGED
@@ -4,6 +4,9 @@ import torch
|
|
4 |
from loadimg import load_img
|
5 |
from torchvision import transforms
|
6 |
from transformers import AutoModelForImageSegmentation
|
|
|
|
|
|
|
7 |
|
8 |
torch.set_float32_matmul_precision(["high", "highest"][0])
|
9 |
|
@@ -20,10 +23,186 @@ transform_image = transforms.Compose(
|
|
20 |
]
|
21 |
)
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
@spaces.GPU
|
25 |
-
def rmbg(image,url):
|
26 |
-
if image is None
|
27 |
image = url
|
28 |
image = load_img(image).convert("RGB")
|
29 |
image_size = image.size
|
@@ -38,11 +217,21 @@ def rmbg(image,url):
|
|
38 |
return image
|
39 |
|
40 |
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
demo = gr.TabbedInterface(
|
44 |
-
[rmbg_tab],
|
45 |
-
["remove background"],
|
46 |
title="Utilities that require GPU",
|
47 |
)
|
48 |
|
|
|
4 |
from loadimg import load_img
|
5 |
from torchvision import transforms
|
6 |
from transformers import AutoModelForImageSegmentation
|
7 |
+
from diffusers import FluxFillPipeline
|
8 |
+
from PIL import Image, ImageDraw
|
9 |
+
from diffusers.utils import load_image
|
10 |
|
11 |
torch.set_float32_matmul_precision(["high", "highest"][0])
|
12 |
|
|
|
23 |
]
|
24 |
)
|
25 |
|
26 |
+
pipe = FluxFillPipeline.from_pretrained(
|
27 |
+
"black-forest-labs/FLUX.1-Fill-dev", torch_dtype=torch.bfloat16
|
28 |
+
).to("cuda")
|
29 |
+
|
30 |
+
|
31 |
+
def can_expand(source_width, source_height, target_width, target_height, alignment):
|
32 |
+
if alignment in ("Left", "Right") and source_width >= target_width:
|
33 |
+
return False
|
34 |
+
if alignment in ("Top", "Bottom") and source_height >= target_height:
|
35 |
+
return False
|
36 |
+
return True
|
37 |
+
|
38 |
+
|
39 |
+
def prepare_image_and_mask(
|
40 |
+
image,
|
41 |
+
width,
|
42 |
+
height,
|
43 |
+
overlap_percentage,
|
44 |
+
resize_percentage,
|
45 |
+
alignment,
|
46 |
+
overlap_left,
|
47 |
+
overlap_right,
|
48 |
+
overlap_top,
|
49 |
+
overlap_bottom,
|
50 |
+
):
|
51 |
+
target_size = (width, height)
|
52 |
+
|
53 |
+
scale_factor = min(target_size[0] / image.width, target_size[1] / image.height)
|
54 |
+
new_width = int(image.width * scale_factor)
|
55 |
+
new_height = int(image.height * scale_factor)
|
56 |
+
|
57 |
+
source = image.resize((new_width, new_height), Image.LANCZOS)
|
58 |
+
|
59 |
+
resize_percentage = 50
|
60 |
+
|
61 |
+
# Calculate new dimensions based on percentage
|
62 |
+
resize_factor = resize_percentage / 100
|
63 |
+
new_width = int(source.width * resize_factor)
|
64 |
+
new_height = int(source.height * resize_factor)
|
65 |
+
|
66 |
+
# Ensure minimum size of 64 pixels
|
67 |
+
new_width = max(new_width, 64)
|
68 |
+
new_height = max(new_height, 64)
|
69 |
+
|
70 |
+
# Resize the image
|
71 |
+
source = source.resize((new_width, new_height), Image.LANCZOS)
|
72 |
+
|
73 |
+
# Calculate the overlap in pixels based on the percentage
|
74 |
+
overlap_x = int(new_width * (overlap_percentage / 100))
|
75 |
+
overlap_y = int(new_height * (overlap_percentage / 100))
|
76 |
+
|
77 |
+
# Ensure minimum overlap of 1 pixel
|
78 |
+
overlap_x = max(overlap_x, 1)
|
79 |
+
overlap_y = max(overlap_y, 1)
|
80 |
+
|
81 |
+
# Calculate margins based on alignment
|
82 |
+
if alignment == "Middle":
|
83 |
+
margin_x = (target_size[0] - new_width) // 2
|
84 |
+
margin_y = (target_size[1] - new_height) // 2
|
85 |
+
elif alignment == "Left":
|
86 |
+
margin_x = 0
|
87 |
+
margin_y = (target_size[1] - new_height) // 2
|
88 |
+
elif alignment == "Right":
|
89 |
+
margin_x = target_size[0] - new_width
|
90 |
+
margin_y = (target_size[1] - new_height) // 2
|
91 |
+
elif alignment == "Top":
|
92 |
+
margin_x = (target_size[0] - new_width) // 2
|
93 |
+
margin_y = 0
|
94 |
+
elif alignment == "Bottom":
|
95 |
+
margin_x = (target_size[0] - new_width) // 2
|
96 |
+
margin_y = target_size[1] - new_height
|
97 |
+
|
98 |
+
# Adjust margins to eliminate gaps
|
99 |
+
margin_x = max(0, min(margin_x, target_size[0] - new_width))
|
100 |
+
margin_y = max(0, min(margin_y, target_size[1] - new_height))
|
101 |
+
|
102 |
+
# Create a new background image and paste the resized source image
|
103 |
+
background = Image.new("RGB", target_size, (255, 255, 255))
|
104 |
+
background.paste(source, (margin_x, margin_y))
|
105 |
+
|
106 |
+
# Create the mask
|
107 |
+
mask = Image.new("L", target_size, 255)
|
108 |
+
mask_draw = ImageDraw.Draw(mask)
|
109 |
+
|
110 |
+
# Calculate overlap areas
|
111 |
+
white_gaps_patch = 2
|
112 |
+
|
113 |
+
left_overlap = margin_x + overlap_x if overlap_left else margin_x + white_gaps_patch
|
114 |
+
right_overlap = (
|
115 |
+
margin_x + new_width - overlap_x
|
116 |
+
if overlap_right
|
117 |
+
else margin_x + new_width - white_gaps_patch
|
118 |
+
)
|
119 |
+
top_overlap = margin_y + overlap_y if overlap_top else margin_y + white_gaps_patch
|
120 |
+
bottom_overlap = (
|
121 |
+
margin_y + new_height - overlap_y
|
122 |
+
if overlap_bottom
|
123 |
+
else margin_y + new_height - white_gaps_patch
|
124 |
+
)
|
125 |
+
|
126 |
+
if alignment == "Left":
|
127 |
+
left_overlap = margin_x + overlap_x if overlap_left else margin_x
|
128 |
+
elif alignment == "Right":
|
129 |
+
right_overlap = (
|
130 |
+
margin_x + new_width - overlap_x if overlap_right else margin_x + new_width
|
131 |
+
)
|
132 |
+
elif alignment == "Top":
|
133 |
+
top_overlap = margin_y + overlap_y if overlap_top else margin_y
|
134 |
+
elif alignment == "Bottom":
|
135 |
+
bottom_overlap = (
|
136 |
+
margin_y + new_height - overlap_y
|
137 |
+
if overlap_bottom
|
138 |
+
else margin_y + new_height
|
139 |
+
)
|
140 |
+
|
141 |
+
# Draw the mask
|
142 |
+
mask_draw.rectangle(
|
143 |
+
[(left_overlap, top_overlap), (right_overlap, bottom_overlap)], fill=0
|
144 |
+
)
|
145 |
+
|
146 |
+
return background, mask
|
147 |
+
|
148 |
+
|
149 |
+
def inpaint(
|
150 |
+
image,
|
151 |
+
width,
|
152 |
+
height,
|
153 |
+
overlap_percentage,
|
154 |
+
num_inference_steps,
|
155 |
+
custom_resize_percentage,
|
156 |
+
prompt_input,
|
157 |
+
alignment,
|
158 |
+
overlap_left,
|
159 |
+
overlap_right,
|
160 |
+
overlap_top,
|
161 |
+
overlap_bottom,
|
162 |
+
progress=gr.Progress(track_tqdm=True),
|
163 |
+
):
|
164 |
+
background, mask = prepare_image_and_mask(
|
165 |
+
image,
|
166 |
+
width,
|
167 |
+
height,
|
168 |
+
overlap_percentage,
|
169 |
+
custom_resize_percentage,
|
170 |
+
alignment,
|
171 |
+
overlap_left,
|
172 |
+
overlap_right,
|
173 |
+
overlap_top,
|
174 |
+
overlap_bottom,
|
175 |
+
)
|
176 |
+
|
177 |
+
if not can_expand(background.width, background.height, width, height, alignment):
|
178 |
+
alignment = "Middle"
|
179 |
+
|
180 |
+
cnet_image = background.copy()
|
181 |
+
cnet_image.paste(0, (0, 0), mask)
|
182 |
+
|
183 |
+
final_prompt = prompt_input
|
184 |
+
|
185 |
+
# generator = torch.Generator(device="cuda").manual_seed(42)
|
186 |
+
|
187 |
+
result = pipe(
|
188 |
+
prompt=final_prompt,
|
189 |
+
height=height,
|
190 |
+
width=width,
|
191 |
+
image=cnet_image,
|
192 |
+
mask_image=mask,
|
193 |
+
num_inference_steps=num_inference_steps,
|
194 |
+
guidance_scale=30,
|
195 |
+
).images[0]
|
196 |
+
|
197 |
+
result = result.convert("RGBA")
|
198 |
+
cnet_image.paste(result, (0, 0), mask)
|
199 |
+
|
200 |
+
return cnet_image
|
201 |
+
|
202 |
|
203 |
@spaces.GPU
|
204 |
+
def rmbg(image, url):
|
205 |
+
if image is None:
|
206 |
image = url
|
207 |
image = load_img(image).convert("RGB")
|
208 |
image_size = image.size
|
|
|
217 |
return image
|
218 |
|
219 |
|
220 |
+
def placeholder(img):
|
221 |
+
return img
|
222 |
+
|
223 |
+
|
224 |
+
rmbg_tab = gr.Interface(
|
225 |
+
fn=rmbg, inputs=["image", "text"], outputs=["image"], api_name="rmbg"
|
226 |
+
)
|
227 |
+
|
228 |
+
outpaint_tab = gr.Interface(
|
229 |
+
fr=placeholder, inputs=["image"], outputs=["image"], api_name="outpainting"
|
230 |
+
)
|
231 |
|
232 |
demo = gr.TabbedInterface(
|
233 |
+
[rmbg_tab, outpaint_tab],
|
234 |
+
["remove background", "outpainting"],
|
235 |
title="Utilities that require GPU",
|
236 |
)
|
237 |
|
requirements.txt
CHANGED
@@ -11,3 +11,4 @@ scikit-image
|
|
11 |
kornia
|
12 |
transformers
|
13 |
huggingface_hub
|
|
|
|
11 |
kornia
|
12 |
transformers
|
13 |
huggingface_hub
|
14 |
+
diffusers
|