Spaces:
Running
on
Zero
Running
on
Zero
Commit
•
29dc892
1
Parent(s):
69b2b0d
Update app.py
Browse files
app.py
CHANGED
@@ -5,6 +5,7 @@ from controlnet_flux import FluxControlNetModel
|
|
5 |
from transformer_flux import FluxTransformer2DModel
|
6 |
from pipeline_flux_controlnet_inpaint import FluxControlNetInpaintingPipeline
|
7 |
from PIL import Image, ImageDraw
|
|
|
8 |
import spaces
|
9 |
|
10 |
# Load models
|
@@ -21,78 +22,370 @@ pipe = FluxControlNetInpaintingPipeline.from_pretrained(
|
|
21 |
pipe.transformer.to(torch.bfloat16)
|
22 |
pipe.controlnet.to(torch.bfloat16)
|
23 |
|
24 |
-
def
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
|
39 |
-
#
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
], fill=0)
|
48 |
-
|
49 |
return background, mask
|
50 |
|
51 |
@spaces.GPU
|
52 |
-
def inpaint(image,
|
53 |
-
|
54 |
-
image, mask = prepare_image_and_mask(image, width, height, overlap_percentage)
|
55 |
|
56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
generator = torch.Generator(device="cuda").manual_seed(42)
|
58 |
-
|
59 |
-
# Run inpainting
|
60 |
result = pipe(
|
61 |
-
prompt=
|
62 |
height=height,
|
63 |
width=width,
|
64 |
-
control_image=
|
65 |
control_mask=mask,
|
66 |
num_inference_steps=num_inference_steps,
|
67 |
generator=generator,
|
68 |
controlnet_conditioning_scale=0.9,
|
69 |
-
guidance_scale=
|
70 |
negative_prompt="",
|
71 |
-
true_guidance_scale=
|
72 |
).images[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
-
return
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
|
92 |
run_button.click(
|
|
|
|
|
|
|
|
|
93 |
fn=inpaint,
|
94 |
-
inputs=[input_image,
|
95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
)
|
97 |
|
98 |
-
demo.launch()
|
|
|
5 |
from transformer_flux import FluxTransformer2DModel
|
6 |
from pipeline_flux_controlnet_inpaint import FluxControlNetInpaintingPipeline
|
7 |
from PIL import Image, ImageDraw
|
8 |
+
import numpy as np
|
9 |
import spaces
|
10 |
|
11 |
# Load models
|
|
|
22 |
pipe.transformer.to(torch.bfloat16)
|
23 |
pipe.controlnet.to(torch.bfloat16)
|
24 |
|
25 |
+
def can_expand(source_width, source_height, target_width, target_height, alignment):
|
26 |
+
if alignment in ("Left", "Right") and source_width >= target_width:
|
27 |
+
return False
|
28 |
+
if alignment in ("Top", "Bottom") and source_height >= target_height:
|
29 |
+
return False
|
30 |
+
return True
|
31 |
+
|
32 |
+
def prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom):
|
33 |
+
target_size = (width, height)
|
34 |
+
|
35 |
+
# Calculate the scaling factor to fit the image within the target size
|
36 |
+
scale_factor = min(target_size[0] / image.width, target_size[1] / image.height)
|
37 |
+
new_width = int(image.width * scale_factor)
|
38 |
+
new_height = int(image.height * scale_factor)
|
39 |
|
40 |
+
# Resize the source image to fit within target size
|
41 |
+
source = image.resize((new_width, new_height), Image.LANCZOS)
|
42 |
+
|
43 |
+
# Apply resize option using percentages
|
44 |
+
if resize_option == "Full":
|
45 |
+
resize_percentage = 100
|
46 |
+
elif resize_option == "50%":
|
47 |
+
resize_percentage = 50
|
48 |
+
elif resize_option == "33%":
|
49 |
+
resize_percentage = 33
|
50 |
+
elif resize_option == "25%":
|
51 |
+
resize_percentage = 25
|
52 |
+
else: # Custom
|
53 |
+
resize_percentage = custom_resize_percentage
|
54 |
+
|
55 |
+
# Calculate new dimensions based on percentage
|
56 |
+
resize_factor = resize_percentage / 100
|
57 |
+
new_width = int(source.width * resize_factor)
|
58 |
+
new_height = int(source.height * resize_factor)
|
59 |
+
|
60 |
+
# Ensure minimum size of 64 pixels
|
61 |
+
new_width = max(new_width, 64)
|
62 |
+
new_height = max(new_height, 64)
|
63 |
+
|
64 |
+
# Resize the image
|
65 |
+
source = source.resize((new_width, new_height), Image.LANCZOS)
|
66 |
+
|
67 |
+
# Calculate the overlap in pixels based on the percentage
|
68 |
+
overlap_x = int(new_width * (overlap_percentage / 100))
|
69 |
+
overlap_y = int(new_height * (overlap_percentage / 100))
|
70 |
+
|
71 |
+
# Ensure minimum overlap of 1 pixel
|
72 |
+
overlap_x = max(overlap_x, 1)
|
73 |
+
overlap_y = max(overlap_y, 1)
|
74 |
+
|
75 |
+
# Calculate margins based on alignment
|
76 |
+
if alignment == "Middle":
|
77 |
+
margin_x = (target_size[0] - new_width) // 2
|
78 |
+
margin_y = (target_size[1] - new_height) // 2
|
79 |
+
elif alignment == "Left":
|
80 |
+
margin_x = 0
|
81 |
+
margin_y = (target_size[1] - new_height) // 2
|
82 |
+
elif alignment == "Right":
|
83 |
+
margin_x = target_size[0] - new_width
|
84 |
+
margin_y = (target_size[1] - new_height) // 2
|
85 |
+
elif alignment == "Top":
|
86 |
+
margin_x = (target_size[0] - new_width) // 2
|
87 |
+
margin_y = 0
|
88 |
+
elif alignment == "Bottom":
|
89 |
+
margin_x = (target_size[0] - new_width) // 2
|
90 |
+
margin_y = target_size[1] - new_height
|
91 |
+
|
92 |
+
# Adjust margins to eliminate gaps
|
93 |
+
margin_x = max(0, min(margin_x, target_size[0] - new_width))
|
94 |
+
margin_y = max(0, min(margin_y, target_size[1] - new_height))
|
95 |
+
|
96 |
+
# Create a new background image and paste the resized source image
|
97 |
+
background = Image.new('RGB', target_size, (255, 255, 255))
|
98 |
+
background.paste(source, (margin_x, margin_y))
|
99 |
+
|
100 |
+
# Create the mask
|
101 |
+
mask = Image.new('L', target_size, 255)
|
102 |
+
mask_draw = ImageDraw.Draw(mask)
|
103 |
+
|
104 |
+
# Calculate overlap areas
|
105 |
+
white_gaps_patch = 2
|
106 |
+
|
107 |
+
left_overlap = margin_x + overlap_x if overlap_left else margin_x + white_gaps_patch
|
108 |
+
right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width - white_gaps_patch
|
109 |
+
top_overlap = margin_y + overlap_y if overlap_top else margin_y + white_gaps_patch
|
110 |
+
bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height - white_gaps_patch
|
111 |
|
112 |
+
if alignment == "Left":
|
113 |
+
left_overlap = margin_x + overlap_x if overlap_left else margin_x
|
114 |
+
elif alignment == "Right":
|
115 |
+
right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width
|
116 |
+
elif alignment == "Top":
|
117 |
+
top_overlap = margin_y + overlap_y if overlap_top else margin_y
|
118 |
+
elif alignment == "Bottom":
|
119 |
+
bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height
|
120 |
+
|
121 |
+
# Draw the mask
|
122 |
+
mask_draw.rectangle([
|
123 |
+
(left_overlap, top_overlap),
|
124 |
+
(right_overlap, bottom_overlap)
|
125 |
], fill=0)
|
126 |
+
|
127 |
return background, mask
|
128 |
|
129 |
@spaces.GPU
|
130 |
+
def inpaint(image, width, height, overlap_percentage, num_inference_steps, resize_option, custom_resize_percentage, prompt_input, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom):
|
131 |
+
background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom)
|
|
|
132 |
|
133 |
+
if not can_expand(background.width, background.height, width, height, alignment):
|
134 |
+
alignment = "Middle"
|
135 |
+
|
136 |
+
cnet_image = background.copy()
|
137 |
+
cnet_image.paste(0, (0, 0), mask)
|
138 |
+
|
139 |
+
final_prompt = f"{prompt_input} , high quality, 4k"
|
140 |
+
|
141 |
generator = torch.Generator(device="cuda").manual_seed(42)
|
142 |
+
|
|
|
143 |
result = pipe(
|
144 |
+
prompt=final_prompt,
|
145 |
height=height,
|
146 |
width=width,
|
147 |
+
control_image=cnet_image,
|
148 |
control_mask=mask,
|
149 |
num_inference_steps=num_inference_steps,
|
150 |
generator=generator,
|
151 |
controlnet_conditioning_scale=0.9,
|
152 |
+
guidance_scale=3.5,
|
153 |
negative_prompt="",
|
154 |
+
true_guidance_scale=3.5
|
155 |
).images[0]
|
156 |
+
|
157 |
+
result = result.convert("RGBA")
|
158 |
+
cnet_image.paste(result, (0, 0), mask)
|
159 |
+
|
160 |
+
return background, cnet_image
|
161 |
+
|
162 |
+
def preview_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom):
|
163 |
+
background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom)
|
164 |
+
|
165 |
+
preview = background.copy().convert('RGBA')
|
166 |
+
red_overlay = Image.new('RGBA', background.size, (255, 0, 0, 64))
|
167 |
+
red_mask = Image.new('RGBA', background.size, (0, 0, 0, 0))
|
168 |
+
red_mask.paste(red_overlay, (0, 0), mask)
|
169 |
+
preview = Image.alpha_composite(preview, red_mask)
|
170 |
|
171 |
+
return preview
|
172 |
+
|
173 |
+
def clear_result():
|
174 |
+
return gr.update(value=None)
|
175 |
+
|
176 |
+
def preload_presets(target_ratio, ui_width, ui_height):
|
177 |
+
if target_ratio == "9:16":
|
178 |
+
return 720, 1280, gr.update()
|
179 |
+
elif target_ratio == "16:9":
|
180 |
+
return 1280, 720, gr.update()
|
181 |
+
elif target_ratio == "1:1":
|
182 |
+
return 1024, 1024, gr.update()
|
183 |
+
elif target_ratio == "Custom":
|
184 |
+
return ui_width, ui_height, gr.update(open=True)
|
185 |
+
|
186 |
+
def select_the_right_preset(user_width, user_height):
|
187 |
+
if user_width == 720 and user_height == 1280:
|
188 |
+
return "9:16"
|
189 |
+
elif user_width == 1280 and user_height == 720:
|
190 |
+
return "16:9"
|
191 |
+
elif user_width == 1024 and user_height == 1024:
|
192 |
+
return "1:1"
|
193 |
+
else:
|
194 |
+
return "Custom"
|
195 |
+
|
196 |
+
def toggle_custom_resize_slider(resize_option):
|
197 |
+
return gr.update(visible=(resize_option == "Custom"))
|
198 |
+
|
199 |
+
def update_history(new_image, history):
|
200 |
+
if history is None:
|
201 |
+
history = []
|
202 |
+
history.insert(0, new_image)
|
203 |
+
return history
|
204 |
+
|
205 |
+
css = """
|
206 |
+
.gradio-container {
|
207 |
+
width: 1200px !important;
|
208 |
+
}
|
209 |
+
"""
|
210 |
+
|
211 |
+
title = """<h1 align="center">FLUX Image Outpaint</h1>
|
212 |
+
<div align="center">Drop an image you would like to extend, pick your expected ratio and hit Generate.</div>
|
213 |
+
"""
|
214 |
+
|
215 |
+
with gr.Blocks(css=css) as demo:
|
216 |
+
with gr.Column():
|
217 |
+
gr.HTML(title)
|
218 |
+
|
219 |
+
with gr.Row():
|
220 |
+
with gr.Column():
|
221 |
+
input_image = gr.Image(
|
222 |
+
type="pil",
|
223 |
+
label="Input Image"
|
224 |
+
)
|
225 |
+
|
226 |
+
with gr.Row():
|
227 |
+
with gr.Column(scale=2):
|
228 |
+
prompt_input = gr.Textbox(label="Prompt (Optional)")
|
229 |
+
with gr.Column(scale=1):
|
230 |
+
run_button = gr.Button("Generate")
|
231 |
+
|
232 |
+
with gr.Row():
|
233 |
+
target_ratio = gr.Radio(
|
234 |
+
label="Expected Ratio",
|
235 |
+
choices=["9:16", "16:9", "1:1", "Custom"],
|
236 |
+
value="9:16",
|
237 |
+
scale=2
|
238 |
+
)
|
239 |
+
|
240 |
+
alignment_dropdown = gr.Dropdown(
|
241 |
+
choices=["Middle", "Left", "Right", "Top", "Bottom"],
|
242 |
+
value="Middle",
|
243 |
+
label="Alignment"
|
244 |
+
)
|
245 |
+
|
246 |
+
with gr.Accordion(label="Advanced settings", open=False) as settings_panel:
|
247 |
+
with gr.Column():
|
248 |
+
with gr.Row():
|
249 |
+
width_slider = gr.Slider(
|
250 |
+
label="Target Width",
|
251 |
+
minimum=720,
|
252 |
+
maximum=1536,
|
253 |
+
step=8,
|
254 |
+
value=720,
|
255 |
+
)
|
256 |
+
height_slider = gr.Slider(
|
257 |
+
label="Target Height",
|
258 |
+
minimum=720,
|
259 |
+
maximum=1536,
|
260 |
+
step=8,
|
261 |
+
value=1280,
|
262 |
+
)
|
263 |
+
|
264 |
+
num_inference_steps = gr.Slider(label="Steps", minimum=4, maximum=12, step=1, value=8)
|
265 |
+
with gr.Group():
|
266 |
+
overlap_percentage = gr.Slider(
|
267 |
+
label="Mask overlap (%)",
|
268 |
+
minimum=1,
|
269 |
+
maximum=50,
|
270 |
+
value=10,
|
271 |
+
step=1
|
272 |
+
)
|
273 |
+
with gr.Row():
|
274 |
+
overlap_top = gr.Checkbox(label="Overlap Top", value=True)
|
275 |
+
overlap_right = gr.Checkbox(label="Overlap Right", value=True)
|
276 |
+
with gr.Row():
|
277 |
+
overlap_left = gr.Checkbox(label="Overlap Left", value=True)
|
278 |
+
overlap_bottom = gr.Checkbox(label="Overlap Bottom", value=True)
|
279 |
+
with gr.Row():
|
280 |
+
resize_option = gr.Radio(
|
281 |
+
label="Resize input image",
|
282 |
+
choices=["Full", "50%", "33%", "25%", "Custom"],
|
283 |
+
value="Full"
|
284 |
+
)
|
285 |
+
custom_resize_percentage = gr.Slider(
|
286 |
+
label="Custom resize (%)",
|
287 |
+
minimum=1,
|
288 |
+
maximum=100,
|
289 |
+
step=1,
|
290 |
+
value=50,
|
291 |
+
visible=False
|
292 |
+
)
|
293 |
+
|
294 |
+
with gr.Column():
|
295 |
+
preview_button = gr.Button("Preview alignment and mask")
|
296 |
+
|
297 |
+
with gr.Column():
|
298 |
+
result = gr.Image(
|
299 |
+
interactive=False,
|
300 |
+
label="Generated Image",
|
301 |
+
)
|
302 |
+
use_as_input_button = gr.Button("Use as Input Image", visible=False)
|
303 |
+
|
304 |
+
history_gallery = gr.Gallery(label="History", columns=6, object_fit="contain", interactive=False)
|
305 |
+
preview_image = gr.Image(label="Preview")
|
306 |
+
|
307 |
+
def use_output_as_input(output_image):
|
308 |
+
return gr.update(value=output_image[1])
|
309 |
+
|
310 |
+
use_as_input_button.click(
|
311 |
+
fn=use_output_as_input,
|
312 |
+
inputs=[result],
|
313 |
+
outputs=[input_image]
|
314 |
+
)
|
315 |
+
|
316 |
+
target_ratio.change(
|
317 |
+
fn=preload_presets,
|
318 |
+
inputs=[target_ratio, width_slider, height_slider],
|
319 |
+
outputs=[width_slider, height_slider, settings_panel],
|
320 |
+
queue=False
|
321 |
+
)
|
322 |
+
|
323 |
+
width_slider.change(
|
324 |
+
fn=select_the_right_preset,
|
325 |
+
inputs=[width_slider, height_slider],
|
326 |
+
outputs=[target_ratio],
|
327 |
+
queue=False
|
328 |
+
)
|
329 |
+
|
330 |
+
height_slider.change(
|
331 |
+
fn=select_the_right_preset,
|
332 |
+
inputs=[width_slider, height_slider],
|
333 |
+
outputs=[target_ratio],
|
334 |
+
queue=False
|
335 |
+
)
|
336 |
+
|
337 |
+
resize_option.change(
|
338 |
+
fn=toggle_custom_resize_slider,
|
339 |
+
inputs=[resize_option],
|
340 |
+
outputs=[custom_resize_percentage],
|
341 |
+
queue=False
|
342 |
+
)
|
343 |
|
344 |
run_button.click(
|
345 |
+
fn=clear_result,
|
346 |
+
inputs=None,
|
347 |
+
outputs=result,
|
348 |
+
).then(
|
349 |
fn=inpaint,
|
350 |
+
inputs=[input_image, width_slider, height_slider, overlap_percentage, num_inference_steps,
|
351 |
+
resize_option, custom_resize_percentage, prompt_input, alignment_dropdown,
|
352 |
+
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
353 |
+
outputs=result,
|
354 |
+
).then(
|
355 |
+
fn=lambda x, history: update_history(x[1], history),
|
356 |
+
inputs=[result, history_gallery],
|
357 |
+
outputs=history_gallery,
|
358 |
+
).then(
|
359 |
+
fn=lambda: gr.update(visible=True),
|
360 |
+
inputs=None,
|
361 |
+
outputs=use_as_input_button,
|
362 |
+
)
|
363 |
+
|
364 |
+
prompt_input.submit(
|
365 |
+
fn=clear_result,
|
366 |
+
inputs=None,
|
367 |
+
outputs=result,
|
368 |
+
).then(
|
369 |
+
fn=inpaint,
|
370 |
+
inputs=[input_image, width_slider, height_slider, overlap_percentage, num_inference_steps, resize_option, custom_resize_percentage, prompt_input, alignment_dropdown,
|
371 |
+
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
372 |
+
outputs=result,
|
373 |
+
).then(
|
374 |
+
fn=lambda x, history: update_history(x[1], history),
|
375 |
+
inputs=[result, history_gallery],
|
376 |
+
outputs=history_gallery,
|
377 |
+
).then(
|
378 |
+
fn=lambda: gr.update(visible=True),
|
379 |
+
inputs=None,
|
380 |
+
outputs=use_as_input_button,
|
381 |
+
)
|
382 |
+
|
383 |
+
preview_button.click(
|
384 |
+
fn=preview_image_and_mask,
|
385 |
+
inputs=[input_image, width_slider, height_slider, overlap_percentage, resize_option, custom_resize_percentage, alignment_dropdown,
|
386 |
+
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
387 |
+
outputs=preview_image,
|
388 |
+
queue=False
|
389 |
)
|
390 |
|
391 |
+
demo.queue(max_size=12).launch(share=False)
|