Spaces:
Sleeping
Sleeping
ZeroGPU (#3)
Browse files- Update for ZeroGPU (c30a512011b66584ac56346cb222264010300aaa)
Co-authored-by: hysts <[email protected]>
app.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
from functools import partial
|
2 |
import os
|
3 |
from PIL import Image, ImageOps
|
4 |
import random
|
@@ -46,6 +45,7 @@ If you have uploaded one of your own images, it is very likely that you will nee
|
|
46 |
You should verify that the preprocessed image is object-centric (i.e., clearly contains a single object) and has white background.
|
47 |
'''
|
48 |
|
|
|
49 |
def center_and_square_image(pil_image_rgba, drags):
|
50 |
image = pil_image_rgba
|
51 |
alpha = np.array(image)[:, :, 3] # Extract the alpha channel
|
@@ -70,11 +70,13 @@ def center_and_square_image(pil_image_rgba, drags):
|
|
70 |
image = image.resize((256, 256), Image.Resampling.LANCZOS)
|
71 |
return image, new_drags
|
72 |
|
|
|
73 |
def sam_init():
|
74 |
sam_checkpoint = os.path.join(os.path.dirname(__file__), "ckpts", "sam_vit_h_4b8939.pth")
|
75 |
predictor = SamPredictor(build_sam(checkpoint=sam_checkpoint).to("cuda"))
|
76 |
return predictor
|
77 |
|
|
|
78 |
def model_init():
|
79 |
model_checkpoint = os.path.join(os.path.dirname(__file__), "ckpts", "drag-a-part-final.pt")
|
80 |
model = UNet2DDragConditionModel.from_pretrained_sd(
|
@@ -94,13 +96,24 @@ def model_init():
|
|
94 |
model.load_state_dict(torch.load(model_checkpoint, map_location="cpu")["model"])
|
95 |
return model.to("cuda")
|
96 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
@spaces.GPU(duration=10)
|
98 |
-
def sam_segment(
|
99 |
image = np.asarray(input_image)
|
100 |
-
|
101 |
|
102 |
with torch.no_grad():
|
103 |
-
masks_bbox, _, _ =
|
104 |
point_coords=foreground_points if foreground_points is not None else None,
|
105 |
point_labels=np.ones(len(foreground_points)) if foreground_points is not None else None,
|
106 |
multimask_output=True
|
@@ -114,6 +127,7 @@ def sam_segment(predictor, input_image, drags, foreground_points=None):
|
|
114 |
|
115 |
return out_image, new_drags
|
116 |
|
|
|
117 |
def get_point(img, sel_pix, evt: gr.SelectData):
|
118 |
sel_pix.append(evt.index)
|
119 |
points = []
|
@@ -136,10 +150,12 @@ def get_point(img, sel_pix, evt: gr.SelectData):
|
|
136 |
points = []
|
137 |
return img if isinstance(img, np.ndarray) else np.array(img)
|
138 |
|
|
|
139 |
def clear_drag():
|
140 |
return []
|
141 |
|
142 |
-
|
|
|
143 |
if img is None:
|
144 |
gr.Warning("No image is specified. Please specify an image before preprocessing.")
|
145 |
return None, drags
|
@@ -157,7 +173,6 @@ def preprocess_image(SAM_predictor, img, chk_group, drags):
|
|
157 |
img_np = np.array(img)
|
158 |
rgb_img = img_np[..., :3]
|
159 |
img, new_drags = sam_segment(
|
160 |
-
SAM_predictor,
|
161 |
rgb_img,
|
162 |
drags,
|
163 |
foreground_points=foreground_points,
|
@@ -173,8 +188,6 @@ def preprocess_image(SAM_predictor, img, chk_group, drags):
|
|
173 |
|
174 |
|
175 |
def single_image_sample(
|
176 |
-
model,
|
177 |
-
diffusion,
|
178 |
x_cond,
|
179 |
x_cond_clip,
|
180 |
rel,
|
@@ -183,7 +196,6 @@ def single_image_sample(
|
|
183 |
drags,
|
184 |
hidden_cls,
|
185 |
num_steps=50,
|
186 |
-
vae=None,
|
187 |
):
|
188 |
z = torch.randn(2, 4, 32, 32).to("cuda")
|
189 |
|
@@ -231,16 +243,11 @@ def single_image_sample(
|
|
231 |
|
232 |
|
233 |
@spaces.GPU(duration=20)
|
234 |
-
def generate_image(
|
235 |
if img_cond is None:
|
236 |
gr.Warning("Please preprocess the image first.")
|
237 |
return None
|
238 |
|
239 |
-
model = model.to("cuda")
|
240 |
-
vae = vae.to("cuda")
|
241 |
-
clip_model = clip_model.to("cuda")
|
242 |
-
clip_vit = clip_vit.to("cuda")
|
243 |
-
|
244 |
with torch.no_grad():
|
245 |
torch.manual_seed(seed)
|
246 |
np.random.seed(seed)
|
@@ -279,8 +286,6 @@ def generate_image(model, image_processor, vae, clip_model, clip_vit, diffusion,
|
|
279 |
break
|
280 |
|
281 |
return single_image_sample(
|
282 |
-
model.to("cuda"),
|
283 |
-
diffusion,
|
284 |
x_cond,
|
285 |
cond_clip_features,
|
286 |
rel,
|
@@ -289,22 +294,9 @@ def generate_image(model, image_processor, vae, clip_model, clip_vit, diffusion,
|
|
289 |
drags,
|
290 |
cls_embedding,
|
291 |
num_steps=50,
|
292 |
-
vae=vae,
|
293 |
)
|
294 |
|
295 |
|
296 |
-
sam_predictor = sam_init()
|
297 |
-
model = model_init()
|
298 |
-
|
299 |
-
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-ema").to('cuda')
|
300 |
-
clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to('cuda')
|
301 |
-
clip_vit = CLIPVisionModel.from_pretrained("openai/clip-vit-large-patch14").to('cuda')
|
302 |
-
image_processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
303 |
-
diffusion = create_diffusion(
|
304 |
-
timestep_respacing="",
|
305 |
-
learn_sigma=False,
|
306 |
-
)
|
307 |
-
|
308 |
with gr.Blocks(title=TITLE) as demo:
|
309 |
gr.Markdown("# " + DESCRIPTION)
|
310 |
|
@@ -378,7 +370,7 @@ with gr.Blocks(title=TITLE) as demo:
|
|
378 |
value="Preprocess Input Image",
|
379 |
)
|
380 |
preprocess_button.click(
|
381 |
-
fn=
|
382 |
inputs=[input_image, preprocess_chk_group, drags],
|
383 |
outputs=[processed_image, drags],
|
384 |
queue=True,
|
@@ -407,7 +399,7 @@ with gr.Blocks(title=TITLE) as demo:
|
|
407 |
value="Generate Image",
|
408 |
)
|
409 |
generate_button.click(
|
410 |
-
fn=
|
411 |
inputs=[processed_image, seed, cfg_scale, drags],
|
412 |
outputs=[generated_image],
|
413 |
)
|
|
|
|
|
1 |
import os
|
2 |
from PIL import Image, ImageOps
|
3 |
import random
|
|
|
45 |
You should verify that the preprocessed image is object-centric (i.e., clearly contains a single object) and has white background.
|
46 |
'''
|
47 |
|
48 |
+
|
49 |
def center_and_square_image(pil_image_rgba, drags):
|
50 |
image = pil_image_rgba
|
51 |
alpha = np.array(image)[:, :, 3] # Extract the alpha channel
|
|
|
70 |
image = image.resize((256, 256), Image.Resampling.LANCZOS)
|
71 |
return image, new_drags
|
72 |
|
73 |
+
|
74 |
def sam_init():
|
75 |
sam_checkpoint = os.path.join(os.path.dirname(__file__), "ckpts", "sam_vit_h_4b8939.pth")
|
76 |
predictor = SamPredictor(build_sam(checkpoint=sam_checkpoint).to("cuda"))
|
77 |
return predictor
|
78 |
|
79 |
+
|
80 |
def model_init():
|
81 |
model_checkpoint = os.path.join(os.path.dirname(__file__), "ckpts", "drag-a-part-final.pt")
|
82 |
model = UNet2DDragConditionModel.from_pretrained_sd(
|
|
|
96 |
model.load_state_dict(torch.load(model_checkpoint, map_location="cpu")["model"])
|
97 |
return model.to("cuda")
|
98 |
|
99 |
+
|
100 |
+
sam_predictor = sam_init()
|
101 |
+
model = model_init()
|
102 |
+
|
103 |
+
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-ema").to('cuda')
|
104 |
+
clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to('cuda')
|
105 |
+
clip_vit = CLIPVisionModel.from_pretrained("openai/clip-vit-large-patch14").to('cuda')
|
106 |
+
image_processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
107 |
+
diffusion = create_diffusion(timestep_respacing="", learn_sigma=False)
|
108 |
+
|
109 |
+
|
110 |
@spaces.GPU(duration=10)
|
111 |
+
def sam_segment(input_image, drags, foreground_points=None):
|
112 |
image = np.asarray(input_image)
|
113 |
+
sam_predictor.set_image(image)
|
114 |
|
115 |
with torch.no_grad():
|
116 |
+
masks_bbox, _, _ = sam_predictor.predict(
|
117 |
point_coords=foreground_points if foreground_points is not None else None,
|
118 |
point_labels=np.ones(len(foreground_points)) if foreground_points is not None else None,
|
119 |
multimask_output=True
|
|
|
127 |
|
128 |
return out_image, new_drags
|
129 |
|
130 |
+
|
131 |
def get_point(img, sel_pix, evt: gr.SelectData):
|
132 |
sel_pix.append(evt.index)
|
133 |
points = []
|
|
|
150 |
points = []
|
151 |
return img if isinstance(img, np.ndarray) else np.array(img)
|
152 |
|
153 |
+
|
154 |
def clear_drag():
|
155 |
return []
|
156 |
|
157 |
+
|
158 |
+
def preprocess_image(img, chk_group, drags):
|
159 |
if img is None:
|
160 |
gr.Warning("No image is specified. Please specify an image before preprocessing.")
|
161 |
return None, drags
|
|
|
173 |
img_np = np.array(img)
|
174 |
rgb_img = img_np[..., :3]
|
175 |
img, new_drags = sam_segment(
|
|
|
176 |
rgb_img,
|
177 |
drags,
|
178 |
foreground_points=foreground_points,
|
|
|
188 |
|
189 |
|
190 |
def single_image_sample(
|
|
|
|
|
191 |
x_cond,
|
192 |
x_cond_clip,
|
193 |
rel,
|
|
|
196 |
drags,
|
197 |
hidden_cls,
|
198 |
num_steps=50,
|
|
|
199 |
):
|
200 |
z = torch.randn(2, 4, 32, 32).to("cuda")
|
201 |
|
|
|
243 |
|
244 |
|
245 |
@spaces.GPU(duration=20)
|
246 |
+
def generate_image(img_cond, seed, cfg_scale, drags_list):
|
247 |
if img_cond is None:
|
248 |
gr.Warning("Please preprocess the image first.")
|
249 |
return None
|
250 |
|
|
|
|
|
|
|
|
|
|
|
251 |
with torch.no_grad():
|
252 |
torch.manual_seed(seed)
|
253 |
np.random.seed(seed)
|
|
|
286 |
break
|
287 |
|
288 |
return single_image_sample(
|
|
|
|
|
289 |
x_cond,
|
290 |
cond_clip_features,
|
291 |
rel,
|
|
|
294 |
drags,
|
295 |
cls_embedding,
|
296 |
num_steps=50,
|
|
|
297 |
)
|
298 |
|
299 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
300 |
with gr.Blocks(title=TITLE) as demo:
|
301 |
gr.Markdown("# " + DESCRIPTION)
|
302 |
|
|
|
370 |
value="Preprocess Input Image",
|
371 |
)
|
372 |
preprocess_button.click(
|
373 |
+
fn=preprocess_image,
|
374 |
inputs=[input_image, preprocess_chk_group, drags],
|
375 |
outputs=[processed_image, drags],
|
376 |
queue=True,
|
|
|
399 |
value="Generate Image",
|
400 |
)
|
401 |
generate_button.click(
|
402 |
+
fn=generate_image,
|
403 |
inputs=[processed_image, seed, cfg_scale, drags],
|
404 |
outputs=[generated_image],
|
405 |
)
|