Spaces:
Running
on
Zero
Running
on
Zero
cuda
Browse files
app.py
CHANGED
@@ -12,8 +12,21 @@ from PIL import Image, ImageDraw
|
|
12 |
import numpy as np
|
13 |
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
14 |
|
15 |
-
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
|
18 |
MODELS = {
|
19 |
"RealVisXL V5.0 Lightning": "SG161222/RealVisXL_V5.0_Lightning",
|
@@ -51,6 +64,8 @@ pipe = StableDiffusionXLFillPipeline.from_pretrained(
|
|
51 |
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
|
52 |
pipe.to("cuda")
|
53 |
print(pipe)
|
|
|
|
|
54 |
PREDICTOR = SAM2ImagePredictor.from_pretrained(SAM_MODEL, device=DEVICE)
|
55 |
|
56 |
def load_default_pipeline():
|
@@ -64,8 +79,11 @@ def load_default_pipeline():
|
|
64 |
return gr.update(value="Default pipeline loaded!")
|
65 |
|
66 |
@spaces.GPU()
|
67 |
-
def predict_masks(
|
68 |
"""Predict a single mask from the image based on selected points."""
|
|
|
|
|
|
|
69 |
if not points:
|
70 |
return image # Return the original image if no points are selected
|
71 |
|
@@ -74,29 +92,28 @@ def predict_masks(image, points):
|
|
74 |
|
75 |
# Ensure points is a list of lists with at least two elements
|
76 |
if isinstance(points, list) and all(isinstance(point, list) and len(point) >= 2 for point in points):
|
77 |
-
|
78 |
else:
|
79 |
return image # Return the original image if points structure is unexpected
|
80 |
|
81 |
-
input_labels = [1] * len(
|
82 |
|
83 |
with torch.inference_mode():
|
84 |
-
PREDICTOR.set_image(
|
85 |
masks, _, _ = PREDICTOR.predict(
|
86 |
-
point_coords=
|
87 |
)
|
88 |
|
89 |
# Prepare the overlay image
|
90 |
-
|
91 |
-
red_mask = np.zeros_like(image_np)
|
92 |
if masks and len(masks) > 0:
|
93 |
red_mask[:, :, 0] = masks[0].astype(np.uint8) * 255 # Apply the red channel
|
94 |
red_mask = PILImage.fromarray(red_mask)
|
95 |
-
original_image = PILImage.fromarray(
|
96 |
blended_image = PILImage.blend(original_image, red_mask, alpha=0.5)
|
97 |
return np.array(blended_image)
|
98 |
else:
|
99 |
-
return
|
100 |
|
101 |
def update_mask(prompts):
|
102 |
"""Update the mask based on the prompts."""
|
|
|
12 |
import numpy as np
|
13 |
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
14 |
|
15 |
+
# class SAM2PredictorSingleton:
|
16 |
+
# _instance = None
|
17 |
+
|
18 |
+
# def __new__(cls):
|
19 |
+
# if cls._instance is None:
|
20 |
+
# cls._instance = super(SAM2PredictorSingleton, cls).__new__(cls)
|
21 |
+
# cls._instance._initialize_predictor()
|
22 |
+
# return cls._instance
|
23 |
+
|
24 |
+
# def _initialize_predictor(self):
|
25 |
+
# MODEL = "facebook/sam2-hiera-large"
|
26 |
+
# DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
27 |
+
# self.predictor = SAM2ImagePredictor.from_pretrained(MODEL, device=DEVICE)
|
28 |
+
|
29 |
+
|
30 |
|
31 |
MODELS = {
|
32 |
"RealVisXL V5.0 Lightning": "SG161222/RealVisXL_V5.0_Lightning",
|
|
|
64 |
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
|
65 |
pipe.to("cuda")
|
66 |
print(pipe)
|
67 |
+
DEVICE = torch.device("cuda")
|
68 |
+
SAM_MODEL = "facebook/sam2.1-hiera-large"
|
69 |
PREDICTOR = SAM2ImagePredictor.from_pretrained(SAM_MODEL, device=DEVICE)
|
70 |
|
71 |
def load_default_pipeline():
|
|
|
79 |
return gr.update(value="Default pipeline loaded!")
|
80 |
|
81 |
@spaces.GPU()
|
82 |
+
def predict_masks(prompts):
|
83 |
"""Predict a single mask from the image based on selected points."""
|
84 |
+
image = np.array(prompts["image"]) # Convert the image to a numpy array
|
85 |
+
points = prompts["points"] # Get the points from prompts
|
86 |
+
|
87 |
if not points:
|
88 |
return image # Return the original image if no points are selected
|
89 |
|
|
|
92 |
|
93 |
# Ensure points is a list of lists with at least two elements
|
94 |
if isinstance(points, list) and all(isinstance(point, list) and len(point) >= 2 for point in points):
|
95 |
+
input_points = [[point[0], point[1]] for point in points]
|
96 |
else:
|
97 |
return image # Return the original image if points structure is unexpected
|
98 |
|
99 |
+
input_labels = [1] * len(input_points)
|
100 |
|
101 |
with torch.inference_mode():
|
102 |
+
PREDICTOR.set_image(image)
|
103 |
masks, _, _ = PREDICTOR.predict(
|
104 |
+
point_coords=input_points, point_labels=input_labels, multimask_output=False
|
105 |
)
|
106 |
|
107 |
# Prepare the overlay image
|
108 |
+
red_mask = np.zeros_like(image)
|
|
|
109 |
if masks and len(masks) > 0:
|
110 |
red_mask[:, :, 0] = masks[0].astype(np.uint8) * 255 # Apply the red channel
|
111 |
red_mask = PILImage.fromarray(red_mask)
|
112 |
+
original_image = PILImage.fromarray(image)
|
113 |
blended_image = PILImage.blend(original_image, red_mask, alpha=0.5)
|
114 |
return np.array(blended_image)
|
115 |
else:
|
116 |
+
return image
|
117 |
|
118 |
def update_mask(prompts):
|
119 |
"""Update the mask based on the prompts."""
|