Spaces:
Running
on
Zero
Running
on
Zero
Commit
·
7655b4c
1
Parent(s):
0d33a1d
update
Browse files- example.py +85 -60
example.py
CHANGED
@@ -1,7 +1,10 @@
|
|
1 |
try:
|
2 |
import spaces
|
3 |
-
except:
|
4 |
-
|
|
|
|
|
|
|
5 |
|
6 |
import os
|
7 |
import torch
|
@@ -39,66 +42,88 @@ if not os.path.exists(adapter_name_or_path):
|
|
39 |
triplane_turbo_pipeline = TriplaneTurboTextTo3DPipeline.from_pretrained(adapter_name_or_path)
|
40 |
triplane_turbo_pipeline.to(device)
|
41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
# Run the pipeline
|
43 |
-
output =
|
44 |
prompt=prompt,
|
45 |
num_results_per_prompt=num_results_per_prompt,
|
46 |
-
|
47 |
-
device=device
|
48 |
)
|
49 |
|
50 |
-
# Initialize a deque with maximum length of 100 to store obj file paths
|
51 |
-
obj_file_queue = deque(maxlen=max_obj_files)
|
52 |
-
|
53 |
-
# Save mesh
|
54 |
-
os.makedirs(output_dir, exist_ok=True)
|
55 |
-
for i, mesh in enumerate(output["mesh"]):
|
56 |
-
vertices = mesh.v_pos
|
57 |
-
|
58 |
-
# 1. First rotate -90 degrees around X-axis to make the model face up
|
59 |
-
vertices = torch.stack([
|
60 |
-
vertices[:, 0], # x remains unchanged
|
61 |
-
vertices[:, 2], # y = z
|
62 |
-
-vertices[:, 1] # z = -y
|
63 |
-
], dim=1)
|
64 |
-
|
65 |
-
# 2. Then rotate 90 degrees around Y-axis to make the model face the observer
|
66 |
-
vertices = torch.stack([
|
67 |
-
-vertices[:, 2], # x = -z
|
68 |
-
vertices[:, 1], # y remains unchanged
|
69 |
-
vertices[:, 0] # z = x
|
70 |
-
], dim=1)
|
71 |
-
|
72 |
-
mesh.v_pos = vertices
|
73 |
-
|
74 |
-
# If mesh has normals, they need to be rotated in the same way
|
75 |
-
if mesh.v_nrm is not None:
|
76 |
-
normals = mesh.v_nrm
|
77 |
-
# 1. Rotate -90 degrees around X-axis
|
78 |
-
normals = torch.stack([
|
79 |
-
normals[:, 0],
|
80 |
-
normals[:, 2],
|
81 |
-
-normals[:, 1]
|
82 |
-
], dim=1)
|
83 |
-
# 2. Rotate 90 degrees around Y-axis
|
84 |
-
normals = torch.stack([
|
85 |
-
-normals[:, 2],
|
86 |
-
normals[:, 1],
|
87 |
-
normals[:, 0]
|
88 |
-
], dim=1)
|
89 |
-
mesh._v_nrm = normals
|
90 |
-
|
91 |
-
# Save obj file and add its path to the queue
|
92 |
-
name = f"{prompt.replace(' ', '_')}_{seed}_{i}"
|
93 |
-
save_paths = export_obj(mesh, f"{output_dir}/{name}.obj")
|
94 |
-
obj_file_queue.append(save_paths[0])
|
95 |
-
|
96 |
-
# If an old file needs to be removed (queue is at max length)
|
97 |
-
# and the file exists, delete it
|
98 |
-
if len(obj_file_queue) == max_obj_files and os.path.exists(obj_file_queue[0]):
|
99 |
-
old_file = obj_file_queue[0]
|
100 |
-
try:
|
101 |
-
os.remove(old_file)
|
102 |
-
except OSError as e:
|
103 |
-
print(f"Error deleting file {old_file}: {e}")
|
104 |
-
|
|
|
1 |
try:
|
2 |
import spaces
|
3 |
+
except ImportError:
|
4 |
+
# Define a dummy decorator if spaces is not available
|
5 |
+
def GPU(func):
|
6 |
+
return func
|
7 |
+
spaces = type('spaces', (), {'GPU': GPU})
|
8 |
|
9 |
import os
|
10 |
import torch
|
|
|
42 |
triplane_turbo_pipeline = TriplaneTurboTextTo3DPipeline.from_pretrained(adapter_name_or_path)
|
43 |
triplane_turbo_pipeline.to(device)
|
44 |
|
45 |
+
@spaces.GPU
|
46 |
+
def generate_3d_model(prompt, num_results_per_prompt=1, seed=42, device="cuda"):
|
47 |
+
"""
|
48 |
+
Generate 3D models using TriplaneTurbo pipeline.
|
49 |
+
|
50 |
+
Args:
|
51 |
+
prompt (str): Text prompt for the 3D model
|
52 |
+
num_results_per_prompt (int): Number of results to generate
|
53 |
+
seed (int): Random seed for generation
|
54 |
+
device (str): Device to use for computation
|
55 |
+
|
56 |
+
Returns:
|
57 |
+
dict: Output from the pipeline
|
58 |
+
"""
|
59 |
+
output = triplane_turbo_pipeline(
|
60 |
+
prompt=prompt,
|
61 |
+
num_results_per_prompt=num_results_per_prompt,
|
62 |
+
generator=torch.Generator(device=device).manual_seed(seed),
|
63 |
+
device=device,
|
64 |
+
)
|
65 |
+
# Initialize a deque with maximum length of 100 to store obj file paths
|
66 |
+
obj_file_queue = deque(maxlen=max_obj_files)
|
67 |
+
|
68 |
+
# Save mesh
|
69 |
+
os.makedirs(output_dir, exist_ok=True)
|
70 |
+
for i, mesh in enumerate(output["mesh"]):
|
71 |
+
vertices = mesh.v_pos
|
72 |
+
|
73 |
+
# 1. First rotate -90 degrees around X-axis to make the model face up
|
74 |
+
vertices = torch.stack([
|
75 |
+
vertices[:, 0], # x remains unchanged
|
76 |
+
vertices[:, 2], # y = z
|
77 |
+
-vertices[:, 1] # z = -y
|
78 |
+
], dim=1)
|
79 |
+
|
80 |
+
# 2. Then rotate 90 degrees around Y-axis to make the model face the observer
|
81 |
+
vertices = torch.stack([
|
82 |
+
-vertices[:, 2], # x = -z
|
83 |
+
vertices[:, 1], # y remains unchanged
|
84 |
+
vertices[:, 0] # z = x
|
85 |
+
], dim=1)
|
86 |
+
|
87 |
+
mesh.v_pos = vertices
|
88 |
+
|
89 |
+
# If mesh has normals, they need to be rotated in the same way
|
90 |
+
if mesh.v_nrm is not None:
|
91 |
+
normals = mesh.v_nrm
|
92 |
+
# 1. Rotate -90 degrees around X-axis
|
93 |
+
normals = torch.stack([
|
94 |
+
normals[:, 0],
|
95 |
+
normals[:, 2],
|
96 |
+
-normals[:, 1]
|
97 |
+
], dim=1)
|
98 |
+
# 2. Rotate 90 degrees around Y-axis
|
99 |
+
normals = torch.stack([
|
100 |
+
-normals[:, 2],
|
101 |
+
normals[:, 1],
|
102 |
+
normals[:, 0]
|
103 |
+
], dim=1)
|
104 |
+
mesh._v_nrm = normals
|
105 |
+
|
106 |
+
# Save obj file and add its path to the queue
|
107 |
+
name = f"{prompt.replace(' ', '_')}_{seed}_{i}"
|
108 |
+
save_paths = export_obj(mesh, f"{output_dir}/{name}.obj")
|
109 |
+
obj_file_queue.append(save_paths[0])
|
110 |
+
|
111 |
+
# If an old file needs to be removed (queue is at max length)
|
112 |
+
# and the file exists, delete it
|
113 |
+
if len(obj_file_queue) == max_obj_files and os.path.exists(obj_file_queue[0]):
|
114 |
+
old_file = obj_file_queue[0]
|
115 |
+
try:
|
116 |
+
os.remove(old_file)
|
117 |
+
except OSError as e:
|
118 |
+
print(f"Error deleting file {old_file}: {e}")
|
119 |
+
|
120 |
+
|
121 |
+
|
122 |
# Run the pipeline
|
123 |
+
output = generate_3d_model(
|
124 |
prompt=prompt,
|
125 |
num_results_per_prompt=num_results_per_prompt,
|
126 |
+
seed=seed,
|
127 |
+
device=device
|
128 |
)
|
129 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|