Spaces:
Paused
Paused
Update app.py
Browse filesThis update refactors the generation process into separate stages — geometry generation and texture generation — instead of running them together in a single step.
Benefits:
* More practical for users: They can inspect and adjust the geometry before applying textures.
* Lower ZeroGPU credit usage: Geometry can be generated once and reused for multiple texture variations without regenerating the mesh, reducing GPU time and credits consumed.
* Better user experience: The UI clearly guides users step-by-step, making the process more transparent and interactive.
This change also includes CUDA environment setup and custom rasterizer installation, ensuring compatibility and optimal performance.
app.py
CHANGED
|
@@ -2,11 +2,22 @@ import os
|
|
| 2 |
import shlex
|
| 3 |
import spaces
|
| 4 |
import subprocess
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
def install_cuda_toolkit():
|
|
|
|
| 6 |
CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_550.54.14_linux.run"
|
| 7 |
CUDA_TOOLKIT_FILE = "/tmp/%s" % os.path.basename(CUDA_TOOLKIT_URL)
|
|
|
|
|
|
|
| 8 |
subprocess.call(["wget", "-q", CUDA_TOOLKIT_URL, "-O", CUDA_TOOLKIT_FILE])
|
| 9 |
subprocess.call(["chmod", "+x", CUDA_TOOLKIT_FILE])
|
|
|
|
|
|
|
| 10 |
subprocess.call([CUDA_TOOLKIT_FILE, "--silent", "--toolkit"])
|
| 11 |
|
| 12 |
os.environ["CUDA_HOME"] = "/usr/local/cuda"
|
|
@@ -16,14 +27,21 @@ def install_cuda_toolkit():
|
|
| 16 |
"" if "LD_LIBRARY_PATH" not in os.environ else os.environ["LD_LIBRARY_PATH"],
|
| 17 |
)
|
| 18 |
os.environ["TORCH_CUDA_ARCH_LIST"] = "8.0;8.6"
|
|
|
|
|
|
|
| 19 |
install_cuda_toolkit()
|
|
|
|
|
|
|
| 20 |
os.system("pip list | grep torch")
|
| 21 |
os.system('nvcc -V')
|
| 22 |
-
|
|
|
|
| 23 |
os.system("cd /home/user/app/step1x3d_texture/differentiable_renderer/ && python setup.py install")
|
| 24 |
|
|
|
|
| 25 |
subprocess.run(shlex.split("pip install custom_rasterizer-0.1-cp310-cp310-linux_x86_64.whl"), check=True)
|
| 26 |
-
|
|
|
|
| 27 |
import uuid
|
| 28 |
import torch
|
| 29 |
import trimesh
|
|
@@ -36,7 +54,10 @@ from step1x3d_texture.pipelines.step1x_3d_texture_synthesis_pipeline import (
|
|
| 36 |
)
|
| 37 |
from step1x3d_geometry.models.pipelines.pipeline_utils import reduce_face, remove_degenerate_face
|
| 38 |
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
| 40 |
parser = argparse.ArgumentParser()
|
| 41 |
parser.add_argument(
|
| 42 |
"--geometry_model", type=str, default="Step1X-3D-Geometry-Label-1300m"
|
|
@@ -48,19 +69,37 @@ parser.add_argument("--cache_dir", type=str, default="cache")
|
|
| 48 |
args = parser.parse_args()
|
| 49 |
|
| 50 |
os.makedirs(args.cache_dir, exist_ok=True)
|
|
|
|
| 51 |
|
|
|
|
| 52 |
geometry_model = Step1X3DGeometryPipeline.from_pretrained(
|
| 53 |
"stepfun-ai/Step1X-3D", subfolder=args.geometry_model
|
| 54 |
-
).to(
|
|
|
|
| 55 |
|
|
|
|
| 56 |
texture_model = Step1X3DTexturePipeline.from_pretrained("stepfun-ai/Step1X-3D", subfolder=args.texture_model)
|
|
|
|
|
|
|
| 57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
|
| 59 |
-
@spaces.GPU(duration=
|
| 60 |
-
def
|
| 61 |
-
input_image_path, guidance_scale, inference_steps, max_facenum, symmetry, edge_type
|
| 62 |
):
|
| 63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
if "Label" in args.geometry_model:
|
| 65 |
symmetry_values = ["x", "asymmetry"]
|
| 66 |
out = geometry_model(
|
|
@@ -79,49 +118,90 @@ def generate_func(
|
|
| 79 |
max_facenum=int(max_facenum),
|
| 80 |
)
|
| 81 |
|
|
|
|
| 82 |
save_name = str(uuid.uuid4())
|
| 83 |
-
print(save_name)
|
| 84 |
geometry_save_path = f"{args.cache_dir}/{save_name}.glb"
|
| 85 |
geometry_mesh = out.mesh[0]
|
| 86 |
geometry_mesh.export(geometry_save_path)
|
| 87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
geometry_mesh = remove_degenerate_face(geometry_mesh)
|
| 89 |
geometry_mesh = reduce_face(geometry_mesh)
|
|
|
|
|
|
|
| 90 |
textured_mesh = texture_model(input_image_path, geometry_mesh)
|
|
|
|
|
|
|
|
|
|
| 91 |
textured_save_path = f"{args.cache_dir}/{save_name}-textured.glb"
|
| 92 |
textured_mesh.export(textured_save_path)
|
| 93 |
-
|
| 94 |
torch.cuda.empty_cache()
|
| 95 |
-
print("
|
| 96 |
-
|
|
|
|
|
|
|
| 97 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
|
| 99 |
with gr.Blocks(title="Step1X-3D demo") as demo:
|
| 100 |
gr.Markdown("# Step1X-3D")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
with gr.Row():
|
| 102 |
with gr.Column(scale=2):
|
| 103 |
-
input_image = gr.Image(label="Image", type="filepath")
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
label="
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
with gr.Column(scale=4):
|
| 123 |
-
textured_preview = gr.Model3D(label="Textured", height=380)
|
| 124 |
-
geometry_preview = gr.Model3D(label="
|
|
|
|
| 125 |
with gr.Column(scale=1):
|
| 126 |
gr.Examples(
|
| 127 |
examples=[
|
|
@@ -138,8 +218,11 @@ with gr.Blocks(title="Step1X-3D demo") as demo:
|
|
| 138 |
cache_examples=False,
|
| 139 |
)
|
| 140 |
|
| 141 |
-
|
| 142 |
-
|
|
|
|
|
|
|
|
|
|
| 143 |
inputs=[
|
| 144 |
input_image,
|
| 145 |
guidance_scale,
|
|
@@ -148,7 +231,22 @@ with gr.Blocks(title="Step1X-3D demo") as demo:
|
|
| 148 |
symmetry,
|
| 149 |
edge_type,
|
| 150 |
],
|
| 151 |
-
outputs=[geometry_preview,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
)
|
| 153 |
|
| 154 |
-
|
|
|
|
|
|
| 2 |
import shlex
|
| 3 |
import spaces
|
| 4 |
import subprocess
|
| 5 |
+
|
| 6 |
+
# --------------------------------------------------------------------------
|
| 7 |
+
# 1. ENVIRONMENT AND DEPENDENCY INSTALLATION
|
| 8 |
+
# This section is crucial for the Hugging Face Space to work.
|
| 9 |
+
# Installs the CUDA toolkit and compiles the necessary C++/CUDA extensions.
|
| 10 |
+
# --------------------------------------------------------------------------
|
| 11 |
def install_cuda_toolkit():
|
| 12 |
+
"""Installs the CUDA toolkit required to compile extensions."""
|
| 13 |
CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_550.54.14_linux.run"
|
| 14 |
CUDA_TOOLKIT_FILE = "/tmp/%s" % os.path.basename(CUDA_TOOLKIT_URL)
|
| 15 |
+
|
| 16 |
+
print("Downloading CUDA Toolkit...")
|
| 17 |
subprocess.call(["wget", "-q", CUDA_TOOLKIT_URL, "-O", CUDA_TOOLKIT_FILE])
|
| 18 |
subprocess.call(["chmod", "+x", CUDA_TOOLKIT_FILE])
|
| 19 |
+
|
| 20 |
+
print("Installing CUDA Toolkit...")
|
| 21 |
subprocess.call([CUDA_TOOLKIT_FILE, "--silent", "--toolkit"])
|
| 22 |
|
| 23 |
os.environ["CUDA_HOME"] = "/usr/local/cuda"
|
|
|
|
| 27 |
"" if "LD_LIBRARY_PATH" not in os.environ else os.environ["LD_LIBRARY_PATH"],
|
| 28 |
)
|
| 29 |
os.environ["TORCH_CUDA_ARCH_LIST"] = "8.0;8.6"
|
| 30 |
+
print("CUDA environment configuration completed.")
|
| 31 |
+
|
| 32 |
install_cuda_toolkit()
|
| 33 |
+
|
| 34 |
+
print("Verifying PyTorch installation and NVCC version:")
|
| 35 |
os.system("pip list | grep torch")
|
| 36 |
os.system('nvcc -V')
|
| 37 |
+
|
| 38 |
+
print("Compiling differentiable renderer extension...")
|
| 39 |
os.system("cd /home/user/app/step1x3d_texture/differentiable_renderer/ && python setup.py install")
|
| 40 |
|
| 41 |
+
print("Installing custom rasterizer...")
|
| 42 |
subprocess.run(shlex.split("pip install custom_rasterizer-0.1-cp310-cp310-linux_x86_64.whl"), check=True)
|
| 43 |
+
print("Installation and compilation completed.")
|
| 44 |
+
|
| 45 |
import uuid
|
| 46 |
import torch
|
| 47 |
import trimesh
|
|
|
|
| 54 |
)
|
| 55 |
from step1x3d_geometry.models.pipelines.pipeline_utils import reduce_face, remove_degenerate_face
|
| 56 |
|
| 57 |
+
# --------------------------------------------------------------------------
|
| 58 |
+
# 2. MODEL CONFIGURATION AND LOADING
|
| 59 |
+
# Here we define the models to be used and load them into memory.
|
| 60 |
+
# --------------------------------------------------------------------------
|
| 61 |
parser = argparse.ArgumentParser()
|
| 62 |
parser.add_argument(
|
| 63 |
"--geometry_model", type=str, default="Step1X-3D-Geometry-Label-1300m"
|
|
|
|
| 69 |
args = parser.parse_args()
|
| 70 |
|
| 71 |
os.makedirs(args.cache_dir, exist_ok=True)
|
| 72 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 73 |
|
| 74 |
+
print(f"Loading geometry model: {args.geometry_model}...")
|
| 75 |
geometry_model = Step1X3DGeometryPipeline.from_pretrained(
|
| 76 |
"stepfun-ai/Step1X-3D", subfolder=args.geometry_model
|
| 77 |
+
).to(device)
|
| 78 |
+
print("Geometry model loaded.")
|
| 79 |
|
| 80 |
+
print(f"Loading texture model: {args.texture_model}...")
|
| 81 |
texture_model = Step1X3DTexturePipeline.from_pretrained("stepfun-ai/Step1X-3D", subfolder=args.texture_model)
|
| 82 |
+
print("Texture model loaded.")
|
| 83 |
+
|
| 84 |
|
| 85 |
+
# --------------------------------------------------------------------------
|
| 86 |
+
# 3. SEPARATE GENERATION FUNCTIONS
|
| 87 |
+
# The logic is split into two functions: one for geometry and one for textures.
|
| 88 |
+
# --------------------------------------------------------------------------
|
| 89 |
|
| 90 |
+
@spaces.GPU(duration=180)
|
| 91 |
+
def generate_geometry(
|
| 92 |
+
input_image_path, guidance_scale, inference_steps, max_facenum, symmetry, edge_type, progress=gr.Progress(track_tqdm=True)
|
| 93 |
):
|
| 94 |
+
"""
|
| 95 |
+
Function that generates geometry only from the input image.
|
| 96 |
+
"""
|
| 97 |
+
if input_image_path is None:
|
| 98 |
+
raise gr.Error("Please upload an image to start.")
|
| 99 |
+
|
| 100 |
+
print("Starting geometry generation...")
|
| 101 |
+
|
| 102 |
+
# Choose the appropriate pipeline based on the model name
|
| 103 |
if "Label" in args.geometry_model:
|
| 104 |
symmetry_values = ["x", "asymmetry"]
|
| 105 |
out = geometry_model(
|
|
|
|
| 118 |
max_facenum=int(max_facenum),
|
| 119 |
)
|
| 120 |
|
| 121 |
+
# Save the result to a temporary file
|
| 122 |
save_name = str(uuid.uuid4())
|
|
|
|
| 123 |
geometry_save_path = f"{args.cache_dir}/{save_name}.glb"
|
| 124 |
geometry_mesh = out.mesh[0]
|
| 125 |
geometry_mesh.export(geometry_save_path)
|
| 126 |
+
|
| 127 |
+
torch.cuda.empty_cache()
|
| 128 |
+
print(f"Geometry saved at: {geometry_save_path}")
|
| 129 |
+
|
| 130 |
+
# Return the path for display in the viewer and to store in state
|
| 131 |
+
return geometry_save_path, geometry_save_path
|
| 132 |
+
|
| 133 |
+
@spaces.GPU(duration=120)
|
| 134 |
+
def generate_texture(input_image_path, geometry_path, progress=gr.Progress(track_tqdm=True)):
|
| 135 |
+
"""
|
| 136 |
+
Function that applies texture to an already generated geometry.
|
| 137 |
+
"""
|
| 138 |
+
if not geometry_path or not os.path.exists(geometry_path):
|
| 139 |
+
raise gr.Error("Please generate the geometry first before texturing.")
|
| 140 |
+
|
| 141 |
+
print(f"Starting texturing for mesh: {geometry_path}")
|
| 142 |
+
geometry_mesh = trimesh.load(geometry_path)
|
| 143 |
+
|
| 144 |
+
# Optional post-processing of the mesh before texturing
|
| 145 |
geometry_mesh = remove_degenerate_face(geometry_mesh)
|
| 146 |
geometry_mesh = reduce_face(geometry_mesh)
|
| 147 |
+
|
| 148 |
+
# Call the texturing pipeline
|
| 149 |
textured_mesh = texture_model(input_image_path, geometry_mesh)
|
| 150 |
+
|
| 151 |
+
# Save the final result
|
| 152 |
+
save_name = os.path.basename(geometry_path).replace(".glb", "")
|
| 153 |
textured_save_path = f"{args.cache_dir}/{save_name}-textured.glb"
|
| 154 |
textured_mesh.export(textured_save_path)
|
| 155 |
+
|
| 156 |
torch.cuda.empty_cache()
|
| 157 |
+
print(f"Textured mesh saved at: {textured_save_path}")
|
| 158 |
+
|
| 159 |
+
return textured_save_path
|
| 160 |
+
|
| 161 |
|
| 162 |
+
# --------------------------------------------------------------------------
|
| 163 |
+
# 4. GRADIO USER INTERFACE
|
| 164 |
+
# Defines the look and behavior of the web app.
|
| 165 |
+
# --------------------------------------------------------------------------
|
| 166 |
|
| 167 |
with gr.Blocks(title="Step1X-3D demo") as demo:
|
| 168 |
gr.Markdown("# Step1X-3D")
|
| 169 |
+
gr.Markdown("### Demo for generating 3D models from a single image")
|
| 170 |
+
|
| 171 |
+
# State component: stores the geometry path between steps
|
| 172 |
+
geometry_path_state = gr.State()
|
| 173 |
+
|
| 174 |
with gr.Row():
|
| 175 |
with gr.Column(scale=2):
|
| 176 |
+
input_image = gr.Image(label="Input Image", type="filepath")
|
| 177 |
+
|
| 178 |
+
with gr.Accordion(label="Generation Parameters", open=True):
|
| 179 |
+
guidance_scale = gr.Number(label="Guidance Scale", value="7.5")
|
| 180 |
+
inference_steps = gr.Slider(
|
| 181 |
+
label="Inference Steps", minimum=1, maximum=100, value=50, step=1
|
| 182 |
+
)
|
| 183 |
+
max_facenum = gr.Number(label="Max. Number of Faces", value="400000")
|
| 184 |
+
symmetry = gr.Radio(
|
| 185 |
+
choices=["symmetry", "asymmetry"],
|
| 186 |
+
label="Symmetry Type",
|
| 187 |
+
value="symmetry",
|
| 188 |
+
type="index",
|
| 189 |
+
)
|
| 190 |
+
edge_type = gr.Radio(
|
| 191 |
+
choices=["sharp", "normal", "smooth"],
|
| 192 |
+
label="Edge Type",
|
| 193 |
+
value="sharp",
|
| 194 |
+
type="value",
|
| 195 |
+
)
|
| 196 |
+
|
| 197 |
+
with gr.Row():
|
| 198 |
+
btn_geo = gr.Button("1. Generate Geometry", variant="primary")
|
| 199 |
+
btn_tex = gr.Button("2. Generate Texture", visible=False, variant="primary")
|
| 200 |
+
|
| 201 |
with gr.Column(scale=4):
|
| 202 |
+
textured_preview = gr.Model3D(label="Textured Model", height=380, clear_color=[0.0, 0.0, 0.0, 0.0])
|
| 203 |
+
geometry_preview = gr.Model3D(label="Model (geometry only)", height=380, clear_color=[0.0, 0.0, 0.0, 0.0])
|
| 204 |
+
|
| 205 |
with gr.Column(scale=1):
|
| 206 |
gr.Examples(
|
| 207 |
examples=[
|
|
|
|
| 218 |
cache_examples=False,
|
| 219 |
)
|
| 220 |
|
| 221 |
+
# --- Button logic and interface flow ---
|
| 222 |
+
|
| 223 |
+
# 1. When the user clicks "Generate Geometry"
|
| 224 |
+
btn_geo.click(
|
| 225 |
+
fn=generate_geometry,
|
| 226 |
inputs=[
|
| 227 |
input_image,
|
| 228 |
guidance_scale,
|
|
|
|
| 231 |
symmetry,
|
| 232 |
edge_type,
|
| 233 |
],
|
| 234 |
+
outputs=[geometry_preview, geometry_path_state]
|
| 235 |
+
).then(
|
| 236 |
+
# 2. When geometry is done, run this part
|
| 237 |
+
fn=lambda: {
|
| 238 |
+
btn_tex: gr.update(visible=True), # Show texture button
|
| 239 |
+
textured_preview: gr.update(value=None) # Clear previous texture preview
|
| 240 |
+
},
|
| 241 |
+
outputs=[btn_tex, textured_preview]
|
| 242 |
+
)
|
| 243 |
+
|
| 244 |
+
# 3. When the user clicks "Generate Texture"
|
| 245 |
+
btn_tex.click(
|
| 246 |
+
fn=generate_texture,
|
| 247 |
+
inputs=[input_image, geometry_path_state],
|
| 248 |
+
outputs=[textured_preview],
|
| 249 |
)
|
| 250 |
|
| 251 |
+
# Launch the app
|
| 252 |
+
demo.launch(ssr_mode=False)
|