Spaces:
Paused
Paused
Commit
Β·
2153bff
1
Parent(s):
0b42bb7
added logs
Browse files- app.py +3 -0
- models/image_generator.py +5 -6
- models/model_3d_generator.py +144 -23
app.py
CHANGED
|
@@ -7,6 +7,9 @@ import gc
|
|
| 7 |
from datetime import datetime
|
| 8 |
from pathlib import Path
|
| 9 |
|
|
|
|
|
|
|
|
|
|
| 10 |
# Initialize directories
|
| 11 |
DATA_DIR = Path("/data") if os.path.exists("/data") else Path("./data")
|
| 12 |
DATA_DIR.mkdir(exist_ok=True)
|
|
|
|
| 7 |
from datetime import datetime
|
| 8 |
from pathlib import Path
|
| 9 |
|
| 10 |
+
# Disable torch dynamo globally to avoid ConstantVariable errors
|
| 11 |
+
torch._dynamo.config.suppress_errors = True
|
| 12 |
+
|
| 13 |
# Initialize directories
|
| 14 |
DATA_DIR = Path("/data") if os.path.exists("/data") else Path("./data")
|
| 15 |
DATA_DIR.mkdir(exist_ok=True)
|
models/image_generator.py
CHANGED
|
@@ -5,6 +5,9 @@ import numpy as np
|
|
| 5 |
from typing import Optional, List, Union
|
| 6 |
import gc
|
| 7 |
|
|
|
|
|
|
|
|
|
|
| 8 |
class OmniGenImageGenerator:
|
| 9 |
"""Image generation using OmniGen2 model"""
|
| 10 |
|
|
@@ -64,12 +67,8 @@ class OmniGenImageGenerator:
|
|
| 64 |
else:
|
| 65 |
self.pipeline = self.pipeline.to(self.device)
|
| 66 |
|
| 67 |
-
#
|
| 68 |
-
|
| 69 |
-
try:
|
| 70 |
-
self.pipeline.unet = torch.compile(self.pipeline.unet, mode="reduce-overhead")
|
| 71 |
-
except:
|
| 72 |
-
pass # Compilation is optional
|
| 73 |
|
| 74 |
except Exception as e:
|
| 75 |
print(f"Failed to load image generation model: {e}")
|
|
|
|
| 5 |
from typing import Optional, List, Union
|
| 6 |
import gc
|
| 7 |
|
| 8 |
+
# Disable torch dynamo to avoid ConstantVariable errors
|
| 9 |
+
torch._dynamo.config.suppress_errors = True
|
| 10 |
+
|
| 11 |
class OmniGenImageGenerator:
|
| 12 |
"""Image generation using OmniGen2 model"""
|
| 13 |
|
|
|
|
| 67 |
else:
|
| 68 |
self.pipeline = self.pipeline.to(self.device)
|
| 69 |
|
| 70 |
+
# Disable torch.compile to avoid dynamo issues that cause ConstantVariable errors
|
| 71 |
+
print("Skipping torch.compile to avoid dynamo compatibility issues")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
|
| 73 |
except Exception as e:
|
| 74 |
print(f"Failed to load image generation model: {e}")
|
models/model_3d_generator.py
CHANGED
|
@@ -6,12 +6,21 @@ import tempfile
|
|
| 6 |
from typing import Union, Optional, Dict, Any
|
| 7 |
from pathlib import Path
|
| 8 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
class Hunyuan3DGenerator:
|
| 11 |
"""3D model generation using Hunyuan3D-2.1"""
|
| 12 |
|
| 13 |
def __init__(self, device: str = "cuda"):
|
|
|
|
|
|
|
| 14 |
self.device = device if torch.cuda.is_available() else "cpu"
|
|
|
|
|
|
|
| 15 |
self.model = None
|
| 16 |
self.preprocessor = None
|
| 17 |
|
|
@@ -25,52 +34,106 @@ class Hunyuan3DGenerator:
|
|
| 25 |
self.resolution = 256 # 3D resolution
|
| 26 |
|
| 27 |
# Use lite model for low VRAM
|
| 28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
def _check_vram(self) -> bool:
|
| 31 |
"""Check if we have enough VRAM for full model"""
|
|
|
|
|
|
|
| 32 |
if not torch.cuda.is_available():
|
|
|
|
| 33 |
return False
|
| 34 |
|
| 35 |
try:
|
| 36 |
vram = torch.cuda.get_device_properties(0).total_memory
|
|
|
|
|
|
|
|
|
|
| 37 |
# Need at least 12GB for full model
|
| 38 |
-
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
| 40 |
return False
|
| 41 |
|
| 42 |
def load_model(self):
|
| 43 |
"""Lazy load the 3D generation model"""
|
| 44 |
if self.model is None:
|
|
|
|
|
|
|
| 45 |
try:
|
| 46 |
# Import Hunyuan3D components
|
|
|
|
| 47 |
from transformers import AutoModel, AutoProcessor
|
| 48 |
|
| 49 |
model_id = self.lite_model_id if self.use_lite else self.model_id
|
|
|
|
| 50 |
|
| 51 |
# Load preprocessor
|
|
|
|
| 52 |
self.preprocessor = AutoProcessor.from_pretrained(model_id)
|
|
|
|
| 53 |
|
| 54 |
# Load model with optimizations
|
| 55 |
torch_dtype = torch.float16 if self.device == "cuda" else torch.float32
|
|
|
|
| 56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
self.model = AutoModel.from_pretrained(
|
| 58 |
model_id,
|
| 59 |
torch_dtype=torch_dtype,
|
| 60 |
low_cpu_mem_usage=True,
|
| 61 |
-
device_map=
|
| 62 |
trust_remote_code=True
|
| 63 |
)
|
|
|
|
| 64 |
|
| 65 |
-
|
| 66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
|
| 68 |
-
# Enable optimizations
|
|
|
|
| 69 |
if hasattr(self.model, 'enable_attention_slicing'):
|
| 70 |
self.model.enable_attention_slicing()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
|
| 72 |
except Exception as e:
|
| 73 |
-
|
|
|
|
|
|
|
| 74 |
# Model loading failed, will use fallback
|
| 75 |
self.model = "fallback"
|
| 76 |
|
|
@@ -80,55 +143,113 @@ class Hunyuan3DGenerator:
|
|
| 80 |
texture_resolution: int = 1024) -> Union[str, trimesh.Trimesh]:
|
| 81 |
"""Convert 2D image to 3D model"""
|
| 82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
try:
|
| 84 |
# Load model if needed
|
|
|
|
| 85 |
if self.model is None:
|
|
|
|
| 86 |
self.load_model()
|
|
|
|
|
|
|
| 87 |
|
| 88 |
# If model loading failed, use fallback
|
| 89 |
if self.model == "fallback":
|
|
|
|
| 90 |
return self._generate_fallback_3d(image)
|
| 91 |
|
| 92 |
# Prepare image
|
|
|
|
| 93 |
if isinstance(image, str):
|
|
|
|
| 94 |
image = Image.open(image)
|
| 95 |
elif isinstance(image, np.ndarray):
|
|
|
|
| 96 |
image = Image.fromarray(image)
|
|
|
|
|
|
|
| 97 |
|
| 98 |
# Ensure RGB
|
|
|
|
| 99 |
if image.mode != 'RGB':
|
|
|
|
| 100 |
image = image.convert('RGB')
|
| 101 |
|
|
|
|
|
|
|
| 102 |
# Resize for processing
|
|
|
|
| 103 |
image = image.resize((512, 512), Image.Resampling.LANCZOS)
|
|
|
|
| 104 |
|
| 105 |
# Remove background if requested
|
| 106 |
if remove_background:
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
|
| 109 |
# Process with model
|
|
|
|
| 110 |
with torch.no_grad():
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
|
| 125 |
# Save mesh
|
|
|
|
| 126 |
mesh_path = self._save_mesh(mesh)
|
|
|
|
| 127 |
|
|
|
|
| 128 |
return mesh_path
|
| 129 |
|
| 130 |
except Exception as e:
|
| 131 |
-
|
|
|
|
|
|
|
| 132 |
return self._generate_fallback_3d(image)
|
| 133 |
|
| 134 |
def _remove_background(self, image: Image.Image) -> Image.Image:
|
|
|
|
| 6 |
from typing import Union, Optional, Dict, Any
|
| 7 |
from pathlib import Path
|
| 8 |
import os
|
| 9 |
+
import logging
|
| 10 |
+
|
| 11 |
+
# Set up detailed logging for 3D generation
|
| 12 |
+
logging.basicConfig(level=logging.INFO)
|
| 13 |
+
logger = logging.getLogger(__name__)
|
| 14 |
|
| 15 |
class Hunyuan3DGenerator:
|
| 16 |
"""3D model generation using Hunyuan3D-2.1"""
|
| 17 |
|
| 18 |
def __init__(self, device: str = "cuda"):
|
| 19 |
+
logger.info(f"π§ Initializing Hunyuan3DGenerator with device: {device}")
|
| 20 |
+
|
| 21 |
self.device = device if torch.cuda.is_available() else "cpu"
|
| 22 |
+
logger.info(f"π§ Final device selection: {self.device}")
|
| 23 |
+
|
| 24 |
self.model = None
|
| 25 |
self.preprocessor = None
|
| 26 |
|
|
|
|
| 34 |
self.resolution = 256 # 3D resolution
|
| 35 |
|
| 36 |
# Use lite model for low VRAM
|
| 37 |
+
vram_check = self._check_vram()
|
| 38 |
+
self.use_lite = self.device == "cpu" or not vram_check
|
| 39 |
+
|
| 40 |
+
logger.info(f"π§ VRAM check result: {vram_check}, using lite model: {self.use_lite}")
|
| 41 |
+
logger.info(f"π§ Model ID to use: {self.lite_model_id if self.use_lite else self.model_id}")
|
| 42 |
|
| 43 |
def _check_vram(self) -> bool:
|
| 44 |
"""Check if we have enough VRAM for full model"""
|
| 45 |
+
logger.info("π Checking VRAM availability...")
|
| 46 |
+
|
| 47 |
if not torch.cuda.is_available():
|
| 48 |
+
logger.info("β CUDA not available")
|
| 49 |
return False
|
| 50 |
|
| 51 |
try:
|
| 52 |
vram = torch.cuda.get_device_properties(0).total_memory
|
| 53 |
+
vram_gb = vram / (1024 * 1024 * 1024)
|
| 54 |
+
logger.info(f"π Available VRAM: {vram_gb:.2f} GB")
|
| 55 |
+
|
| 56 |
# Need at least 12GB for full model
|
| 57 |
+
has_enough = vram > 12 * 1024 * 1024 * 1024
|
| 58 |
+
logger.info(f"π Has enough VRAM (>12GB): {has_enough}")
|
| 59 |
+
return has_enough
|
| 60 |
+
except Exception as e:
|
| 61 |
+
logger.error(f"β Error checking VRAM: {e}")
|
| 62 |
return False
|
| 63 |
|
| 64 |
def load_model(self):
|
| 65 |
"""Lazy load the 3D generation model"""
|
| 66 |
if self.model is None:
|
| 67 |
+
logger.info("π Starting 3D model loading process...")
|
| 68 |
+
|
| 69 |
try:
|
| 70 |
# Import Hunyuan3D components
|
| 71 |
+
logger.info("π¦ Importing transformers components...")
|
| 72 |
from transformers import AutoModel, AutoProcessor
|
| 73 |
|
| 74 |
model_id = self.lite_model_id if self.use_lite else self.model_id
|
| 75 |
+
logger.info(f"π¦ Loading model: {model_id}")
|
| 76 |
|
| 77 |
# Load preprocessor
|
| 78 |
+
logger.info("π¦ Loading preprocessor...")
|
| 79 |
self.preprocessor = AutoProcessor.from_pretrained(model_id)
|
| 80 |
+
logger.info("β
Preprocessor loaded successfully")
|
| 81 |
|
| 82 |
# Load model with optimizations
|
| 83 |
torch_dtype = torch.float16 if self.device == "cuda" else torch.float32
|
| 84 |
+
logger.info(f"π¦ Using torch dtype: {torch_dtype}")
|
| 85 |
|
| 86 |
+
# Disable torch.compile to avoid dynamo issues
|
| 87 |
+
logger.info("π¦ Disabling torch compile to avoid dynamo issues...")
|
| 88 |
+
torch._dynamo.config.suppress_errors = True
|
| 89 |
+
|
| 90 |
+
logger.info("π¦ Loading 3D model with safe device handling...")
|
| 91 |
self.model = AutoModel.from_pretrained(
|
| 92 |
model_id,
|
| 93 |
torch_dtype=torch_dtype,
|
| 94 |
low_cpu_mem_usage=True,
|
| 95 |
+
device_map=None, # Avoid auto device mapping to prevent meta tensor issues
|
| 96 |
trust_remote_code=True
|
| 97 |
)
|
| 98 |
+
logger.info("β
3D model loaded from pretrained")
|
| 99 |
|
| 100 |
+
# Safe device movement
|
| 101 |
+
logger.info(f"π¦ Moving model to device: {self.device}")
|
| 102 |
+
try:
|
| 103 |
+
if hasattr(self.model, 'to_empty'):
|
| 104 |
+
# Use to_empty for meta tensors
|
| 105 |
+
logger.info("π¦ Using to_empty() for safe device movement...")
|
| 106 |
+
self.model = self.model.to_empty(device=self.device, dtype=torch_dtype)
|
| 107 |
+
else:
|
| 108 |
+
# Standard device movement
|
| 109 |
+
logger.info("π¦ Using standard to() for device movement...")
|
| 110 |
+
self.model = self.model.to(self.device, dtype=torch_dtype)
|
| 111 |
+
logger.info("β
Model successfully moved to device")
|
| 112 |
+
|
| 113 |
+
except RuntimeError as device_error:
|
| 114 |
+
logger.error(f"β Device movement failed: {device_error}")
|
| 115 |
+
if "meta tensor" in str(device_error):
|
| 116 |
+
logger.info("π Attempting CPU fallback for meta tensor issue...")
|
| 117 |
+
self.device = "cpu"
|
| 118 |
+
self.model = self.model.to("cpu")
|
| 119 |
+
logger.info("β
Fallback to CPU successful")
|
| 120 |
+
else:
|
| 121 |
+
raise device_error
|
| 122 |
|
| 123 |
+
# Enable optimizations safely
|
| 124 |
+
logger.info("π¦ Applying model optimizations...")
|
| 125 |
if hasattr(self.model, 'enable_attention_slicing'):
|
| 126 |
self.model.enable_attention_slicing()
|
| 127 |
+
logger.info("β
Attention slicing enabled")
|
| 128 |
+
else:
|
| 129 |
+
logger.info("β οΈ Attention slicing not available")
|
| 130 |
+
|
| 131 |
+
logger.info("π 3D model loading completed successfully!")
|
| 132 |
|
| 133 |
except Exception as e:
|
| 134 |
+
logger.error(f"β Failed to load Hunyuan3D model: {e}")
|
| 135 |
+
logger.error(f"β Error type: {type(e).__name__}")
|
| 136 |
+
logger.info("π Falling back to simple 3D generation...")
|
| 137 |
# Model loading failed, will use fallback
|
| 138 |
self.model = "fallback"
|
| 139 |
|
|
|
|
| 143 |
texture_resolution: int = 1024) -> Union[str, trimesh.Trimesh]:
|
| 144 |
"""Convert 2D image to 3D model"""
|
| 145 |
|
| 146 |
+
logger.info("π― Starting image-to-3D conversion process...")
|
| 147 |
+
logger.info(f"π― Input type: {type(image)}")
|
| 148 |
+
logger.info(f"π― Remove background: {remove_background}")
|
| 149 |
+
logger.info(f"π― Texture resolution: {texture_resolution}")
|
| 150 |
+
|
| 151 |
try:
|
| 152 |
# Load model if needed
|
| 153 |
+
logger.info("π Checking if model needs loading...")
|
| 154 |
if self.model is None:
|
| 155 |
+
logger.info("π¦ Model not loaded, initiating loading...")
|
| 156 |
self.load_model()
|
| 157 |
+
else:
|
| 158 |
+
logger.info("β
Model already loaded")
|
| 159 |
|
| 160 |
# If model loading failed, use fallback
|
| 161 |
if self.model == "fallback":
|
| 162 |
+
logger.info("π Using fallback 3D generation...")
|
| 163 |
return self._generate_fallback_3d(image)
|
| 164 |
|
| 165 |
# Prepare image
|
| 166 |
+
logger.info("πΌοΈ Preparing input image...")
|
| 167 |
if isinstance(image, str):
|
| 168 |
+
logger.info(f"πΌοΈ Loading image from path: {image}")
|
| 169 |
image = Image.open(image)
|
| 170 |
elif isinstance(image, np.ndarray):
|
| 171 |
+
logger.info("πΌοΈ Converting numpy array to PIL Image")
|
| 172 |
image = Image.fromarray(image)
|
| 173 |
+
else:
|
| 174 |
+
logger.info("πΌοΈ Input is already PIL Image")
|
| 175 |
|
| 176 |
# Ensure RGB
|
| 177 |
+
logger.info(f"πΌοΈ Image mode: {image.mode}")
|
| 178 |
if image.mode != 'RGB':
|
| 179 |
+
logger.info("πΌοΈ Converting image to RGB mode")
|
| 180 |
image = image.convert('RGB')
|
| 181 |
|
| 182 |
+
logger.info(f"πΌοΈ Final image size: {image.size}")
|
| 183 |
+
|
| 184 |
# Resize for processing
|
| 185 |
+
logger.info("πΌοΈ Resizing image for processing (512x512)...")
|
| 186 |
image = image.resize((512, 512), Image.Resampling.LANCZOS)
|
| 187 |
+
logger.info("β
Image resized successfully")
|
| 188 |
|
| 189 |
# Remove background if requested
|
| 190 |
if remove_background:
|
| 191 |
+
logger.info("π Removing background from image...")
|
| 192 |
+
try:
|
| 193 |
+
image = self._remove_background(image)
|
| 194 |
+
logger.info("β
Background removal completed")
|
| 195 |
+
except Exception as bg_error:
|
| 196 |
+
logger.error(f"β Background removal failed: {bg_error}")
|
| 197 |
+
logger.info("π Continuing with original image...")
|
| 198 |
|
| 199 |
# Process with model
|
| 200 |
+
logger.info("π§ Starting model inference...")
|
| 201 |
with torch.no_grad():
|
| 202 |
+
try:
|
| 203 |
+
# Preprocess image
|
| 204 |
+
logger.info("π Preprocessing image for model...")
|
| 205 |
+
inputs = self.preprocessor(images=image, return_tensors="pt")
|
| 206 |
+
logger.info(f"π Input tensor shape: {inputs['pixel_values'].shape if 'pixel_values' in inputs else 'unknown'}")
|
| 207 |
+
|
| 208 |
+
# Move inputs to device safely
|
| 209 |
+
logger.info(f"π Moving inputs to device: {self.device}")
|
| 210 |
+
try:
|
| 211 |
+
# Avoid device-related dynamo issues
|
| 212 |
+
device_str = str(self.device) # Convert to string to avoid torch.device in dynamo
|
| 213 |
+
inputs = {k: v.to(device_str) for k, v in inputs.items() if hasattr(v, 'to')}
|
| 214 |
+
logger.info("β
Inputs moved to device successfully")
|
| 215 |
+
except Exception as device_error:
|
| 216 |
+
logger.error(f"β Failed to move inputs to device: {device_error}")
|
| 217 |
+
raise device_error
|
| 218 |
+
|
| 219 |
+
# Generate 3D
|
| 220 |
+
logger.info("π Starting 3D generation inference...")
|
| 221 |
+
logger.info(f"π Parameters: steps={self.num_inference_steps}, guidance={self.guidance_scale}")
|
| 222 |
+
|
| 223 |
+
outputs = self.model.generate(
|
| 224 |
+
**inputs,
|
| 225 |
+
num_inference_steps=self.num_inference_steps,
|
| 226 |
+
guidance_scale=self.guidance_scale,
|
| 227 |
+
texture_resolution=texture_resolution
|
| 228 |
+
)
|
| 229 |
+
logger.info("β
3D generation completed successfully")
|
| 230 |
+
|
| 231 |
+
# Extract mesh
|
| 232 |
+
logger.info("π§ Extracting mesh from model outputs...")
|
| 233 |
+
mesh = self._extract_mesh(outputs)
|
| 234 |
+
logger.info("β
Mesh extraction completed")
|
| 235 |
+
|
| 236 |
+
except Exception as inference_error:
|
| 237 |
+
logger.error(f"β Model inference failed: {inference_error}")
|
| 238 |
+
logger.error(f"β Inference error type: {type(inference_error).__name__}")
|
| 239 |
+
raise inference_error
|
| 240 |
|
| 241 |
# Save mesh
|
| 242 |
+
logger.info("πΎ Saving generated mesh...")
|
| 243 |
mesh_path = self._save_mesh(mesh)
|
| 244 |
+
logger.info(f"β
Mesh saved to: {mesh_path}")
|
| 245 |
|
| 246 |
+
logger.info("π 3D generation process completed successfully!")
|
| 247 |
return mesh_path
|
| 248 |
|
| 249 |
except Exception as e:
|
| 250 |
+
logger.error(f"β 3D generation error: {e}")
|
| 251 |
+
logger.error(f"β Error type: {type(e).__name__}")
|
| 252 |
+
logger.info("π Falling back to simple 3D generation...")
|
| 253 |
return self._generate_fallback_3d(image)
|
| 254 |
|
| 255 |
def _remove_background(self, image: Image.Image) -> Image.Image:
|