Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -16,947 +16,204 @@ from flask_cors import CORS
|
|
16 |
import numpy as np
|
17 |
import trimesh
|
18 |
from transformers import pipeline
|
19 |
-
from
|
20 |
-
|
21 |
-
import
|
22 |
|
23 |
app = Flask(__name__)
|
24 |
-
CORS(app)
|
25 |
|
26 |
-
#
|
27 |
UPLOAD_FOLDER = '/tmp/uploads'
|
28 |
RESULTS_FOLDER = '/tmp/results'
|
29 |
CACHE_DIR = '/tmp/huggingface'
|
30 |
ALLOWED_EXTENSIONS = {'png', 'jpg', 'jpeg'}
|
|
|
31 |
|
32 |
-
# Create necessary directories
|
33 |
os.makedirs(UPLOAD_FOLDER, exist_ok=True)
|
34 |
os.makedirs(RESULTS_FOLDER, exist_ok=True)
|
35 |
os.makedirs(CACHE_DIR, exist_ok=True)
|
36 |
|
37 |
-
#
|
38 |
os.environ['HF_HOME'] = CACHE_DIR
|
39 |
os.environ['TRANSFORMERS_CACHE'] = os.path.join(CACHE_DIR, 'transformers')
|
40 |
-
os.environ['HF_DATASETS_CACHE'] = os.path.join(CACHE_DIR, 'datasets')
|
41 |
|
42 |
app.config['UPLOAD_FOLDER'] = UPLOAD_FOLDER
|
43 |
-
app.config['MAX_CONTENT_LENGTH'] = 16 * 1024 * 1024
|
44 |
|
45 |
-
#
|
46 |
-
|
47 |
-
|
48 |
-
# Global model variables
|
49 |
depth_estimator = None
|
50 |
model_loaded = False
|
51 |
model_loading = False
|
52 |
|
53 |
-
|
54 |
-
TIMEOUT_SECONDS = 240 # 4 minutes max for processing
|
55 |
-
MAX_DIMENSION = 512 # Max image dimension to process
|
56 |
|
57 |
-
# TimeoutError for handling timeouts
|
58 |
class TimeoutError(Exception):
|
59 |
pass
|
60 |
|
61 |
-
# Thread-safe timeout implementation
|
62 |
-
def process_with_timeout(function, args, timeout):
|
63 |
-
result = [None]
|
64 |
-
error = [None]
|
65 |
-
completed = [False]
|
66 |
-
|
67 |
-
def target():
|
68 |
-
try:
|
69 |
-
result[0] = function(*args)
|
70 |
-
completed[0] = True
|
71 |
-
except Exception as e:
|
72 |
-
error[0] = e
|
73 |
-
|
74 |
-
thread = threading.Thread(target=target)
|
75 |
-
thread.daemon = True
|
76 |
-
thread.start()
|
77 |
-
|
78 |
-
thread.join(timeout)
|
79 |
-
|
80 |
-
if not completed[0]:
|
81 |
-
if thread.is_alive():
|
82 |
-
return None, TimeoutError(f"Processing timed out after {timeout} seconds")
|
83 |
-
elif error[0]:
|
84 |
-
return None, error[0]
|
85 |
-
|
86 |
-
if error[0]:
|
87 |
-
return None, error[0]
|
88 |
-
|
89 |
-
return result[0], None
|
90 |
-
|
91 |
def allowed_file(filename):
|
92 |
return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS
|
93 |
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
# Resize if the image is too large
|
100 |
-
if img.width > MAX_DIMENSION or img.height > MAX_DIMENSION:
|
101 |
-
# Calculate new dimensions while preserving aspect ratio
|
102 |
-
if img.width > img.height:
|
103 |
-
new_width = MAX_DIMENSION
|
104 |
-
new_height = int(img.height * (MAX_DIMENSION / img.width))
|
105 |
-
else:
|
106 |
-
new_height = MAX_DIMENSION
|
107 |
-
new_width = int(img.width * (MAX_DIMENSION / img.height))
|
108 |
-
|
109 |
-
# Use high-quality Lanczos resampling for better detail preservation
|
110 |
-
img = img.resize((new_width, new_height), Image.LANCZOS)
|
111 |
-
|
112 |
-
# Convert to numpy array for additional preprocessing
|
113 |
-
img_array = np.array(img)
|
114 |
-
|
115 |
-
# Optional: Apply adaptive histogram equalization for better contrast
|
116 |
-
# This helps the depth model detect more details
|
117 |
-
if len(img_array.shape) == 3 and img_array.shape[2] == 3:
|
118 |
-
# Convert to LAB color space
|
119 |
-
lab = cv2.cvtColor(img_array, cv2.COLOR_RGB2LAB)
|
120 |
-
l, a, b = cv2.split(lab)
|
121 |
-
|
122 |
-
# Apply CLAHE to L channel
|
123 |
-
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
|
124 |
-
cl = clahe.apply(l)
|
125 |
-
|
126 |
-
# Merge channels back
|
127 |
-
enhanced_lab = cv2.merge((cl, a, b))
|
128 |
-
|
129 |
-
# Convert back to RGB
|
130 |
-
img_array = cv2.cvtColor(enhanced_lab, cv2.COLOR_LAB2RGB)
|
131 |
-
|
132 |
-
# Convert back to PIL Image
|
133 |
-
img = Image.fromarray(img_array)
|
134 |
-
|
135 |
-
return img
|
136 |
|
137 |
-
def
|
138 |
-
global depth_estimator, model_loaded
|
139 |
-
|
140 |
if model_loaded:
|
141 |
-
return
|
142 |
-
|
143 |
-
if model_loading:
|
144 |
-
# Wait for model to load if it's already in progress
|
145 |
-
while model_loading and not model_loaded:
|
146 |
-
time.sleep(0.5)
|
147 |
-
return depth_estimator
|
148 |
-
|
149 |
try:
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
#
|
158 |
-
max_retries = 3
|
159 |
-
retry_delay = 5
|
160 |
-
|
161 |
-
for attempt in range(max_retries):
|
162 |
-
try:
|
163 |
-
snapshot_download(
|
164 |
-
repo_id=model_name,
|
165 |
-
cache_dir=CACHE_DIR,
|
166 |
-
resume_download=True,
|
167 |
-
)
|
168 |
-
break
|
169 |
-
except Exception as e:
|
170 |
-
if attempt < max_retries - 1:
|
171 |
-
print(f"Download attempt {attempt+1} failed: {str(e)}. Retrying in {retry_delay} seconds...")
|
172 |
-
time.sleep(retry_delay)
|
173 |
-
retry_delay *= 2
|
174 |
-
else:
|
175 |
-
raise
|
176 |
-
|
177 |
-
# Initialize model with appropriate precision
|
178 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
179 |
-
|
180 |
-
# Load depth estimator pipeline
|
181 |
depth_estimator = pipeline(
|
182 |
-
"depth-estimation",
|
183 |
-
model=
|
184 |
-
device=device if device == "cuda" else -1,
|
185 |
cache_dir=CACHE_DIR
|
186 |
)
|
187 |
-
|
188 |
-
# Optimize memory usage
|
189 |
-
if device == "cuda":
|
190 |
-
torch.cuda.empty_cache()
|
191 |
-
|
192 |
model_loaded = True
|
193 |
-
print(
|
194 |
-
return depth_estimator
|
195 |
-
|
196 |
except Exception as e:
|
197 |
-
print(f"Error loading
|
198 |
-
print(traceback.format_exc())
|
199 |
raise
|
200 |
-
finally:
|
201 |
-
model_loading = False
|
202 |
-
|
203 |
-
# Enhanced depth processing function to improve detail quality
|
204 |
-
def enhance_depth_map(depth_map, detail_level='medium'):
|
205 |
-
"""Apply sophisticated processing to enhance depth map details"""
|
206 |
-
# Convert to numpy array if needed
|
207 |
-
if isinstance(depth_map, Image.Image):
|
208 |
-
depth_map = np.array(depth_map)
|
209 |
-
|
210 |
-
# Make sure the depth map is 2D
|
211 |
-
if len(depth_map.shape) > 2:
|
212 |
-
depth_map = np.mean(depth_map, axis=2) if depth_map.shape[2] > 1 else depth_map[:,:,0]
|
213 |
-
|
214 |
-
# Create a copy for processing
|
215 |
-
enhanced_depth = depth_map.copy().astype(np.float32)
|
216 |
-
|
217 |
-
# Remove outliers using percentile clipping (more stable than min/max)
|
218 |
-
p_low, p_high = np.percentile(enhanced_depth, [1, 99])
|
219 |
-
enhanced_depth = np.clip(enhanced_depth, p_low, p_high)
|
220 |
-
|
221 |
-
# Normalize to 0-1 range for processing
|
222 |
-
enhanced_depth = (enhanced_depth - p_low) / (p_high - p_low) if p_high > p_low else enhanced_depth
|
223 |
-
|
224 |
-
# Apply different enhancement methods based on detail level
|
225 |
-
if detail_level == 'high':
|
226 |
-
# Apply unsharp masking for edge enhancement - simulating Hunyuan's detail technique
|
227 |
-
# First apply gaussian blur
|
228 |
-
blurred = gaussian_filter(enhanced_depth, sigma=1.5)
|
229 |
-
# Create the unsharp mask
|
230 |
-
mask = enhanced_depth - blurred
|
231 |
-
# Apply the mask with strength factor
|
232 |
-
enhanced_depth = enhanced_depth + 1.5 * mask
|
233 |
-
|
234 |
-
# Apply bilateral filter to preserve edges while smoothing noise
|
235 |
-
# Simulate using gaussian combinations
|
236 |
-
smooth1 = gaussian_filter(enhanced_depth, sigma=0.5)
|
237 |
-
smooth2 = gaussian_filter(enhanced_depth, sigma=2.0)
|
238 |
-
edge_mask = enhanced_depth - smooth2
|
239 |
-
enhanced_depth = smooth1 + 1.2 * edge_mask
|
240 |
-
|
241 |
-
elif detail_level == 'medium':
|
242 |
-
# Less aggressive but still effective enhancement
|
243 |
-
# Apply mild unsharp masking
|
244 |
-
blurred = gaussian_filter(enhanced_depth, sigma=1.0)
|
245 |
-
mask = enhanced_depth - blurred
|
246 |
-
enhanced_depth = enhanced_depth + 0.8 * mask
|
247 |
-
|
248 |
-
# Apply mild smoothing to reduce noise but preserve edges
|
249 |
-
enhanced_depth = gaussian_filter(enhanced_depth, sigma=0.5)
|
250 |
-
|
251 |
-
else: # low
|
252 |
-
# Just apply noise reduction without too much detail enhancement
|
253 |
-
enhanced_depth = gaussian_filter(enhanced_depth, sigma=0.7)
|
254 |
-
|
255 |
-
# Normalize again after processing
|
256 |
-
enhanced_depth = np.clip(enhanced_depth, 0, 1)
|
257 |
-
|
258 |
-
return enhanced_depth
|
259 |
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
)
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
# Normalize z-values with advanced scaling for better depth impression
|
297 |
-
z_min, z_max = np.percentile(z_values, [2, 98]) # Remove outliers
|
298 |
-
z_values = (z_values - z_min) / (z_max - z_min) if z_max > z_min else z_values
|
299 |
-
|
300 |
-
# Apply depth scaling appropriate to the detail level
|
301 |
-
if detail_level == 'high':
|
302 |
-
z_scaling = 2.5 # More pronounced depth variations
|
303 |
-
elif detail_level == 'medium':
|
304 |
-
z_scaling = 2.0 # Standard depth
|
305 |
-
else:
|
306 |
-
z_scaling = 1.5 # More subtle depth variations
|
307 |
-
|
308 |
-
z_values = z_values * z_scaling
|
309 |
-
|
310 |
-
# Normalize x and y coordinates
|
311 |
-
x_grid = (x_grid / w - 0.5) * 2.0 # Map to -1 to 1
|
312 |
-
y_grid = (y_grid / h - 0.5) * 2.0 # Map to -1 to 1
|
313 |
-
|
314 |
-
# Create vertices
|
315 |
-
vertices = np.vstack([x_grid.flatten(), -y_grid.flatten(), -z_values.flatten()]).T
|
316 |
-
|
317 |
-
# Create faces (triangles) with optimized winding for better normals
|
318 |
-
faces = []
|
319 |
-
for i in range(resolution-1):
|
320 |
-
for j in range(resolution-1):
|
321 |
-
p1 = i * resolution + j
|
322 |
-
p2 = i * resolution + (j + 1)
|
323 |
-
p3 = (i + 1) * resolution + j
|
324 |
-
p4 = (i + 1) * resolution + (j + 1)
|
325 |
-
|
326 |
-
# Calculate normals to ensure consistent orientation
|
327 |
-
v1 = vertices[p1]
|
328 |
-
v2 = vertices[p2]
|
329 |
-
v3 = vertices[p3]
|
330 |
-
v4 = vertices[p4]
|
331 |
-
|
332 |
-
# Calculate normals for both possible triangulations
|
333 |
-
# and choose the one that's more consistent
|
334 |
-
norm1 = np.cross(v2-v1, v4-v1)
|
335 |
-
norm2 = np.cross(v4-v3, v1-v3)
|
336 |
-
|
337 |
-
if np.dot(norm1, norm2) >= 0:
|
338 |
-
# Standard triangulation
|
339 |
-
faces.append([p1, p2, p4])
|
340 |
-
faces.append([p1, p4, p3])
|
341 |
-
else:
|
342 |
-
# Alternative triangulation for smoother surface
|
343 |
-
faces.append([p1, p2, p3])
|
344 |
-
faces.append([p2, p4, p3])
|
345 |
-
|
346 |
-
faces = np.array(faces)
|
347 |
-
|
348 |
-
# Create mesh
|
349 |
-
mesh = trimesh.Trimesh(vertices=vertices, faces=faces)
|
350 |
-
|
351 |
-
# Apply advanced texturing if image is provided
|
352 |
-
if image:
|
353 |
-
# Convert to numpy array if needed
|
354 |
-
if isinstance(image, Image.Image):
|
355 |
-
img_array = np.array(image)
|
356 |
-
else:
|
357 |
-
img_array = image
|
358 |
-
|
359 |
-
# Create vertex colors with improved sampling
|
360 |
-
if resolution <= img_array.shape[0] and resolution <= img_array.shape[1]:
|
361 |
-
# Create vertex colors by sampling the image with bilinear interpolation
|
362 |
-
vertex_colors = np.zeros((vertices.shape[0], 4), dtype=np.uint8)
|
363 |
-
|
364 |
-
# Get normalized coordinates for sampling
|
365 |
-
for i in range(resolution):
|
366 |
-
for j in range(resolution):
|
367 |
-
# Calculate exact image coordinates with proper scaling
|
368 |
-
img_x = j * (img_array.shape[1] - 1) / (resolution - 1)
|
369 |
-
img_y = i * (img_array.shape[0] - 1) / (resolution - 1)
|
370 |
-
|
371 |
-
# Bilinear interpolation for smooth color transitions
|
372 |
-
x0, y0 = int(img_x), int(img_y)
|
373 |
-
x1, y1 = min(x0 + 1, img_array.shape[1] - 1), min(y0 + 1, img_array.shape[0] - 1)
|
374 |
-
|
375 |
-
# Calculate interpolation weights
|
376 |
-
wx = img_x - x0
|
377 |
-
wy = img_y - y0
|
378 |
-
|
379 |
-
vertex_idx = i * resolution + j
|
380 |
-
|
381 |
-
if len(img_array.shape) == 3 and img_array.shape[2] == 3: # RGB
|
382 |
-
# Perform bilinear interpolation for each color channel
|
383 |
-
r = int((1-wx)*(1-wy)*img_array[y0, x0, 0] + wx*(1-wy)*img_array[y0, x1, 0] +
|
384 |
-
(1-wx)*wy*img_array[y1, x0, 0] + wx*wy*img_array[y1, x1, 0])
|
385 |
-
g = int((1-wx)*(1-wy)*img_array[y0, x0, 1] + wx*(1-wy)*img_array[y0, x1, 1] +
|
386 |
-
(1-wx)*wy*img_array[y1, x0, 1] + wx*wy*img_array[y1, x1, 1])
|
387 |
-
b = int((1-wx)*(1-wy)*img_array[y0, x0, 2] + wx*(1-wy)*img_array[y0, x1, 2] +
|
388 |
-
(1-wx)*wy*img_array[y1, x0, 2] + wx*wy*img_array[y1, x1, 2])
|
389 |
-
|
390 |
-
vertex_colors[vertex_idx, :3] = [r, g, b]
|
391 |
-
vertex_colors[vertex_idx, 3] = 255 # Alpha
|
392 |
-
elif len(img_array.shape) == 3 and img_array.shape[2] == 4: # RGBA
|
393 |
-
for c in range(4): # For each RGBA channel
|
394 |
-
vertex_colors[vertex_idx, c] = int((1-wx)*(1-wy)*img_array[y0, x0, c] +
|
395 |
-
wx*(1-wy)*img_array[y0, x1, c] +
|
396 |
-
(1-wx)*wy*img_array[y1, x0, c] +
|
397 |
-
wx*wy*img_array[y1, x1, c])
|
398 |
-
else:
|
399 |
-
# Handle grayscale with bilinear interpolation
|
400 |
-
gray = int((1-wx)*(1-wy)*img_array[y0, x0] + wx*(1-wy)*img_array[y0, x1] +
|
401 |
-
(1-wx)*wy*img_array[y1, x0] + wx*wy*img_array[y1, x1])
|
402 |
-
vertex_colors[vertex_idx, :3] = [gray, gray, gray]
|
403 |
-
vertex_colors[vertex_idx, 3] = 255
|
404 |
-
|
405 |
-
mesh.visual.vertex_colors = vertex_colors
|
406 |
-
|
407 |
-
# Apply smoothing to get rid of staircase artifacts
|
408 |
-
if detail_level != 'high':
|
409 |
-
# For medium and low detail, apply Laplacian smoothing
|
410 |
-
# but preserve the overall shape
|
411 |
-
mesh = mesh.smoothed(method='laplacian', iterations=1)
|
412 |
-
|
413 |
-
# Calculate and fix normals for better rendering
|
414 |
-
mesh.fix_normals()
|
415 |
-
|
416 |
return mesh
|
417 |
|
418 |
-
@app.route('/health', methods=['GET'])
|
419 |
-
def health_check():
|
420 |
-
return jsonify({
|
421 |
-
"status": "healthy",
|
422 |
-
"model": "Enhanced Depth-Based 3D Model Generator (DPT-Large)",
|
423 |
-
"device": "cuda" if torch.cuda.is_available() else "cpu"
|
424 |
-
}), 200
|
425 |
-
|
426 |
-
@app.route('/progress/<job_id>', methods=['GET'])
|
427 |
-
def progress(job_id):
|
428 |
-
def generate():
|
429 |
-
if job_id not in processing_jobs:
|
430 |
-
yield f"data: {json.dumps({'error': 'Job not found'})}\n\n"
|
431 |
-
return
|
432 |
-
|
433 |
-
job = processing_jobs[job_id]
|
434 |
-
|
435 |
-
# Send initial progress
|
436 |
-
yield f"data: {json.dumps({'status': 'processing', 'progress': job['progress']})}\n\n"
|
437 |
-
|
438 |
-
# Wait for job to complete or update
|
439 |
-
last_progress = job['progress']
|
440 |
-
check_count = 0
|
441 |
-
while job['status'] == 'processing':
|
442 |
-
if job['progress'] != last_progress:
|
443 |
-
yield f"data: {json.dumps({'status': 'processing', 'progress': job['progress']})}\n\n"
|
444 |
-
last_progress = job['progress']
|
445 |
-
|
446 |
-
time.sleep(0.5)
|
447 |
-
check_count += 1
|
448 |
-
|
449 |
-
# If client hasn't received updates for a while, check if job is still running
|
450 |
-
if check_count > 60: # 30 seconds with no updates
|
451 |
-
if 'thread_alive' in job and not job['thread_alive']():
|
452 |
-
job['status'] = 'error'
|
453 |
-
job['error'] = 'Processing thread died unexpectedly'
|
454 |
-
break
|
455 |
-
check_count = 0
|
456 |
-
|
457 |
-
# Send final status
|
458 |
-
if job['status'] == 'completed':
|
459 |
-
yield f"data: {json.dumps({'status': 'completed', 'progress': 100, 'result_url': job['result_url'], 'preview_url': job['preview_url']})}\n\n"
|
460 |
-
else:
|
461 |
-
yield f"data: {json.dumps({'status': 'error', 'error': job['error']})}\n\n"
|
462 |
-
|
463 |
-
return Response(stream_with_context(generate()), mimetype='text/event-stream')
|
464 |
-
|
465 |
@app.route('/convert', methods=['POST'])
|
466 |
def convert_image_to_3d():
|
467 |
-
# Check if image is in the request
|
468 |
if 'image' not in request.files:
|
469 |
return jsonify({"error": "No image provided"}), 400
|
470 |
|
471 |
file = request.files['image']
|
472 |
-
if file.filename == '':
|
473 |
-
return jsonify({"error": "No image selected"}), 400
|
474 |
-
|
475 |
if not allowed_file(file.filename):
|
476 |
-
return jsonify({"error":
|
477 |
-
|
478 |
-
# Get optional parameters with defaults
|
479 |
-
try:
|
480 |
-
mesh_resolution = min(int(request.form.get('mesh_resolution', 100)), 200) # Limit max resolution
|
481 |
-
output_format = request.form.get('output_format', 'obj').lower()
|
482 |
-
detail_level = request.form.get('detail_level', 'medium').lower() # Parameter for detail level
|
483 |
-
texture_quality = request.form.get('texture_quality', 'medium').lower() # New parameter for texture quality
|
484 |
-
except ValueError:
|
485 |
-
return jsonify({"error": "Invalid parameter values"}), 400
|
486 |
-
|
487 |
-
# Validate output format
|
488 |
-
if output_format not in ['obj', 'glb']:
|
489 |
-
return jsonify({"error": "Unsupported output format. Use 'obj' or 'glb'"}), 400
|
490 |
-
|
491 |
-
# Adjust mesh resolution based on detail level
|
492 |
-
if detail_level == 'high':
|
493 |
-
mesh_resolution = min(int(mesh_resolution * 1.5), 200)
|
494 |
-
elif detail_level == 'low':
|
495 |
-
mesh_resolution = max(int(mesh_resolution * 0.7), 50)
|
496 |
-
|
497 |
-
# Create a job ID
|
498 |
job_id = str(uuid.uuid4())
|
499 |
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
500 |
os.makedirs(output_dir, exist_ok=True)
|
501 |
-
|
502 |
-
# Save the uploaded file
|
503 |
filename = secure_filename(file.filename)
|
504 |
filepath = os.path.join(app.config['UPLOAD_FOLDER'], f"{job_id}_{filename}")
|
505 |
file.save(filepath)
|
506 |
-
|
507 |
-
# Initialize job tracking
|
508 |
processing_jobs[job_id] = {
|
509 |
'status': 'processing',
|
510 |
'progress': 0,
|
511 |
'result_url': None,
|
512 |
-
'
|
513 |
-
'error': None,
|
514 |
-
'output_format': output_format,
|
515 |
-
'created_at': time.time()
|
516 |
}
|
517 |
-
|
518 |
-
# Start processing in a separate thread
|
519 |
def process_image():
|
520 |
-
thread = threading.current_thread()
|
521 |
-
processing_jobs[job_id]['thread_alive'] = lambda: thread.is_alive()
|
522 |
-
|
523 |
try:
|
524 |
-
# Preprocess image
|
525 |
-
|
526 |
-
|
527 |
-
|
528 |
-
|
529 |
-
|
530 |
-
|
531 |
-
|
532 |
-
|
533 |
-
|
534 |
-
|
535 |
-
|
536 |
-
|
537 |
-
|
538 |
-
# Process image with thread-safe timeout
|
539 |
-
try:
|
540 |
-
def estimate_depth():
|
541 |
-
# Get depth map
|
542 |
-
result = model(image)
|
543 |
-
depth_map = result["depth"]
|
544 |
-
|
545 |
-
# Convert to numpy array if needed
|
546 |
-
if isinstance(depth_map, torch.Tensor):
|
547 |
-
depth_map = depth_map.cpu().numpy()
|
548 |
-
elif hasattr(depth_map, 'numpy'):
|
549 |
-
depth_map = depth_map.numpy()
|
550 |
-
elif isinstance(depth_map, Image.Image):
|
551 |
-
depth_map = np.array(depth_map)
|
552 |
-
|
553 |
-
return depth_map
|
554 |
-
|
555 |
-
depth_map, error = process_with_timeout(estimate_depth, [], TIMEOUT_SECONDS)
|
556 |
|
557 |
-
|
558 |
-
|
559 |
-
|
560 |
-
|
561 |
-
return
|
562 |
-
else:
|
563 |
-
raise error
|
564 |
-
|
565 |
-
processing_jobs[job_id]['progress'] = 60
|
566 |
-
|
567 |
-
# Create mesh from depth map with enhanced detail handling
|
568 |
-
mesh_resolution_int = int(mesh_resolution)
|
569 |
-
mesh = depth_to_mesh(depth_map, image, resolution=mesh_resolution_int, detail_level=detail_level)
|
570 |
-
processing_jobs[job_id]['progress'] = 80
|
571 |
-
|
572 |
-
except Exception as e:
|
573 |
-
error_details = traceback.format_exc()
|
574 |
-
processing_jobs[job_id]['status'] = 'error'
|
575 |
-
processing_jobs[job_id]['error'] = f"Error during processing: {str(e)}"
|
576 |
-
print(f"Error processing job {job_id}: {str(e)}")
|
577 |
-
print(error_details)
|
578 |
-
return
|
579 |
-
|
580 |
-
# Export based on requested format with enhanced quality settings
|
581 |
-
try:
|
582 |
-
if output_format == 'obj':
|
583 |
-
obj_path = os.path.join(output_dir, "model.obj")
|
584 |
-
|
585 |
-
# Export with normal and texture coordinates
|
586 |
-
mesh.export(
|
587 |
-
obj_path,
|
588 |
-
file_type='obj',
|
589 |
-
include_normals=True,
|
590 |
-
include_texture=True
|
591 |
-
)
|
592 |
-
|
593 |
-
# Create a zip file with OBJ and MTL
|
594 |
-
zip_path = os.path.join(output_dir, "model.zip")
|
595 |
-
with zipfile.ZipFile(zip_path, 'w') as zipf:
|
596 |
-
zipf.write(obj_path, arcname="model.obj")
|
597 |
-
mtl_path = os.path.join(output_dir, "model.mtl")
|
598 |
-
if os.path.exists(mtl_path):
|
599 |
-
zipf.write(mtl_path, arcname="model.mtl")
|
600 |
-
|
601 |
-
# Include texture file if it exists
|
602 |
-
texture_path = os.path.join(output_dir, "model.png")
|
603 |
-
if os.path.exists(texture_path):
|
604 |
-
zipf.write(texture_path, arcname="model.png")
|
605 |
-
|
606 |
-
processing_jobs[job_id]['result_url'] = f"/download/{job_id}"
|
607 |
-
processing_jobs[job_id]['preview_url'] = f"/preview/{job_id}"
|
608 |
-
|
609 |
-
elif output_format == 'glb':
|
610 |
-
# Export as GLB with enhanced settings
|
611 |
-
glb_path = os.path.join(output_dir, "model.glb")
|
612 |
-
mesh.export(
|
613 |
-
glb_path,
|
614 |
-
file_type='glb'
|
615 |
-
)
|
616 |
-
|
617 |
-
processing_jobs[job_id]['result_url'] = f"/download/{job_id}"
|
618 |
-
processing_jobs[job_id]['preview_url'] = f"/preview/{job_id}"
|
619 |
-
|
620 |
-
# Update job status
|
621 |
-
processing_jobs[job_id]['status'] = 'completed'
|
622 |
-
processing_jobs[job_id]['progress'] = 100
|
623 |
-
print(f"Job {job_id} completed successfully")
|
624 |
-
except Exception as e:
|
625 |
-
error_details = traceback.format_exc()
|
626 |
-
processing_jobs[job_id]['status'] = 'error'
|
627 |
-
processing_jobs[job_id]['error'] = f"Error exporting model: {str(e)}"
|
628 |
-
print(f"Error exporting model for job {job_id}: {str(e)}")
|
629 |
-
print(error_details)
|
630 |
-
|
631 |
-
# Clean up temporary file
|
632 |
-
if os.path.exists(filepath):
|
633 |
-
os.remove(filepath)
|
634 |
-
|
635 |
-
# Force garbage collection to free memory
|
636 |
-
gc.collect()
|
637 |
-
if torch.cuda.is_available():
|
638 |
-
torch.cuda.empty_cache()
|
639 |
-
|
640 |
-
except Exception as e:
|
641 |
-
# Handle errors
|
642 |
-
error_details = traceback.format_exc()
|
643 |
-
processing_jobs[job_id]['status'] = 'error'
|
644 |
-
processing_jobs[job_id]['error'] = f"{str(e)}\n{error_details}"
|
645 |
-
print(f"Error processing job {job_id}: {str(e)}")
|
646 |
-
print(error_details)
|
647 |
-
|
648 |
-
# Clean up on error
|
649 |
-
if os.path.exists(filepath):
|
650 |
-
os.remove(filepath)
|
651 |
-
|
652 |
-
# Start processing thread
|
653 |
-
processing_thread = threading.Thread(target=process_image)
|
654 |
-
processing_thread.daemon = True
|
655 |
-
processing_thread.start()
|
656 |
-
|
657 |
-
# Return job ID immediately
|
658 |
-
return jsonify({"job_id": job_id}), 202 # 202 Accepted
|
659 |
-
|
660 |
-
@app.route('/download/<job_id>', methods=['GET'])
|
661 |
-
def download_model(job_id):
|
662 |
-
if job_id not in processing_jobs or processing_jobs[job_id]['status'] != 'completed':
|
663 |
-
return jsonify({"error": "Model not found or processing not complete"}), 404
|
664 |
-
|
665 |
-
# Get the output directory for this job
|
666 |
-
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
667 |
-
|
668 |
-
# Determine file format from the job data
|
669 |
-
output_format = processing_jobs[job_id].get('output_format', 'obj')
|
670 |
-
|
671 |
-
if output_format == 'obj':
|
672 |
-
zip_path = os.path.join(output_dir, "model.zip")
|
673 |
-
if os.path.exists(zip_path):
|
674 |
-
return send_file(zip_path, as_attachment=True, download_name="model.zip")
|
675 |
-
else: # glb
|
676 |
-
glb_path = os.path.join(output_dir, "model.glb")
|
677 |
-
if os.path.exists(glb_path):
|
678 |
-
return send_file(glb_path, as_attachment=True, download_name="model.glb")
|
679 |
-
|
680 |
-
return jsonify({"error": "File not found"}), 404
|
681 |
-
|
682 |
-
@app.route('/preview/<job_id>', methods=['GET'])
|
683 |
-
def preview_model(job_id):
|
684 |
-
if job_id not in processing_jobs or processing_jobs[job_id]['status'] != 'completed':
|
685 |
-
return jsonify({"error": "Model not found or processing not complete"}), 404
|
686 |
-
|
687 |
-
# Get the output directory for this job
|
688 |
-
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
689 |
-
output_format = processing_jobs[job_id].get('output_format', 'obj')
|
690 |
-
|
691 |
-
if output_format == 'obj':
|
692 |
-
obj_path = os.path.join(output_dir, "model.obj")
|
693 |
-
if os.path.exists(obj_path):
|
694 |
-
return send_file(obj_path, mimetype='model/obj')
|
695 |
-
else: # glb
|
696 |
-
glb_path = os.path.join(output_dir, "model.glb")
|
697 |
-
if os.path.exists(glb_path):
|
698 |
-
return send_file(glb_path, mimetype='model/gltf-binary')
|
699 |
-
|
700 |
-
return jsonify({"error": "Model file not found"}), 404
|
701 |
-
|
702 |
-
# Cleanup old jobs periodically
|
703 |
-
def cleanup_old_jobs():
|
704 |
-
current_time = time.time()
|
705 |
-
job_ids_to_remove = []
|
706 |
-
|
707 |
-
for job_id, job_data in processing_jobs.items():
|
708 |
-
# Remove completed jobs after 1 hour
|
709 |
-
if job_data['status'] == 'completed' and (current_time - job_data.get('created_at', 0)) > 3600:
|
710 |
-
job_ids_to_remove.append(job_id)
|
711 |
-
# Remove error jobs after 30 minutes
|
712 |
-
elif job_data['status'] == 'error' and (current_time - job_data.get('created_at', 0)) > 1800:
|
713 |
-
job_ids_to_remove.append(job_id)
|
714 |
-
|
715 |
-
# Remove the jobs
|
716 |
-
for job_id in job_ids_to_remove:
|
717 |
-
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
718 |
-
try:
|
719 |
-
import shutil
|
720 |
-
if os.path.exists(output_dir):
|
721 |
-
shutil.rmtree(output_dir)
|
722 |
-
except Exception as e:
|
723 |
-
print(f"Error cleaning up job {job_id}: {str(e)}")
|
724 |
-
|
725 |
-
# Remove from tracking dictionary
|
726 |
-
if job_id in processing_jobs:
|
727 |
-
del processing_jobs[job_id]
|
728 |
-
|
729 |
-
# Schedule the next cleanup
|
730 |
-
threading.Timer(300, cleanup_old_jobs).start() # Run every 5 minutes
|
731 |
-
|
732 |
-
# New endpoint to get detailed information about a model
|
733 |
-
@app.route('/model-info/<job_id>', methods=['GET'])
|
734 |
-
def model_info(job_id):
|
735 |
-
if job_id not in processing_jobs:
|
736 |
-
return jsonify({"error": "Model not found"}), 404
|
737 |
-
|
738 |
-
job = processing_jobs[job_id]
|
739 |
-
|
740 |
-
if job['status'] != 'completed':
|
741 |
-
return jsonify({
|
742 |
-
"status": job['status'],
|
743 |
-
"progress": job['progress'],
|
744 |
-
"error": job.get('error')
|
745 |
-
}), 200
|
746 |
-
|
747 |
-
# For completed jobs, include information about the model
|
748 |
-
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
749 |
-
model_stats = {}
|
750 |
-
|
751 |
-
# Get file size
|
752 |
-
if job['output_format'] == 'obj':
|
753 |
-
obj_path = os.path.join(output_dir, "model.obj")
|
754 |
-
zip_path = os.path.join(output_dir, "model.zip")
|
755 |
-
|
756 |
-
if os.path.exists(obj_path):
|
757 |
-
model_stats['obj_size'] = os.path.getsize(obj_path)
|
758 |
-
|
759 |
-
if os.path.exists(zip_path):
|
760 |
-
model_stats['package_size'] = os.path.getsize(zip_path)
|
761 |
-
|
762 |
-
else: # glb
|
763 |
-
glb_path = os.path.join(output_dir, "model.glb")
|
764 |
-
if os.path.exists(glb_path):
|
765 |
-
model_stats['model_size'] = os.path.getsize(glb_path)
|
766 |
-
|
767 |
-
# Return detailed info
|
768 |
-
return jsonify({
|
769 |
-
"status": job['status'],
|
770 |
-
"model_format": job['output_format'],
|
771 |
-
"download_url": job['result_url'],
|
772 |
-
"preview_url": job['preview_url'],
|
773 |
-
"model_stats": model_stats,
|
774 |
-
"created_at": job.get('created_at'),
|
775 |
-
"completed_at": job.get('completed_at')
|
776 |
-
}), 200
|
777 |
|
778 |
-
|
779 |
-
|
780 |
-
|
781 |
-
"message": "Enhanced Image to 3D API (DPT-Large Model)",
|
782 |
-
"endpoints": [
|
783 |
-
"/convert",
|
784 |
-
"/progress/<job_id>",
|
785 |
-
"/download/<job_id>",
|
786 |
-
"/preview/<job_id>",
|
787 |
-
"/model-info/<job_id>"
|
788 |
-
],
|
789 |
-
"parameters": {
|
790 |
-
"mesh_resolution": "Integer (50-200), controls mesh density",
|
791 |
-
"output_format": "obj or glb",
|
792 |
-
"detail_level": "low, medium, or high - controls the level of detail in the final model",
|
793 |
-
"texture_quality": "low, medium, or high - controls the quality of textures"
|
794 |
-
},
|
795 |
-
"description": "This API creates high-quality 3D models from 2D images with enhanced detail finishing similar to Hunyuan model"
|
796 |
-
}), 200
|
797 |
|
798 |
-
#
|
799 |
-
|
800 |
-
def compare_detail_levels():
|
801 |
-
# Check if image is in the request
|
802 |
-
if 'image' not in request.files:
|
803 |
-
return jsonify({"error": "No image provided"}), 400
|
804 |
-
|
805 |
-
file = request.files['image']
|
806 |
-
if file.filename == '':
|
807 |
-
return jsonify({"error": "No image selected"}), 400
|
808 |
-
|
809 |
-
if not allowed_file(file.filename):
|
810 |
-
return jsonify({"error": f"File type not allowed. Supported types: {', '.join(ALLOWED_EXTENSIONS)}"}), 400
|
811 |
-
|
812 |
-
# Create a job ID
|
813 |
-
job_id = str(uuid.uuid4())
|
814 |
-
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
815 |
-
os.makedirs(output_dir, exist_ok=True)
|
816 |
-
|
817 |
-
# Save the uploaded file
|
818 |
-
filename = secure_filename(file.filename)
|
819 |
-
filepath = os.path.join(app.config['UPLOAD_FOLDER'], f"{job_id}_{filename}")
|
820 |
-
file.save(filepath)
|
821 |
-
|
822 |
-
# Initialize job tracking
|
823 |
-
processing_jobs[job_id] = {
|
824 |
-
'status': 'processing',
|
825 |
-
'progress': 0,
|
826 |
-
'result_url': None,
|
827 |
-
'preview_url': None,
|
828 |
-
'error': None,
|
829 |
-
'output_format': 'glb', # Use GLB for comparison
|
830 |
-
'created_at': time.time(),
|
831 |
-
'comparison': True
|
832 |
-
}
|
833 |
-
|
834 |
-
# Process in separate thread to create 3 different detail levels
|
835 |
-
def process_comparison():
|
836 |
-
thread = threading.current_thread()
|
837 |
-
processing_jobs[job_id]['thread_alive'] = lambda: thread.is_alive()
|
838 |
-
|
839 |
-
try:
|
840 |
-
# Preprocess image
|
841 |
-
image = preprocess_image(filepath)
|
842 |
-
processing_jobs[job_id]['progress'] = 10
|
843 |
|
844 |
-
#
|
845 |
-
|
846 |
-
|
847 |
-
processing_jobs[job_id]['progress'] = 20
|
848 |
-
except Exception as e:
|
849 |
-
processing_jobs[job_id]['status'] = 'error'
|
850 |
-
processing_jobs[job_id]['error'] = f"Error loading model: {str(e)}"
|
851 |
-
return
|
852 |
|
853 |
-
# Process image to get depth map
|
854 |
-
try:
|
855 |
-
depth_map = model(image)["depth"]
|
856 |
-
if isinstance(depth_map, torch.Tensor):
|
857 |
-
depth_map = depth_map.cpu().numpy()
|
858 |
-
elif hasattr(depth_map, 'numpy'):
|
859 |
-
depth_map = depth_map.numpy()
|
860 |
-
elif isinstance(depth_map, Image.Image):
|
861 |
-
depth_map = np.array(depth_map)
|
862 |
-
|
863 |
-
processing_jobs[job_id]['progress'] = 40
|
864 |
-
except Exception as e:
|
865 |
-
processing_jobs[job_id]['status'] = 'error'
|
866 |
-
processing_jobs[job_id]['error'] = f"Error estimating depth: {str(e)}"
|
867 |
-
return
|
868 |
-
|
869 |
-
# Create meshes at different detail levels
|
870 |
-
result_urls = {}
|
871 |
-
|
872 |
-
for detail_level in ['low', 'medium', 'high']:
|
873 |
-
try:
|
874 |
-
# Update progress
|
875 |
-
if detail_level == 'low':
|
876 |
-
processing_jobs[job_id]['progress'] = 50
|
877 |
-
elif detail_level == 'medium':
|
878 |
-
processing_jobs[job_id]['progress'] = 70
|
879 |
-
else:
|
880 |
-
processing_jobs[job_id]['progress'] = 90
|
881 |
-
|
882 |
-
# Create mesh with appropriate detail level
|
883 |
-
mesh_resolution = 100 # Fixed resolution for fair comparison
|
884 |
-
if detail_level == 'high':
|
885 |
-
mesh_resolution = 150
|
886 |
-
elif detail_level == 'low':
|
887 |
-
mesh_resolution = 80
|
888 |
-
|
889 |
-
mesh = depth_to_mesh(depth_map, image,
|
890 |
-
resolution=mesh_resolution,
|
891 |
-
detail_level=detail_level)
|
892 |
-
|
893 |
-
# Export as GLB
|
894 |
-
model_path = os.path.join(output_dir, f"model_{detail_level}.glb")
|
895 |
-
mesh.export(model_path, file_type='glb')
|
896 |
-
|
897 |
-
# Add to result URLs
|
898 |
-
result_urls[detail_level] = f"/compare-download/{job_id}/{detail_level}"
|
899 |
-
|
900 |
-
except Exception as e:
|
901 |
-
print(f"Error processing {detail_level} detail level: {str(e)}")
|
902 |
-
# Continue with other detail levels even if one fails
|
903 |
-
|
904 |
-
# Update job status
|
905 |
processing_jobs[job_id]['status'] = 'completed'
|
|
|
906 |
processing_jobs[job_id]['progress'] = 100
|
907 |
-
|
908 |
-
processing_jobs[job_id]['completed_at'] = time.time()
|
909 |
-
|
910 |
-
# Clean up temporary file
|
911 |
-
if os.path.exists(filepath):
|
912 |
-
os.remove(filepath)
|
913 |
-
|
914 |
-
# Force garbage collection
|
915 |
-
gc.collect()
|
916 |
-
if torch.cuda.is_available():
|
917 |
-
torch.cuda.empty_cache()
|
918 |
-
|
919 |
except Exception as e:
|
920 |
-
# Handle errors
|
921 |
processing_jobs[job_id]['status'] = 'error'
|
922 |
-
processing_jobs[job_id]['error'] =
|
923 |
-
|
924 |
-
# Clean up on error
|
925 |
if os.path.exists(filepath):
|
926 |
os.remove(filepath)
|
927 |
-
|
928 |
-
|
929 |
-
processing_thread = threading.Thread(target=process_comparison)
|
930 |
-
processing_thread.daemon = True
|
931 |
-
processing_thread.start()
|
932 |
-
|
933 |
-
# Return job ID immediately
|
934 |
-
return jsonify({"job_id": job_id, "check_progress_at": f"/progress/{job_id}"}), 202
|
935 |
|
936 |
-
|
937 |
-
|
|
|
|
|
|
|
|
|
938 |
if job_id not in processing_jobs or processing_jobs[job_id]['status'] != 'completed':
|
939 |
-
return jsonify({"error": "
|
940 |
|
941 |
-
|
942 |
-
|
943 |
-
|
944 |
-
|
945 |
-
|
946 |
-
|
947 |
-
|
948 |
-
|
949 |
-
|
950 |
-
|
951 |
-
|
952 |
-
return send_file(model_path, as_attachment=True, download_name=f"model_{detail_level}.glb")
|
953 |
-
|
954 |
-
return jsonify({"error": "File not found"}), 404
|
955 |
|
956 |
if __name__ == '__main__':
|
957 |
-
|
958 |
-
|
959 |
-
|
960 |
-
# Use port 7860 which is standard for Hugging Face Spaces
|
961 |
-
port = int(os.environ.get('PORT', 7860))
|
962 |
-
app.run(host='0.0.0.0', port=port)
|
|
|
16 |
import numpy as np
|
17 |
import trimesh
|
18 |
from transformers import pipeline
|
19 |
+
from diffusers import StableDiffusionZero123Pipeline
|
20 |
+
import imageio
|
21 |
+
from scipy.spatial.transform import Rotation
|
22 |
|
23 |
app = Flask(__name__)
|
24 |
+
CORS(app)
|
25 |
|
26 |
+
# Configuration
|
27 |
UPLOAD_FOLDER = '/tmp/uploads'
|
28 |
RESULTS_FOLDER = '/tmp/results'
|
29 |
CACHE_DIR = '/tmp/huggingface'
|
30 |
ALLOWED_EXTENSIONS = {'png', 'jpg', 'jpeg'}
|
31 |
+
VIEW_ANGLES = [(30, 0), (30, 90), (30, 180), (30, 270)] # (elevation, azimuth)
|
32 |
|
|
|
33 |
os.makedirs(UPLOAD_FOLDER, exist_ok=True)
|
34 |
os.makedirs(RESULTS_FOLDER, exist_ok=True)
|
35 |
os.makedirs(CACHE_DIR, exist_ok=True)
|
36 |
|
37 |
+
# Environment variables for caching
|
38 |
os.environ['HF_HOME'] = CACHE_DIR
|
39 |
os.environ['TRANSFORMERS_CACHE'] = os.path.join(CACHE_DIR, 'transformers')
|
|
|
40 |
|
41 |
app.config['UPLOAD_FOLDER'] = UPLOAD_FOLDER
|
42 |
+
app.config['MAX_CONTENT_LENGTH'] = 16 * 1024 * 1024
|
43 |
|
44 |
+
# Global models
|
45 |
+
view_generator = None
|
|
|
|
|
46 |
depth_estimator = None
|
47 |
model_loaded = False
|
48 |
model_loading = False
|
49 |
|
50 |
+
processing_jobs = {}
|
|
|
|
|
51 |
|
|
|
52 |
class TimeoutError(Exception):
|
53 |
pass
|
54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
def allowed_file(filename):
|
56 |
return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS
|
57 |
|
58 |
+
def preprocess_image(image_path, size=256):
|
59 |
+
img = Image.open(image_path).convert("RGB")
|
60 |
+
img = img.resize((size, size), Image.LANCZOS)
|
61 |
+
return img
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
63 |
+
def load_models():
|
64 |
+
global view_generator, depth_estimator, model_loaded
|
|
|
65 |
if model_loaded:
|
66 |
+
return
|
67 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
try:
|
69 |
+
# Load view generator
|
70 |
+
view_generator = StableDiffusionZero123Pipeline.from_pretrained(
|
71 |
+
"stabilityai/stable-zero123-6dof",
|
72 |
+
torch_dtype=torch.float16,
|
73 |
+
cache_dir=CACHE_DIR
|
74 |
+
).to("cuda" if torch.cuda.is_available() else "cpu")
|
75 |
+
|
76 |
+
# Load depth estimator
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
depth_estimator = pipeline(
|
78 |
+
"depth-estimation",
|
79 |
+
model="Intel/dpt-hybrid-midas",
|
|
|
80 |
cache_dir=CACHE_DIR
|
81 |
)
|
82 |
+
|
|
|
|
|
|
|
|
|
83 |
model_loaded = True
|
84 |
+
print("Models loaded successfully")
|
|
|
|
|
85 |
except Exception as e:
|
86 |
+
print(f"Error loading models: {str(e)}")
|
|
|
87 |
raise
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
|
89 |
+
def generate_novel_views(image, num_views=4):
|
90 |
+
views = []
|
91 |
+
for elevation, azimuth in VIEW_ANGLES:
|
92 |
+
result = view_generator(
|
93 |
+
image,
|
94 |
+
num_inference_steps=50,
|
95 |
+
elevation=elevation,
|
96 |
+
azimuth=azimuth,
|
97 |
+
guidance_scale=3.0
|
98 |
+
).images[0]
|
99 |
+
views.append((result, (elevation, azimuth)))
|
100 |
+
return views
|
101 |
+
|
102 |
+
def depth_to_pointcloud(depth_map, pose, fov=60):
|
103 |
+
h, w = depth_map.shape
|
104 |
+
f = w / (2 * np.tan(np.radians(fov/2)))
|
105 |
+
|
106 |
+
xx, yy = np.meshgrid(np.arange(w), np.arange(h))
|
107 |
+
x = (xx - w/2) * depth_map / f
|
108 |
+
y = (yy - h/2) * depth_map / f
|
109 |
+
z = depth_map
|
110 |
+
|
111 |
+
points = np.vstack((x.flatten(), y.flatten(), z.flatten())).T
|
112 |
+
|
113 |
+
# Apply pose transformation
|
114 |
+
rot = Rotation.from_euler('zyx', [pose[1], pose[0], 0], degrees=True)
|
115 |
+
points = rot.apply(points)
|
116 |
+
|
117 |
+
return points
|
118 |
+
|
119 |
+
def create_mesh_from_pointcloud(points, image):
|
120 |
+
pcd = trimesh.PointCloud(points)
|
121 |
+
scene = pcd.scene()
|
122 |
+
mesh = scene.delaunay_3d.triangulate_pcd(pcd)
|
123 |
+
mesh.visual.vertex_colors = image.resize((mesh.vertices.shape[0], 3))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
return mesh
|
125 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
@app.route('/convert', methods=['POST'])
|
127 |
def convert_image_to_3d():
|
|
|
128 |
if 'image' not in request.files:
|
129 |
return jsonify({"error": "No image provided"}), 400
|
130 |
|
131 |
file = request.files['image']
|
|
|
|
|
|
|
132 |
if not allowed_file(file.filename):
|
133 |
+
return jsonify({"error": "Invalid file type"}), 400
|
134 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
job_id = str(uuid.uuid4())
|
136 |
output_dir = os.path.join(RESULTS_FOLDER, job_id)
|
137 |
os.makedirs(output_dir, exist_ok=True)
|
138 |
+
|
|
|
139 |
filename = secure_filename(file.filename)
|
140 |
filepath = os.path.join(app.config['UPLOAD_FOLDER'], f"{job_id}_{filename}")
|
141 |
file.save(filepath)
|
142 |
+
|
|
|
143 |
processing_jobs[job_id] = {
|
144 |
'status': 'processing',
|
145 |
'progress': 0,
|
146 |
'result_url': None,
|
147 |
+
'error': None
|
|
|
|
|
|
|
148 |
}
|
149 |
+
|
|
|
150 |
def process_image():
|
|
|
|
|
|
|
151 |
try:
|
152 |
+
# Preprocess input image
|
153 |
+
img = preprocess_image(filepath)
|
154 |
+
processing_jobs[job_id]['progress'] = 20
|
155 |
+
|
156 |
+
# Generate novel views
|
157 |
+
views = generate_novel_views(img)
|
158 |
+
processing_jobs[job_id]['progress'] = 40
|
159 |
+
|
160 |
+
# Process each view
|
161 |
+
all_points = []
|
162 |
+
for view_img, pose in views:
|
163 |
+
# Estimate depth
|
164 |
+
depth_result = depth_estimator(view_img)
|
165 |
+
depth_map = np.array(depth_result["depth"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
166 |
|
167 |
+
# Convert to point cloud
|
168 |
+
points = depth_to_pointcloud(depth_map, pose)
|
169 |
+
all_points.append(points)
|
170 |
+
processing_jobs[job_id]['progress'] += 10
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
|
172 |
+
# Combine point clouds
|
173 |
+
combined_points = np.vstack(all_points)
|
174 |
+
processing_jobs[job_id]['progress'] = 80
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
|
176 |
+
# Create mesh
|
177 |
+
mesh = create_mesh_from_pointcloud(combined_points, img)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
|
179 |
+
# Export
|
180 |
+
obj_path = os.path.join(output_dir, "model.obj")
|
181 |
+
mesh.export(obj_path)
|
|
|
|
|
|
|
|
|
|
|
182 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
processing_jobs[job_id]['status'] = 'completed'
|
184 |
+
processing_jobs[job_id]['result_url'] = f"/download/{job_id}"
|
185 |
processing_jobs[job_id]['progress'] = 100
|
186 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
187 |
except Exception as e:
|
|
|
188 |
processing_jobs[job_id]['status'] = 'error'
|
189 |
+
processing_jobs[job_id]['error'] = str(e)
|
190 |
+
finally:
|
|
|
191 |
if os.path.exists(filepath):
|
192 |
os.remove(filepath)
|
193 |
+
gc.collect()
|
194 |
+
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
195 |
|
196 |
+
thread = threading.Thread(target=process_image)
|
197 |
+
thread.start()
|
198 |
+
return jsonify({"job_id": job_id}), 202
|
199 |
+
|
200 |
+
@app.route('/download/<job_id>')
|
201 |
+
def download_model(job_id):
|
202 |
if job_id not in processing_jobs or processing_jobs[job_id]['status'] != 'completed':
|
203 |
+
return jsonify({"error": "Job not complete"}), 404
|
204 |
|
205 |
+
obj_path = os.path.join(RESULTS_FOLDER, job_id, "model.obj")
|
206 |
+
return send_file(obj_path, as_attachment=True)
|
207 |
+
|
208 |
+
@app.route('/progress/<job_id>')
|
209 |
+
def get_progress(job_id):
|
210 |
+
job = processing_jobs.get(job_id, {})
|
211 |
+
return jsonify({
|
212 |
+
'status': job.get('status'),
|
213 |
+
'progress': job.get('progress'),
|
214 |
+
'error': job.get('error')
|
215 |
+
})
|
|
|
|
|
|
|
216 |
|
217 |
if __name__ == '__main__':
|
218 |
+
load_models()
|
219 |
+
app.run(host='0.0.0.0', port=7860)
|
|
|
|
|
|
|
|