# Set comprehensive OpenMP environment variables to prevent ANY fork() errors # This must be done BEFORE importing any libraries that use OpenMP (torch, numpy, rembg, etc.) import os os.environ["OMP_NUM_THREADS"] = "1" # Limit OpenMP to single thread os.environ["MKL_NUM_THREADS"] = "1" # Intel MKL threading os.environ["NUMEXPR_NUM_THREADS"] = "1" # NumExpr threading os.environ["OPENBLAS_NUM_THREADS"] = "1" # OpenBLAS threading os.environ["VECLIB_MAXIMUM_THREADS"] = "1" # Apple vecLib threading os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE" # Avoid Intel MKL errors os.environ["OMP_THREAD_LIMIT"] = "1" # Limit total number of OpenMP threads os.environ["OMP_DISPLAY_ENV"] = "FALSE" # Suppress OpenMP environment display os.environ["KMP_WARNINGS"] = "FALSE" # Suppress KMP warnings os.environ["PYTHONFAULTHANDLER"] = "1" # Better error reporting # Additional settings to prevent subprocess/multiprocessing issues in rembg and other libraries os.environ["TOKENIZERS_PARALLELISM"] = "false" # Disable tokenizer parallelism os.environ["OMP_WAIT_POLICY"] = "PASSIVE" # Use passive waiting os.environ["KMP_INIT_AT_FORK"] = "FALSE" # Don't initialize OpenMP at fork # Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT # except for the third-party components listed below. # Hunyuan 3D does not impose any additional limitations beyond what is outlined # in the repsective licenses of these third-party components. # Users must comply with all terms and conditions of original licenses of these third-party # components and must ensure that the usage of the third party components adheres to # all relevant laws and regulations. # For avoidance of doubts, Hunyuan 3D means the large language models and # their software and algorithms, including trained model weights, parameters (including # optimizer states), machine-learning model code, inference-enabling code, training-enabling code, # fine-tuning enabling code and other elements of the foregoing made publicly available # by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. # Apply torchvision compatibility fix before other imports import sys sys.path.insert(0, './hy3dshape') sys.path.insert(0, './hy3dpaint') pythonpath = sys.executable print(pythonpath) try: from torchvision_fix import apply_fix apply_fix() except ImportError: print("Warning: torchvision_fix module not found, proceeding without compatibility fix") except Exception as e: print(f"Warning: Failed to apply torchvision fix: {e}") import os import random import shutil import subprocess import time from glob import glob from pathlib import Path import base64 import json import threading from typing import Dict, Optional, Any from enum import Enum import gradio as gr import torch import trimesh import uvicorn from fastapi import FastAPI, HTTPException, BackgroundTasks, File, Form, UploadFile from fastapi.staticfiles import StaticFiles from fastapi.responses import JSONResponse from pydantic import BaseModel import uuid import numpy as np from PIL import Image import io from hy3dshape.utils import logger from hy3dpaint.convert_utils import create_glb_with_pbr_materials # API Models class JobStatus(Enum): QUEUED = "queued" PROCESSING = "processing" COMPLETED = "completed" FAILED = "failed" class GenerateOptions(BaseModel): enable_pbr: bool = True should_remesh: bool = True should_texture: bool = True # Critical for 3D model quality class JobInfo: def __init__(self, job_id: str): self.job_id = job_id self.status = JobStatus.QUEUED self.progress = 0 self.stage = "queued" self.start_time = time.time() self.end_time = None self.error_message = None self.model_urls = {} self.images = {} self.options = {} # Global job storage jobs: Dict[str, JobInfo] = {} def create_job() -> str: """Create a new job and return its ID.""" job_id = str(uuid.uuid4()) jobs[job_id] = JobInfo(job_id) return job_id def update_job_status(job_id: str, status: JobStatus, progress: int = None, stage: str = None, error_message: str = None): """Update job status and progress.""" if job_id in jobs: jobs[job_id].status = status if progress is not None: jobs[job_id].progress = progress if stage is not None: jobs[job_id].stage = stage if error_message is not None: jobs[job_id].error_message = error_message if status in [JobStatus.COMPLETED, JobStatus.FAILED]: jobs[job_id].end_time = time.time() def base64_to_pil_image(base64_string: str) -> Image.Image: """Convert base64 string to PIL Image.""" try: # Remove data URL prefix if present if base64_string.startswith('data:image'): base64_string = base64_string.split(',')[1] # Ensure we have valid base64 data # Add padding if necessary missing_padding = len(base64_string) % 4 if missing_padding: base64_string += '=' * (4 - missing_padding) # Decode base64 data try: image_data = base64.b64decode(base64_string) except Exception as e: raise ValueError(f"Failed to decode base64 string: {str(e)}") # Ensure we have valid image data if not image_data or len(image_data) == 0: raise ValueError("Empty image data after base64 decoding") # Open as PIL Image image = Image.open(io.BytesIO(image_data)) # Ensure consistent format - convert to RGBA image = image.convert("RGBA") return image except Exception as e: raise HTTPException(status_code=400, detail=f"Invalid image data: {str(e)}") def process_generation_job(job_id: str, images: Dict[str, Image.Image], options: Dict[str, Any]): """Background task to process generation job.""" global face_reduce_worker, tex_pipeline, HAS_TEXTUREGEN, SAVE_DIR try: update_job_status(job_id, JobStatus.PROCESSING, progress=10, stage="initializing") # Images are already PIL Images pil_images = images # Extract options enable_pbr = options.get("enable_pbr", True) should_remesh = options.get("should_remesh", True) should_texture = options.get("should_texture", True) update_job_status(job_id, JobStatus.PROCESSING, progress=20, stage="preprocessing") # Generate 3D mesh # For non-MV mode, use the front image as the main image, or the first available image main_image = pil_images.get('front') if main_image is None and pil_images: # If no front image, use the first available image main_image = next(iter(pil_images.values())) mesh, main_image, save_folder, stats, seed = _gen_shape( caption=None, image=main_image, mv_image_front=pil_images.get('front'), mv_image_back=pil_images.get('back'), mv_image_left=pil_images.get('left'), mv_image_right=pil_images.get('right'), steps=75, guidance_scale=9.0, seed=1234, octree_resolution=384, check_box_rembg=True, num_chunks=200000, randomize_seed=False, ) update_job_status(job_id, JobStatus.PROCESSING, progress=50, stage="shape_generation") # After mesh generation and before exporting, print and store stats try: number_of_faces = mesh.faces.shape[0] if hasattr(mesh, 'faces') else None number_of_vertices = mesh.vertices.shape[0] if hasattr(mesh, 'vertices') else None logger.info(f"Mesh stats: faces={number_of_faces}, vertices={number_of_vertices}") except Exception as e: logger.warning(f"Failed to log mesh stats: {e}") # Print generation parameters for traceability try: logger.info(f"Generation parameters: seed={seed}, steps=75, octree_resolution=384, guidance_scale=9.0, num_chunks=200000, target_face_count=15000") except Exception as e: logger.warning(f"Failed to log generation parameters: {e}") # Export white mesh white_mesh_path = export_mesh(mesh, save_folder, textured=False, type='obj') # Face reduction mesh = face_reduce_worker(mesh) reduced_mesh_path = export_mesh(mesh, save_folder, textured=False, type='obj') update_job_status(job_id, JobStatus.PROCESSING, progress=70, stage="face_reduction") # Texture generation if enabled textured_mesh_path = None if should_texture: if HAS_TEXTUREGEN: try: text_path = os.path.join(save_folder, 'textured_mesh.obj') # Use GPU function for texture generation with lazy initialization try: logger.info(f"Starting texture generation for job {job_id}") # Count available images to adapt texture generation settings num_images = len(pil_images) logger.info(f"Job {job_id} has {num_images} images available") except Exception as e: logger.warning(f"Failed to log texture generation start: {e}") num_images = len(pil_images) if pil_images else 1 # Try texture generation with adaptive settings based on available images textured_mesh_path = generate_texture_lazy_adaptive( mesh_path=reduced_mesh_path, image_path=main_image, output_mesh_path=text_path, num_available_images=num_images ) if textured_mesh_path and os.path.exists(textured_mesh_path): try: logger.info(f"Texture generation completed for job {job_id}") except Exception as e: logger.warning(f"Failed to log texture completion: {e}") # Convert to GLB glb_path_textured = os.path.join(save_folder, 'textured_mesh.glb') quick_convert_with_obj2gltf(textured_mesh_path, glb_path_textured) textured_mesh_path = glb_path_textured else: try: logger.warning(f"Texture generation returned None or file doesn't exist for job {job_id}") except Exception as e: logger.warning(f"Failed to log texture warning: {e}") textured_mesh_path = None except Exception as e: logger.error(f"Texture generation failed for job {job_id}: {e}") # Continue without texture - user will get the white mesh textured_mesh_path = None else: try: logger.warning(f"Texture generation requested for job {job_id} but texture pipeline is not available") except Exception as e: logger.warning(f"Failed to log texture unavailable warning: {e}") update_job_status(job_id, JobStatus.PROCESSING, progress=75, stage="texture_generation_unavailable", message="Texture generation is not available - returning mesh without texture") update_job_status(job_id, JobStatus.PROCESSING, progress=90, stage="finalizing") # Prepare model URLs model_urls = {} if textured_mesh_path and os.path.exists(textured_mesh_path): model_urls["glb"] = f"/static/{os.path.relpath(textured_mesh_path, SAVE_DIR)}" else: # Fallback to white mesh white_glb_path = export_mesh(mesh, save_folder, textured=False, type='glb') model_urls["glb"] = f"/static/{os.path.relpath(white_glb_path, SAVE_DIR)}" # Add mesh stats to API output model_urls["number_of_faces"] = number_of_faces model_urls["number_of_vertices"] = number_of_vertices # Update job with results jobs[job_id].model_urls = model_urls update_job_status(job_id, JobStatus.COMPLETED, progress=100, stage="completed") except Exception as e: logger.error(f"Job {job_id} failed: {e}") update_job_status(job_id, JobStatus.FAILED, stage="failed", error_message=str(e)) MAX_SEED = 1e7 ENV = "Huggingface" # "Huggingface" if ENV == 'Huggingface': """ Setup environment for running on Huggingface platform. This block performs the following: - Changes directory to the differentiable renderer folder and runs a shell script to compile the mesh painter. - Installs a custom rasterizer wheel package via pip. Note: This setup assumes the script is running in the Huggingface environment with the specified directory structure. """ import os, spaces, subprocess, sys, shlex from spaces import zero def install_cuda_toolkit(): # CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run" CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/12.2.0/local_installers/cuda_12.2.0_535.54.03_linux.run" CUDA_TOOLKIT_FILE = "/tmp/%s" % os.path.basename(CUDA_TOOLKIT_URL) subprocess.call(["wget", "-q", CUDA_TOOLKIT_URL, "-O", CUDA_TOOLKIT_FILE]) subprocess.call(["chmod", "+x", CUDA_TOOLKIT_FILE]) subprocess.call([CUDA_TOOLKIT_FILE, "--silent", "--toolkit"]) os.environ["CUDA_HOME"] = "/usr/local/cuda" os.environ["PATH"] = "%s/bin:%s" % (os.environ["CUDA_HOME"], os.environ["PATH"]) os.environ["LD_LIBRARY_PATH"] = "%s/lib:%s" % ( os.environ["CUDA_HOME"], "" if "LD_LIBRARY_PATH" not in os.environ else os.environ["LD_LIBRARY_PATH"], ) # Fix: arch_list[-1] += '+PTX'; IndexError: list index out of range os.environ["TORCH_CUDA_ARCH_LIST"] = "8.0;8.6" def prepare_env(): # print('install custom') # os.system(f"cd /home/user/app/hy3dpaint/custom_rasterizer && {pythonpath} -m pip install -e .") # os.system(f"cd /home/user/app/hy3dpaint/packages/custom_rasterizer && pip install -e .") subprocess.run(shlex.split("pip install custom_rasterizer-0.1-cp310-cp310-linux_x86_64.whl"), check=True) print("cd /home/user/app/hy3dpaint/differentiable_renderer/ && bash compile_mesh_painter.sh") os.system("cd /home/user/app/hy3dpaint/DifferentiableRenderer && bash compile_mesh_painter.sh") print("Downloading RealESRGAN model for texture enhancement...") os.makedirs("/home/user/app/hy3dpaint/ckpt", exist_ok=True) os.system("wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P /home/user/app/hy3dpaint/ckpt") def check(): import custom_rasterizer print(type(custom_rasterizer)) print(dir(custom_rasterizer)) print(getattr(custom_rasterizer, '__file__', None)) package_dir = None if hasattr(custom_rasterizer, '__file__') and custom_rasterizer.__file__: package_dir = os.path.dirname(custom_rasterizer.__file__) elif hasattr(custom_rasterizer, '__path__'): package_dir = list(custom_rasterizer.__path__)[0] else: raise RuntimeError("Cannot determine package path") print(package_dir) for root, dirs, files in os.walk(package_dir): level = root.replace(package_dir, '').count(os.sep) indent = ' ' * 4 * level print(f"{indent}{os.path.basename(root)}/") subindent = ' ' * 4 * (level + 1) for f in files: print(f"{subindent}{f}") # print(torch.__version__) # install_cuda_toolkit() print(torch.__version__) prepare_env() check() else: """ Define a dummy `spaces` module with a GPU decorator class for local environment. The GPU decorator is a no-op that simply returns the decorated function unchanged. This allows code that uses the `spaces.GPU` decorator to run without modification locally. """ class spaces: class GPU: def __init__(self, duration=60): self.duration = duration def __call__(self, func): return func def get_example_img_list(): """ Load and return a sorted list of example image file paths. Searches recursively for PNG images under the './assets/example_images/' directory. Returns: list[str]: Sorted list of file paths to example PNG images. """ print('Loading example img list ...') return sorted(glob('./assets/example_images/**/*.png', recursive=True)) def get_example_txt_list(): """ Load and return a list of example text prompts. Reads lines from the './assets/example_prompts.txt' file, stripping whitespace. Returns: list[str]: List of example text prompts. """ print('Loading example txt list ...') txt_list = list() for line in open('./assets/example_prompts.txt', encoding='utf-8'): txt_list.append(line.strip()) return txt_list def gen_save_folder(max_size=200): """ Generate a new save folder inside SAVE_DIR, maintaining a maximum number of folders. If the number of existing folders in SAVE_DIR exceeds `max_size`, the oldest folder is removed. Args: max_size (int, optional): Maximum number of folders to keep in SAVE_DIR. Defaults to 200. Returns: str: Path to the newly created save folder. """ os.makedirs(SAVE_DIR, exist_ok=True) dirs = [f for f in Path(SAVE_DIR).iterdir() if f.is_dir()] if len(dirs) >= max_size: oldest_dir = min(dirs, key=lambda x: x.stat().st_ctime) shutil.rmtree(oldest_dir) print(f"Removed the oldest folder: {oldest_dir}") new_folder = os.path.join(SAVE_DIR, str(uuid.uuid4())) os.makedirs(new_folder, exist_ok=True) print(f"Created new folder: {new_folder}") return new_folder # Removed complex PBR conversion functions - using simple trimesh-based conversion def export_mesh(mesh, save_folder, textured=False, type='glb'): """ Export a mesh to a file in the specified folder, optionally including textures. Args: mesh (trimesh.Trimesh): The mesh object to export. save_folder (str): Directory path where the mesh file will be saved. textured (bool, optional): Whether to include textures/normals in the export. Defaults to False. type (str, optional): File format to export ('glb' or 'obj' supported). Defaults to 'glb'. Returns: str: The full path to the exported mesh file. """ if textured: path = os.path.join(save_folder, f'textured_mesh.{type}') else: path = os.path.join(save_folder, f'white_mesh.{type}') if type not in ['glb', 'obj']: mesh.export(path) else: mesh.export(path, include_normals=textured) return path def quick_convert_with_obj2gltf(obj_path: str, glb_path: str) -> bool: # 执行转换 textures = { 'albedo': obj_path.replace('.obj', '.jpg'), 'metallic': obj_path.replace('.obj', '_metallic.jpg'), 'roughness': obj_path.replace('.obj', '_roughness.jpg') } create_glb_with_pbr_materials(obj_path, textures, glb_path) def randomize_seed_fn(seed: int, randomize_seed: bool) -> int: if randomize_seed: seed = random.randint(0, MAX_SEED) return seed def build_model_viewer_html(save_folder, height=660, width=790, textured=False): # Remove first folder from path to make relative path if textured: related_path = f"./textured_mesh.glb" template_name = './assets/modelviewer-textured-template.html' output_html_path = os.path.join(save_folder, f'textured_mesh.html') else: related_path = f"./white_mesh.glb" template_name = './assets/modelviewer-template.html' output_html_path = os.path.join(save_folder, f'white_mesh.html') offset = 50 if textured else 10 with open(os.path.join(CURRENT_DIR, template_name), 'r', encoding='utf-8') as f: template_html = f.read() with open(output_html_path, 'w', encoding='utf-8') as f: template_html = template_html.replace('#height#', f'{height - offset}') template_html = template_html.replace('#width#', f'{width}') template_html = template_html.replace('#src#', f'{related_path}/') f.write(template_html) rel_path = os.path.relpath(output_html_path, SAVE_DIR) iframe_tag = f'' print(f'Find html file {output_html_path}, \ {os.path.exists(output_html_path)}, relative HTML path is /static/{rel_path}') return f"""
{iframe_tag}
""" @spaces.GPU(duration=60) def _gen_shape( caption=None, image=None, mv_image_front=None, mv_image_back=None, mv_image_left=None, mv_image_right=None, steps=75, guidance_scale=9.0, seed=1234, octree_resolution=384, check_box_rembg=False, num_chunks=200000, randomize_seed: bool = False, ): # Check if we're using multi-view mode based on inputs # Only consider non-None AND non-empty images for multi-view detection using_multiview = ((mv_image_front is not None and mv_image_front is not False) or (mv_image_back is not None and mv_image_back is not False) or (mv_image_left is not None and mv_image_left is not False) or (mv_image_right is not None and mv_image_right is not False)) # Single image mode validation if not using_multiview and image is None and caption is None: raise gr.Error("Please provide either a caption or an image.") # Multi-view mode validation and processing if using_multiview: # Check if any valid images were provided has_valid_image = False image = {} if mv_image_front is not None: image['front'] = mv_image_front has_valid_image = True if mv_image_back is not None: image['back'] = mv_image_back has_valid_image = True if mv_image_left is not None: image['left'] = mv_image_left has_valid_image = True if mv_image_right is not None: image['right'] = mv_image_right has_valid_image = True if not has_valid_image: raise gr.Error("Please provide at least one view image.") seed = int(randomize_seed_fn(seed, randomize_seed)) octree_resolution = int(octree_resolution) if caption: print('prompt is', caption) save_folder = gen_save_folder() stats = { 'model': { 'shapegen': f'{args.model_path}/{args.subfolder}', 'texgen': f'{args.texgen_model_path}', }, 'params': { 'caption': caption, 'steps': steps, 'guidance_scale': guidance_scale, 'seed': seed, 'octree_resolution': octree_resolution, 'check_box_rembg': check_box_rembg, 'num_chunks': num_chunks, } } time_meta = {} if image is None: start_time = time.time() try: image = t2i_worker(caption) except Exception as e: raise gr.Error(f"Text to 3D is disable. \ Please enable it by `python gradio_app.py --enable_t23d`.") time_meta['text2image'] = time.time() - start_time # remove disk io to make responding faster, uncomment at your will. # image.save(os.path.join(save_folder, 'input.png')) # Process images based on whether we're using multi-view mode start_time = time.time() if isinstance(image, dict): # Multi-view mode for k, v in image.items(): if v is not None and (check_box_rembg or v.mode == "RGB"): try: img = rmbg_worker(v.convert('RGB')) image[k] = img except Exception as e: print(f"Error processing {k} view: {e}") # Keep the original image if background removal fails pass else: # Single image mode if image is not None and (check_box_rembg or image.mode == "RGB"): try: image = rmbg_worker(image.convert('RGB')) except Exception as e: print(f"Error removing background: {e}") # Keep the original image if background removal fails pass time_meta['remove background'] = time.time() - start_time # remove disk io to make responding faster, uncomment at your will. # image.save(os.path.join(save_folder, 'rembg.png')) # image to white model start_time = time.time() generator = torch.Generator() generator = generator.manual_seed(int(seed)) outputs = i23d_worker( image=image, num_inference_steps=steps, guidance_scale=guidance_scale, generator=generator, octree_resolution=octree_resolution, num_chunks=num_chunks, output_type='mesh' ) time_meta['shape generation'] = time.time() - start_time logger.info("---Shape generation takes %s seconds ---" % (time.time() - start_time)) tmp_start = time.time() mesh = export_to_trimesh(outputs)[0] time_meta['export to trimesh'] = time.time() - tmp_start stats['number_of_faces'] = mesh.faces.shape[0] stats['number_of_vertices'] = mesh.vertices.shape[0] stats['time'] = time_meta # Select the main image for display based on what's available if isinstance(image, dict) and 'front' in image: main_image = image['front'] # Use front view as main image in multi-view mode else: main_image = image # Use the single image in single-image mode return mesh, main_image, save_folder, stats, seed @spaces.GPU(duration=180) def generation_all( caption=None, image=None, mv_image_front=None, mv_image_back=None, mv_image_left=None, mv_image_right=None, steps=75, guidance_scale=9.0, seed=1234, octree_resolution=384, check_box_rembg=False, num_chunks=200000, randomize_seed: bool = False, ): start_time_0 = time.time() mesh, image, save_folder, stats, seed = _gen_shape( caption, image, mv_image_front=mv_image_front, mv_image_back=mv_image_back, mv_image_left=mv_image_left, mv_image_right=mv_image_right, steps=steps, guidance_scale=guidance_scale, seed=seed, octree_resolution=octree_resolution, check_box_rembg=check_box_rembg, num_chunks=num_chunks, randomize_seed=randomize_seed, ) path = export_mesh(mesh, save_folder, textured=False) print(path) print('='*40) # tmp_time = time.time() # mesh = floater_remove_worker(mesh) # mesh = degenerate_face_remove_worker(mesh) # logger.info("---Postprocessing takes %s seconds ---" % (time.time() - tmp_time)) # stats['time']['postprocessing'] = time.time() - tmp_time tmp_time = time.time() mesh = face_reduce_worker(mesh) # path = export_mesh(mesh, save_folder, textured=False, type='glb') path = export_mesh(mesh, save_folder, textured=False, type='obj') # 这样操作也会 core dump logger.info("---Face Reduction takes %s seconds ---" % (time.time() - tmp_time)) stats['time']['face reduction'] = time.time() - tmp_time tmp_time = time.time() text_path = os.path.join(save_folder, f'textured_mesh.obj') path_textured = tex_pipeline(mesh_path=path, image_path=image, output_mesh_path=text_path, save_glb=False) logger.info("---Texture Generation takes %s seconds ---" % (time.time() - tmp_time)) stats['time']['texture generation'] = time.time() - tmp_time tmp_time = time.time() # Convert textured OBJ to GLB using obj2gltf with PBR support glb_path_textured = os.path.join(save_folder, 'textured_mesh.glb') conversion_success = quick_convert_with_obj2gltf(path_textured, glb_path_textured) logger.info("---Convert textured OBJ to GLB takes %s seconds ---" % (time.time() - tmp_time)) stats['time']['convert textured OBJ to GLB'] = time.time() - tmp_time stats['time']['total'] = time.time() - start_time_0 model_viewer_html_textured = build_model_viewer_html(save_folder, height=HTML_HEIGHT, width=HTML_WIDTH, textured=True) if args.low_vram_mode: torch.cuda.empty_cache() return ( gr.update(value=path), gr.update(value=glb_path_textured), model_viewer_html_textured, stats, seed, ) @spaces.GPU(duration=180) def generate_texture_lazy_adaptive(mesh_path, image_path, output_mesh_path, num_available_images=1): """Generate texture for a mesh with adaptive settings based on available images.""" try: # Lazy initialization of texture pipeline inside GPU function from hy3dpaint.textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig # Use the same high-quality settings as the Gradio app max_views = 9 resolution = 768 logger.info(f"Using high quality settings: {max_views} views, {resolution} resolution (same as Gradio app)") conf = Hunyuan3DPaintConfig(max_num_view=max_views, resolution=resolution) conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth" conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml" conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr" # Initialize texture pipeline inside GPU function local_tex_pipeline = Hunyuan3DPaintPipeline(conf) # Generate texture with timeout handling try: textured_mesh_path = local_tex_pipeline( mesh_path=mesh_path, image_path=image_path, output_mesh_path=output_mesh_path, save_glb=False ) return textured_mesh_path except Exception as texture_error: logger.error(f"Texture generation pipeline failed: {texture_error}") # Try with medium quality settings as fallback try: fallback_views = 4 fallback_resolution = 384 logger.info(f"Trying fallback settings: {fallback_views} views, {fallback_resolution} resolution") conf = Hunyuan3DPaintConfig(max_num_view=fallback_views, resolution=fallback_resolution) conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth" conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml" conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr" local_tex_pipeline = Hunyuan3DPaintPipeline(conf) textured_mesh_path = local_tex_pipeline( mesh_path=mesh_path, image_path=image_path, output_mesh_path=output_mesh_path, save_glb=False ) return textured_mesh_path except Exception as fallback_error: logger.error(f"Fallback texture generation also failed: {fallback_error}") return None except Exception as e: logger.error(f"Texture generation initialization failed: {e}") return None @spaces.GPU(duration=300) def generate_texture_lazy(mesh_path, image_path, output_mesh_path): """Generate texture for a mesh using lazy initialization to avoid CUDA startup issues.""" try: # Lazy initialization of texture pipeline inside GPU function from hy3dpaint.textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig # Use fast settings optimized for 5-minute Hugging Face Spaces limit conf = Hunyuan3DPaintConfig(max_num_view=2, resolution=256) conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth" conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml" conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr" # Initialize texture pipeline inside GPU function local_tex_pipeline = Hunyuan3DPaintPipeline(conf) # Generate texture with timeout handling try: textured_mesh_path = local_tex_pipeline( mesh_path=mesh_path, image_path=image_path, output_mesh_path=output_mesh_path, save_glb=False ) return textured_mesh_path except Exception as texture_error: logger.error(f"Texture generation pipeline failed: {texture_error}") # Try with even faster settings as fallback try: conf = Hunyuan3DPaintConfig(max_num_view=1, resolution=128) conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth" conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml" conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr" local_tex_pipeline = Hunyuan3DPaintPipeline(conf) textured_mesh_path = local_tex_pipeline( mesh_path=mesh_path, image_path=image_path, output_mesh_path=output_mesh_path, save_glb=False ) return textured_mesh_path except Exception as fallback_error: logger.error(f"Fallback texture generation also failed: {fallback_error}") return None except Exception as e: logger.error(f"Texture generation initialization failed: {e}") return None @spaces.GPU(duration=60) def shape_generation( caption=None, image=None, mv_image_front=None, mv_image_back=None, mv_image_left=None, mv_image_right=None, steps=75, guidance_scale=9.0, seed=1234, octree_resolution=384, check_box_rembg=False, num_chunks=200000, randomize_seed: bool = False, ): start_time_0 = time.time() mesh, image, save_folder, stats, seed = _gen_shape( caption, image, mv_image_front=mv_image_front, mv_image_back=mv_image_back, mv_image_left=mv_image_left, mv_image_right=mv_image_right, steps=steps, guidance_scale=guidance_scale, seed=seed, octree_resolution=octree_resolution, check_box_rembg=check_box_rembg, num_chunks=num_chunks, randomize_seed=randomize_seed, ) stats['time']['total'] = time.time() - start_time_0 mesh.metadata['extras'] = stats path = export_mesh(mesh, save_folder, textured=False) model_viewer_html = build_model_viewer_html(save_folder, height=HTML_HEIGHT, width=HTML_WIDTH) if args.low_vram_mode: torch.cuda.empty_cache() return ( gr.update(value=path), model_viewer_html, stats, seed, ) def build_app(): title = 'Hunyuan3D-2: High Resolution Textured 3D Assets Generation' if MV_MODE: title = 'Hunyuan3D-2mv: Image to 3D Generation with 1-4 Views' if 'mini' in args.subfolder: title = 'Hunyuan3D-2mini: Strong 0.6B Image to Shape Generator' title = 'Hunyuan-3D-2.1' if TURBO_MODE: title = title.replace(':', '-Turbo: Fast ') title_html = f"""
{title}
Tencent Hunyuan3D Team
""" custom_css = """ .app.svelte-wpkpf6.svelte-wpkpf6:not(.fill_width) { max-width: 1480px; } .mv-image button .wrap { font-size: 10px; } .mv-image .icon-wrap { width: 20px; } """ with gr.Blocks(theme=gr.themes.Base(), title='Hunyuan-3D-2.1', analytics_enabled=False, css=custom_css) as demo: gr.HTML(title_html) with gr.Row(): with gr.Column(scale=3): with gr.Tabs(selected='tab_img_prompt') as tabs_prompt: with gr.Tab('Image Prompt', id='tab_img_prompt', visible=True) as tab_ip: image = gr.Image(label='Image', type='pil', image_mode='RGBA', height=290) caption = gr.State(None) # with gr.Tab('Text Prompt', id='tab_txt_prompt', visible=HAS_T2I and not MV_MODE) as tab_tp: # caption = gr.Textbox(label='Text Prompt', # placeholder='HunyuanDiT will be used to generate image.', # info='Example: A 3D model of a cute cat, white background.') with gr.Tab('MultiView Prompt', visible=True) as tab_mv: # gr.Label('Please upload at least one front image.') with gr.Row(): mv_image_front = gr.Image(label='Front', type='pil', image_mode='RGBA', height=140, min_width=100, elem_classes='mv-image') mv_image_back = gr.Image(label='Back', type='pil', image_mode='RGBA', height=140, min_width=100, elem_classes='mv-image') with gr.Row(): mv_image_left = gr.Image(label='Left', type='pil', image_mode='RGBA', height=140, min_width=100, elem_classes='mv-image') mv_image_right = gr.Image(label='Right', type='pil', image_mode='RGBA', height=140, min_width=100, elem_classes='mv-image') with gr.Row(): btn = gr.Button(value='Gen Shape', variant='primary', min_width=100) btn_all = gr.Button(value='Gen Textured Shape', variant='primary', visible=True, # Force visible for now, was: HAS_TEXTUREGEN min_width=100) with gr.Group(): file_out = gr.File(label="File", visible=False) file_out2 = gr.File(label="File", visible=False) with gr.Tabs(selected='tab_options' if TURBO_MODE else 'tab_export'): with gr.Tab("Options", id='tab_options', visible=TURBO_MODE): gen_mode = gr.Radio( label='Generation Mode', info='Recommendation: Turbo for most cases, \ Fast for very complex cases, Standard seldom use.', choices=['Turbo', 'Fast', 'Standard'], value='Turbo') decode_mode = gr.Radio( label='Decoding Mode', info='The resolution for exporting mesh from generated vectset', choices=['Low', 'Standard', 'High'], value='Standard') with gr.Tab('Advanced Options', id='tab_advanced_options'): with gr.Row(): check_box_rembg = gr.Checkbox( value=True, label='Remove Background', min_width=100) randomize_seed = gr.Checkbox( label="Randomize seed", value=True, min_width=100) seed = gr.Slider( label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=1234, min_width=100, ) with gr.Row(): num_steps = gr.Slider(maximum=100, minimum=1, value=5 if 'turbo' in args.subfolder else 75, step=1, label='Inference Steps') octree_resolution = gr.Slider(maximum=512, minimum=16, value=384, label='Octree Resolution') with gr.Row(): cfg_scale = gr.Number(value=9.0, label='Guidance Scale', min_width=100) num_chunks = gr.Slider(maximum=5000000, minimum=1000, value=8000, label='Number of Chunks', min_width=100) with gr.Tab("Export", id='tab_export'): with gr.Row(): file_type = gr.Dropdown(label='File Type', choices=SUPPORTED_FORMATS, value='glb', min_width=100) reduce_face = gr.Checkbox(label='Simplify Mesh', value=False, min_width=100) export_texture = gr.Checkbox(label='Include Texture', value=False, visible=False, min_width=100) target_face_num = gr.Slider(maximum=1000000, minimum=100, value=15000, label='Target Face Number') with gr.Row(): confirm_export = gr.Button(value="Transform", min_width=100) file_export = gr.DownloadButton(label="Download", variant='primary', interactive=False, min_width=100) with gr.Column(scale=6): with gr.Tabs(selected='gen_mesh_panel') as tabs_output: with gr.Tab('Generated Mesh', id='gen_mesh_panel'): html_gen_mesh = gr.HTML(HTML_OUTPUT_PLACEHOLDER, label='Output') with gr.Tab('Exporting Mesh', id='export_mesh_panel'): html_export_mesh = gr.HTML(HTML_OUTPUT_PLACEHOLDER, label='Output') with gr.Tab('Mesh Statistic', id='stats_panel'): stats = gr.Json({}, label='Mesh Stats') with gr.Column(scale=3 if MV_MODE else 2): with gr.Tabs(selected='tab_img_gallery') as gallery: with gr.Tab('Image to 3D Gallery', id='tab_img_gallery', visible=not MV_MODE) as tab_gi: with gr.Row(): gr.Examples(examples=example_is, inputs=[image], label=None, examples_per_page=18) tab_ip.select(fn=lambda: gr.update(selected='tab_img_gallery'), outputs=gallery) #if HAS_T2I: # tab_tp.select(fn=lambda: gr.update(selected='tab_txt_gallery'), outputs=gallery) btn.click( shape_generation, inputs=[ caption, image, mv_image_front, mv_image_back, mv_image_left, mv_image_right, num_steps, cfg_scale, seed, octree_resolution, check_box_rembg, num_chunks, randomize_seed, ], outputs=[file_out, html_gen_mesh, stats, seed] ).then( lambda: (gr.update(visible=False, value=False), gr.update(interactive=True), gr.update(interactive=True), gr.update(interactive=False)), outputs=[export_texture, reduce_face, confirm_export, file_export], ).then( lambda: gr.update(selected='gen_mesh_panel'), outputs=[tabs_output], ) btn_all.click( generation_all, inputs=[ caption, image, mv_image_front, mv_image_back, mv_image_left, mv_image_right, num_steps, cfg_scale, seed, octree_resolution, check_box_rembg, num_chunks, randomize_seed, ], outputs=[file_out, file_out2, html_gen_mesh, stats, seed] ).then( lambda: (gr.update(visible=True, value=True), gr.update(interactive=False), gr.update(interactive=True), gr.update(interactive=False)), outputs=[export_texture, reduce_face, confirm_export, file_export], ).then( lambda: gr.update(selected='gen_mesh_panel'), outputs=[tabs_output], ) def on_gen_mode_change(value): if value == 'Turbo': return gr.update(value=5) elif value == 'Fast': return gr.update(value=10) else: return gr.update(value=30) gen_mode.change(on_gen_mode_change, inputs=[gen_mode], outputs=[num_steps]) def on_decode_mode_change(value): if value == 'Low': return gr.update(value=196) elif value == 'Standard': return gr.update(value=256) else: return gr.update(value=384) decode_mode.change(on_decode_mode_change, inputs=[decode_mode], outputs=[octree_resolution]) def on_export_click(file_out, file_out2, file_type, reduce_face, export_texture, target_face_num): if file_out is None: raise gr.Error('Please generate a mesh first.') print(f'exporting {file_out}') print(f'reduce face to {target_face_num}') if export_texture: mesh = trimesh.load(file_out2) save_folder = gen_save_folder() path = export_mesh(mesh, save_folder, textured=True, type=file_type) # for preview save_folder = gen_save_folder() _ = export_mesh(mesh, save_folder, textured=True) model_viewer_html = build_model_viewer_html(save_folder, height=HTML_HEIGHT, width=HTML_WIDTH, textured=True) else: mesh = trimesh.load(file_out) mesh = floater_remove_worker(mesh) mesh = degenerate_face_remove_worker(mesh) if reduce_face: mesh = face_reduce_worker(mesh, target_face_num) save_folder = gen_save_folder() path = export_mesh(mesh, save_folder, textured=False, type=file_type) # for preview save_folder = gen_save_folder() _ = export_mesh(mesh, save_folder, textured=False) model_viewer_html = build_model_viewer_html(save_folder, height=HTML_HEIGHT, width=HTML_WIDTH, textured=False) print(f'export to {path}') return model_viewer_html, gr.update(value=path, interactive=True) confirm_export.click( lambda: gr.update(selected='export_mesh_panel'), outputs=[tabs_output], ).then( on_export_click, inputs=[file_out, file_out2, file_type, reduce_face, export_texture, target_face_num], outputs=[html_export_mesh, file_export] ) return demo if __name__ == '__main__': import argparse parser = argparse.ArgumentParser() parser.add_argument("--model_path", type=str, default='tencent/Hunyuan3D-2.1') parser.add_argument("--subfolder", type=str, default='hunyuan3d-dit-v2-1') parser.add_argument("--texgen_model_path", type=str, default='tencent/Hunyuan3D-2.1') parser.add_argument('--port', type=int, default=7860) parser.add_argument('--host', type=str, default='0.0.0.0') parser.add_argument('--device', type=str, default='cuda') parser.add_argument('--mc_algo', type=str, default='mc') parser.add_argument('--cache-path', type=str, default='/root/save_dir') parser.add_argument('--enable_t23d', action='store_true') parser.add_argument('--disable_tex', action='store_true') parser.add_argument('--enable_flashvdm', action='store_true') parser.add_argument('--compile', action='store_true') parser.add_argument('--low_vram_mode', action='store_true') args = parser.parse_args() args.enable_flashvdm = False SAVE_DIR = args.cache_path os.makedirs(SAVE_DIR, exist_ok=True) CURRENT_DIR = os.path.dirname(os.path.abspath(__file__)) MV_MODE = True # Force multi-view mode to be enabled TURBO_MODE = 'turbo' in args.subfolder HTML_HEIGHT = 690 if MV_MODE else 650 HTML_WIDTH = 500 HTML_OUTPUT_PLACEHOLDER = f"""

Welcome to Hunyuan3D!

No mesh here.

""" INPUT_MESH_HTML = """
""" example_is = get_example_img_list() example_ts = get_example_txt_list() SUPPORTED_FORMATS = ['glb', 'obj', 'ply', 'stl'] HAS_TEXTUREGEN = False if not args.disable_tex: try: print("Initializing texture generation pipeline...") # Apply torchvision fix before importing basicsr/RealESRGAN print("Applying torchvision compatibility fix for texture generation...") try: from torchvision_fix import apply_fix fix_result = apply_fix() if not fix_result: print("Warning: Torchvision fix may not have been applied successfully") else: print("Torchvision fix applied successfully") except ImportError as ie: print(f"Warning: Could not import torchvision_fix: {ie}") except Exception as fix_error: print(f"Warning: Failed to apply torchvision fix: {fix_error}") import traceback traceback.print_exc() # Import texture generation components print("Importing texture generation components...") try: from hy3dpaint.textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig print("Successfully imported texture generation components") except ImportError as ie: print(f"Failed to import texture generation components: {ie}") raise except Exception as e: print(f"Error importing texture generation components: {e}") import traceback traceback.print_exc() raise # Configure texture pipeline print("Configuring texture pipeline...") conf = Hunyuan3DPaintConfig(max_num_view=9, resolution=768) conf.realesrgan_ckpt_path = "hy3dpaint/ckpt/RealESRGAN_x4plus.pth" conf.multiview_cfg_path = "hy3dpaint/cfgs/hunyuan-paint-pbr.yaml" conf.custom_pipeline = "hy3dpaint/hunyuanpaintpbr" # Initialize texture pipeline print("Initializing texture pipeline...") tex_pipeline = Hunyuan3DPaintPipeline(conf) print("Texture pipeline initialized successfully") # Not help much, ignore for now. # if args.compile: # texgen_worker.models['delight_model'].pipeline.unet.compile() # texgen_worker.models['delight_model'].pipeline.vae.compile() # texgen_worker.models['multiview_model'].pipeline.unet.compile() # texgen_worker.models['multiview_model'].pipeline.vae.compile() HAS_TEXTUREGEN = True print("Texture generation is ENABLED - Gen Textured Shape button will be visible") except Exception as e: print(f"Error loading texture generator: {e}") print("Failed to load texture generator.") print('Please try to install requirements by following README.md') import traceback traceback.print_exc() HAS_TEXTUREGEN = False print("Texture generation is DISABLED - Gen Textured Shape button will be hidden") # HAS_T2I = True # if args.enable_t23d: # from hy3dgen.text2image import HunyuanDiTPipeline # t2i_worker = HunyuanDiTPipeline('Tencent-Hunyuan/HunyuanDiT-v1.1-Diffusers-Distilled') # HAS_T2I = True from hy3dshape import FaceReducer, FloaterRemover, DegenerateFaceRemover, MeshSimplifier, \ Hunyuan3DDiTFlowMatchingPipeline from hy3dshape.pipelines import export_to_trimesh from hy3dshape.rembg import BackgroundRemover rmbg_worker = BackgroundRemover() i23d_worker = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained( args.model_path, subfolder=args.subfolder, use_safetensors=False, device=args.device, ) if args.enable_flashvdm: mc_algo = 'mc' if args.device in ['cpu', 'mps'] else args.mc_algo i23d_worker.enable_flashvdm(mc_algo=mc_algo) if args.compile: i23d_worker.compile() floater_remove_worker = FloaterRemover() degenerate_face_remove_worker = DegenerateFaceRemover() face_reduce_worker = FaceReducer() # https://discuss.huggingface.co/t/how-to-serve-an-html-file/33921/2 # create a FastAPI app app = FastAPI() # API Endpoints @app.post("/api/generate") async def generate_3d_model( front: UploadFile = File(None), back: UploadFile = File(None), left: UploadFile = File(None), right: UploadFile = File(None), options: str = Form("{}"), background_tasks: BackgroundTasks = BackgroundTasks() ): """Generate 3D model from images using multipart/form-data.""" try: # Parse options options_dict = json.loads(options) if options else {} generate_options = { "enable_pbr": options_dict.get("enable_pbr", True), "should_remesh": options_dict.get("should_remesh", True), "should_texture": options_dict.get("should_texture", True) } # Process uploaded files images = {} # Validate input - at least one image is required if not front: raise HTTPException(status_code=400, detail="Front image is required") # Process each uploaded file for view, file in { "front": front, "back": back, "left": left, "right": right }.items(): if file and file.filename: # Read file content contents = await file.read() # Convert to PIL Image try: img = Image.open(io.BytesIO(contents)) images[view] = img except Exception as e: raise HTTPException(status_code=400, detail=f"Invalid image for {view}: {str(e)}") # Create job job_id = create_job() # Store job data (store file paths instead of actual images) jobs[job_id].images = {k: "" for k in images.keys()} jobs[job_id].options = generate_options # Start background task background_tasks.add_task( process_generation_job, job_id, images, generate_options ) return JSONResponse({ "job_id": job_id, "status": "queued" }) except Exception as e: raise HTTPException(status_code=500, detail=str(e)) @app.get("/api/status") async def get_job_status(job_id: str): """Get job status and results.""" if job_id not in jobs: raise HTTPException(status_code=404, detail="Job not found") job = jobs[job_id] response = { "status": job.status.value, "progress": job.progress, "stage": job.stage } if job.status == JobStatus.COMPLETED: # Transform relative URLs to absolute URLs absolute_model_urls = {} # Determine base URL based on environment if ENV == 'Huggingface': base_url = "https://asimfayaz-hunyuan3d-2-1.hf.space" # TODO: Refactor this URL else: # For local development host_for_url = "localhost" if args.host == "0.0.0.0" else args.host base_url = f"http://{host_for_url}:{args.port}" for key, relative_url in job.model_urls.items(): # Handle both string and non-string values if isinstance(relative_url, str): if relative_url.startswith('/'): absolute_model_urls[key] = f"{base_url}{relative_url}" else: absolute_model_urls[key] = relative_url else: # For non-string values (like integers), keep as is absolute_model_urls[key] = relative_url response["model_urls"] = absolute_model_urls elif job.status == JobStatus.FAILED: response["error"] = job.error_message return JSONResponse(response) @app.get("/api/health") async def health_check(): """Health check endpoint.""" return JSONResponse({ "status": "ok", "version": "2.1" }) # create a static directory to store the static files static_dir = Path(SAVE_DIR).absolute() static_dir.mkdir(parents=True, exist_ok=True) app.mount("/static", StaticFiles(directory=static_dir, html=True), name="static") shutil.copytree('./assets/env_maps', os.path.join(static_dir, 'env_maps'), dirs_exist_ok=True) if args.low_vram_mode: torch.cuda.empty_cache() demo = build_app() app = gr.mount_gradio_app(app, demo, path="/") if ENV == 'Huggingface': # for Zerogpu from spaces import zero zero.startup() uvicorn.run(app, host=args.host, port=args.port)