import torch import numpy as np from PIL import Image import pymeshlab as ml from pytorch3d.renderer import TexturesVertex from pytorch3d.structures import Meshes from rembg import new_session, remove import trimesh from typing import List, Tuple import torch.nn.functional as F # Constants NEG_PROMPT = "sketch, sculpture, hand drawing, outline, single color, NSFW, lowres, bad anatomy, bad hands, text, error, missing fingers, yellow sleeves, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, (worst quality:1.4), (low quality:1.4)" # CUDA Configuration CUDA_PROVIDERS = [ ('CUDAExecutionProvider', { 'device_id': 0, 'arena_extend_strategy': 'kSameAsRequested', 'gpu_mem_limit': 8 * 1024 * 1024 * 1024, 'cudnn_conv_algo_search': 'HEURISTIC', }) ] # Initialize rembg session rembg_session = new_session(providers=CUDA_PROVIDERS) # Mesh Loading and Conversion Functions def load_mesh_with_trimesh(file_name, file_type=None): mesh = trimesh.load(file_name, file_type=file_type) if isinstance(mesh, trimesh.Scene): mesh = _process_trimesh_scene(mesh) vertices = torch.from_numpy(mesh.vertices).T faces = torch.from_numpy(mesh.faces).T colors = _get_mesh_colors(mesh) return vertices, faces, colors def _process_trimesh_scene(mesh): from io import BytesIO with BytesIO() as f: mesh.export(f, file_type="obj") f.seek(0) mesh = trimesh.load(f, file_type="obj") if isinstance(mesh, trimesh.Scene): mesh = trimesh.util.concatenate( tuple(trimesh.Trimesh(vertices=g.vertices, faces=g.faces) for g in mesh.geometry.values())) return mesh def _get_mesh_colors(mesh): if mesh.visual is not None and hasattr(mesh.visual, 'vertex_colors'): return torch.from_numpy(mesh.visual.vertex_colors)[..., :3].T / 255. return torch.ones_like(mesh.vertices.T) * 0.5 # Mesh Conversion Functions def meshlab_mesh_to_py3dmesh(mesh: ml.Mesh) -> Meshes: verts = torch.from_numpy(mesh.vertex_matrix()).float() faces = torch.from_numpy(mesh.face_matrix()).long() colors = torch.from_numpy(mesh.vertex_color_matrix()[..., :3]).float() textures = TexturesVertex(verts_features=[colors]) return Meshes(verts=[verts], faces=[faces], textures=textures) def py3dmesh_to_meshlab_mesh(meshes: Meshes) -> ml.Mesh: colors_in = F.pad(meshes.textures.verts_features_packed().cpu().float(), [0,1], value=1).numpy().astype(np.float64) return ml.Mesh( vertex_matrix=meshes.verts_packed().cpu().float().numpy().astype(np.float64), face_matrix=meshes.faces_packed().cpu().long().numpy().astype(np.int32), v_normals_matrix=meshes.verts_normals_packed().cpu().float().numpy().astype(np.float64), v_color_matrix=colors_in) # Normal Map Rotation Functions def rotate_normalmap_by_angle(normal_map: np.ndarray, angle: float): angle_rad = np.radians(angle) R = np.array([ [np.cos(angle_rad), 0, np.sin(angle_rad)], [0, 1, 0], [-np.sin(angle_rad), 0, np.cos(angle_rad)] ]) return np.dot(normal_map.reshape(-1, 3), R.T).reshape(normal_map.shape) def rotate_normals(normal_pils, return_types='np', rotate_direction=1): n_views = len(normal_pils) ret = [] for idx, rgba_normal in enumerate(normal_pils): normal_np = _process_normal_map(rgba_normal, idx, n_views, rotate_direction) ret.append(_format_output(normal_np, return_types)) return ret def _process_normal_map(rgba_normal, idx, n_views, rotate_direction): normal_np = np.array(rgba_normal)[:, :, :3] / 255 * 2 - 1 alpha_np = np.array(rgba_normal)[:, :, 3] / 255 normal_np = rotate_normalmap_by_angle(normal_np, rotate_direction * idx * (360 / n_views)) normal_np = (normal_np + 1) / 2 * alpha_np[..., None] return np.concatenate([normal_np * 255, alpha_np[:, :, None] * 255], axis=-1) def _format_output(normal_np, return_types): if return_types == 'np': return normal_np elif return_types == 'pil': return Image.fromarray(normal_np.astype(np.uint8)) else: raise ValueError(f"return_types should be 'np' or 'pil', but got {return_types}") # Background Change Functions def change_bkgd(img_pils, new_bkgd=(0., 0., 0.)): new_bkgd = np.array(new_bkgd).reshape(1, 1, 3) return [_process_image(rgba_img, new_bkgd) for rgba_img in img_pils] def _process_image(rgba_img, new_bkgd): img_np = np.array(rgba_img)[:, :, :3] / 255 alpha_np = np.array(rgba_img)[:, :, 3] / 255 ori_bkgd = img_np[:1, :1] alpha_np_clamp = np.clip(alpha_np, 1e-6, 1) ori_img_np = (img_np - ori_bkgd * (1 - alpha_np[..., None])) / alpha_np_clamp[..., None] img_np = np.where(alpha_np[..., None] > 0.05, ori_img_np * alpha_np[..., None] + new_bkgd * (1 - alpha_np[..., None]), new_bkgd) rgba_img_np = np.concatenate([img_np * 255, alpha_np[..., None] * 255], axis=-1) return Image.fromarray(rgba_img_np.astype(np.uint8)) # Mesh Cleaning Function def simple_clean_mesh(pyml_mesh: ml.Mesh, apply_smooth=True, stepsmoothnum=1, apply_sub_divide=False, sub_divide_threshold=0.25): ms = ml.MeshSet() ms.add_mesh(pyml_mesh, "cube_mesh") if apply_smooth: ms.apply_filter("apply_coord_laplacian_smoothing", stepsmoothnum=stepsmoothnum, cotangentweight=False) if apply_sub_divide: ms.apply_filter("meshing_repair_non_manifold_vertices") ms.apply_filter("meshing_repair_non_manifold_edges", method='Remove Faces') ms.apply_filter("meshing_surface_subdivision_loop", iterations=2, threshold=ml.PercentageValue(sub_divide_threshold)) return meshlab_mesh_to_py3dmesh(ms.current_mesh()) # Image Processing Functions def expand2square(pil_img, background_color): width, height = pil_img.size if width == height: return pil_img new_size = max(width, height) result = Image.new(pil_img.mode, (new_size, new_size), background_color) offset = ((new_size - width) // 2, (new_size - height) // 2) result.paste(pil_img, offset) return result def simple_preprocess(input_image, rembg_session=rembg_session, background_color=255): RES = 2048 input_image.thumbnail([RES, RES], Image.Resampling.LANCZOS) if input_image.mode != 'RGBA': image_rem = input_image.convert('RGBA') input_image = remove(image_rem, alpha_matting=False, session=rembg_session) arr = np.asarray(input_image) alpha = arr[:, :, -1] x_nonzero, y_nonzero = np.nonzero(alpha > 60) x_min, x_max = x_nonzero.min(), x_nonzero.max() y_min, y_max = y_nonzero.min(), y_nonzero.max() arr = arr[x_min:x_max+1, y_min:y_max+1] input_image = Image.fromarray(arr) return expand2square(input_image, (background_color, background_color, background_color, 0)) # Mesh Saving Functions def save_py3dmesh_with_trimesh_fast(meshes: Meshes, save_glb_path, apply_sRGB_to_LinearRGB=True): vertices = meshes.verts_packed().cpu().float().numpy() triangles = meshes.faces_packed().cpu().long().numpy() np_color = meshes.textures.verts_features_packed().cpu().float().numpy() if save_glb_path.endswith(".glb"): vertices[:, [0, 2]] = -vertices[:, [0, 2]] if apply_sRGB_to_LinearRGB: np_color = srgb_to_linear(np_color) mesh = trimesh.Trimesh(vertices=vertices, faces=triangles, vertex_colors=np_color) mesh.remove_unreferenced_vertices() mesh.export(save_glb_path) if save_glb_path.endswith(".glb"): fix_vert_color_glb(save_glb_path) print(f"Saved to {save_glb_path}") def save_glb_and_video(save_mesh_prefix: str, meshes: Meshes, with_timestamp=True, **kwargs) -> Tuple[str, str]: import time if '.' in save_mesh_prefix: save_mesh_prefix = ".".join(save_mesh_prefix.split('.')[:-1]) if with_timestamp: save_mesh_prefix = save_mesh_prefix + f"_{int(time.time())}" ret_mesh = save_mesh_prefix + ".glb" save_py3dmesh_with_trimesh_fast(meshes, ret_mesh) return ret_mesh, None # Utility Functions def srgb_to_linear(c_srgb): return np.where(c_srgb <= 0.04045, c_srgb / 12.92, ((c_srgb + 0.055) / 1.055) ** 2.4).clip(0, 1.) def fix_vert_color_glb(mesh_path): from pygltflib import GLTF2, Material, PbrMetallicRoughness obj1 = GLTF2().load(mesh_path) obj1.meshes[0].primitives[0].material = 0 obj1.materials.append(Material( pbrMetallicRoughness = PbrMetallicRoughness( baseColorFactor = [1.0, 1.0, 1.0, 1.0], metallicFactor = 0., roughnessFactor = 1.0, ), emissiveFactor = [0.0, 0.0, 0.0], doubleSided = True, )) obj1.save(mesh_path) def init_target(img_pils, new_bkgd=(0., 0., 0.), device="cuda"): new_bkgd = torch.tensor(new_bkgd, dtype=torch.float32).view(1, 1, 3).to(device) imgs = torch.stack([torch.from_numpy(np.array(img, dtype=np.float32)) for img in img_pils]).to(device) / 255 img_nps, alpha_nps = imgs[..., :3], imgs[..., 3] ori_bkgds = img_nps[:, :1, :1] alpha_nps_clamp = torch.clamp(alpha_nps, 1e-6, 1) ori_img_nps = (img_nps - ori_bkgds * (1 - alpha_nps.unsqueeze(-1))) / alpha_nps_clamp.unsqueeze(-1) ori_img_nps = torch.clamp(ori_img_nps, 0, 1) img_nps = torch.where(alpha_nps.unsqueeze(-1) > 0.05, ori_img_nps * alpha_nps.unsqueeze(-1) + new_bkgd * (1 - alpha_nps.unsqueeze(-1)), new_bkgd) return torch.cat([img_nps, alpha_nps.unsqueeze(-1)], dim=-1) def save_obj_and_video(save_mesh_prefix: str, meshes: Meshes, with_timestamp=True, **kwargs) -> Tuple[str, str]: if '.' in save_mesh_prefix: save_mesh_prefix = ".".join(save_mesh_prefix.split('.')[:-1]) if with_timestamp: save_mesh_prefix = save_mesh_prefix + f"_{int(time.time())}" ret_mesh = save_mesh_prefix + ".obj" vertices = meshes.verts_packed().cpu().float().numpy() triangles = meshes.faces_packed().cpu().long().numpy() np_color = meshes.textures.verts_features_packed().cpu().float().numpy() # Apply sRGB to LinearRGB conversion np_color = srgb_to_linear(np_color) mesh = trimesh.Trimesh(vertices=vertices, faces=triangles, vertex_colors=np_color) mesh.remove_unreferenced_vertices() mesh.export(ret_mesh) print(f"Saved to {ret_mesh}") return ret_mesh, None