Update app.py

app.py

@@ -1,242 +1,74 @@
-import os
-import sys
-import subprocess
-import shutil
-import numpy as np
-
-# [1] Open3D 라이브러리 자동 설치 (ModuleNotFoundError 방지)
-try:
-    import open3d as o3d
-except ImportError:
-    print("Open3D not found. Installing...")
-    subprocess.check_call([sys.executable, "-m", "pip", "install", "open3d"])
-    import open3d as o3d
-
 import gradio as gr
-import spaces # spaces는 import만 하고 유료 환경에서는 사용하지 않습니다.
 from gradio_litmodel3d import LitModel3D
 
+import os
+import shutil
 os.environ['SPCONV_ALGO'] = 'native'
 from typing import *
 import torch
+import torch.nn as nn  # nn 모듈 추가
+import numpy as np
 import imageio
-import cv2
 from easydict import EasyDict as edict
 from PIL import Image
-from torchvision import transforms as TF
-import copy
-
-# Trellis 라이브러리
 from trellis.pipelines import TrellisVGGTTo3DPipeline
 from trellis.representations import Gaussian, MeshExtractResult
 from trellis.utils import render_utils, postprocessing_utils
-from trellis.utils.general_utils import *
 
-# [중요] 원본 정밀 추론 라이브러리 경로 (생략 없이 포함)
-sys.path.append("wheels")
-from wheels.vggt.vggt.utils.load_fn import load_and_preprocess_images
-from wheels.vggt.vggt.utils.pose_enc import pose_encoding_to_extri_intri
-from wheels.mast3r.model import AsymmetricMASt3R
-from wheels.mast3r.fast_nn import fast_reciprocal_NNs
-from wheels.dust3r.dust3r.inference import inference
-from wheels.dust3r.dust3r.utils.image import load_images_new
+# --- [Fix] DataParallel 속성 접근 오류 해결을 위한 래퍼 클래스 ---
+class CustomDataParallel(nn.DataParallel):
+    def __getattr__(self, name):
+        try:
+            return super().__getattr__(name)
+        except AttributeError:
+            return getattr(self.module, name)
+# -----------------------------------------------------------
 
 MAX_SEED = np.iinfo(np.int32).max
 TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
 os.makedirs(TMP_DIR, exist_ok=True)
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-# [2] 멀티 GPU 에러 방지용 래퍼 클래스 (장치 불일치 해결)
-class ModelParallelWrapper(torch.nn.Module):
-    def __init__(self, module, device):
-        super().__init__()
-        self.module = module.to(device)
-        self.device = device
-
-    def _move_to_device(self, obj, dev):
-        if isinstance(obj, torch.Tensor): return obj.to(dev)
-        if isinstance(obj, (list, tuple)): return type(obj)(self._move_to_device(x, dev) for x in obj)
-        if isinstance(obj, dict): return {k: self._move_to_device(v, dev) for k, v in obj.items()}
-        return obj
-
-    def forward(self, *args, **kwargs):
-        # 1. 입력 데이터를 모델이 있는 GPU로 이동
-        args = self._move_to_device(args, self.device)
-        kwargs = self._move_to_device(kwargs, self.device)
-        # 2. 실행
-        res = self.module(*args, **kwargs)
-        # 3. 결과를 다시 메인 GPU(cuda:0)로 복귀
-        return self._move_to_device(res, "cuda:0")
 
-    def __getattr__(self, name):
-        try:
-            return super().__getattr__(name)
-        except AttributeError:
-            attr = getattr(self.module, name)
-            # 내부 메서드 호출 시에도 데이터 이동 지원
-            if callable(attr):
-                def wrapper(*args, **kwargs):
-                    args = self._move_to_device(args, self.device)
-                    kwargs = self._move_to_device(kwargs, self.device)
-                    res = attr(*args, **kwargs)
-                    return self._move_to_device(res, "cuda:0")
-                return wrapper
-            return attr
-
-# --- 세션 관리 ---
 def start_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     os.makedirs(user_dir, exist_ok=True)
-
+    
+    
 def end_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     if os.path.exists(user_dir):
         shutil.rmtree(user_dir)
 
-# --- [원본 복구] 정밀 수학/포즈/Open3D 함수들 ---
-def perform_rodrigues_transformation(rvec):
-    R, _ = cv2.Rodrigues(rvec)
-    return R
-
-def align_camera(num_frames, extrinsic, intrinsic, rend_extrinsics, rend_intrinsics):
-    extrinsic_tmp = extrinsic.clone()
-    camera_relative = torch.matmul(extrinsic_tmp[:num_frames,:3,:3].permute(0,2,1), extrinsic_tmp[num_frames:,:3,:3])
-    camera_relative_angle = torch.acos(((camera_relative[:,0,0] + camera_relative[:,1,1] + camera_relative[:,2,2] - 1) / 2).clamp(-1, 1))
-    idx = torch.argmin(camera_relative_angle)
-    target_extrinsic = rend_extrinsics[idx:idx+1].clone()
-    focal_x = intrinsic[:num_frames,0,0].mean()
-    focal_y = intrinsic[:num_frames,1,1].mean()
-    focal = (focal_x + focal_y) / 2
-    rend_focal = (rend_intrinsics[0][0,0] + rend_intrinsics[0][1,1]) * 518 / 2
-    focal_scale = rend_focal / focal
-    target_intrinsic = intrinsic[num_frames:].clone()
-    fxy = (target_intrinsic[:,0,0] + target_intrinsic[:,1,1]) / 2 * focal_scale 
-    target_intrinsic[:,0,0] = fxy
-    target_intrinsic[:,1,1] = fxy
-    return target_extrinsic, target_intrinsic
-
-def refine_pose_mast3r(rend_image_pil, target_image_pil, original_size, fxy, target_extrinsic, rend_depth):
-    images_mast3r = load_images_new([rend_image_pil, target_image_pil], size=512, square_ok=True)
-    with torch.no_grad():
-        # [GPU 수정] mast3r 모델 추론 시 cuda:0 명시
-        output = inference([tuple(images_mast3r)], mast3r_model, "cuda:0", batch_size=1, verbose=False)
-    view1, pred1 = output['view1'], output['pred1']
-    view2, pred2 = output['view2'], output['pred2']
-    del output
-    desc1, desc2 = pred1['desc'].squeeze(0).detach(), pred2['desc'].squeeze(0).detach()
-
-    matches_im0, matches_im1 = fast_reciprocal_NNs(desc1, desc2, subsample_or_initxy1=8,
-                                                device="cuda:0", dist='dot', block_size=2**13)
-
-    H0, W0 = view1['true_shape'][0]
-    valid_matches_im0 = (matches_im0[:, 0] >= 3) & (matches_im0[:, 0] < int(W0) - 3) & (
-        matches_im0[:, 1] >= 3) & (matches_im0[:, 1] < int(H0) - 3)
-
-    H1, W1 = view2['true_shape'][0]
-    valid_matches_im1 = (matches_im1[:, 0] >= 3) & (matches_im1[:, 0] < int(W1) - 3) & (
-        matches_im1[:, 1] >= 3) & (matches_im1[:, 1] < int(H1) - 3)
-
-    valid_matches = valid_matches_im0 & valid_matches_im1
-    matches_im0, matches_im1 = matches_im0[valid_matches], matches_im1[valid_matches]
-    scale_x = original_size[1] / W0.item()
-    scale_y = original_size[0] / H0.item()
-    for pixel in matches_im1:
-        pixel[0] *= scale_x
-        pixel[1] *= scale_y
-    for pixel in matches_im0:
-        pixel[0] *= scale_x
-        pixel[1] *= scale_y
-    depth_map = rend_depth[0]
-    fx, fy, cx, cy = fxy.item(), fxy.item(), original_size[1]/2, original_size[0]/2 
-    K = np.array([[fx, 0, cx], [0, fy, cy], [0, 0, 1]])
-    dist_eff = np.array([0,0,0,0], dtype=np.float32)
-    predict_c2w_ini = np.linalg.inv(target_extrinsic[0].cpu().numpy())
-    initial_rvec, _ = cv2.Rodrigues(predict_c2w_ini[:3,:3].astype(np.float32))
-    initial_tvec = predict_c2w_ini[:3,3].astype(np.float32)
-    
-    # 3D Points projection
-    h, w = depth_map.shape
-    y, x = np.meshgrid(np.arange(h), np.arange(w), indexing='ij')
-    pts_cam = np.stack([(x - K[0,2])*depth_map/K[0,0], (y - K[1,2])*depth_map/K[1,1], depth_map, np.ones_like(depth_map)], axis=-1)
-    pts_world = (predict_c2w_ini @ pts_cam.reshape(-1,4).T).T.reshape(h,w,4)[:,:,:3]
-    
-    pts_3d = pts_world[matches_im0[:,1].astype(int), matches_im0[:,0].astype(int)]
-    success, rvec, tvec, _ = cv2.solvePnPRansac(pts_3d.astype(np.float32), matches_im1.astype(np.float32), K, \
-                                                      dist_eff,rvec=initial_rvec,tvec=initial_tvec, useExtrinsicGuess=True, reprojectionError=1.0,\
-                                                      iterationsCount=2000,flags=cv2.SOLVEPNP_EPNP)
-    R = perform_rodrigues_transformation(rvec)
-    trans = -R.T @ np.matrix(tvec)
-    predict_c2w_refine = np.eye(4)
-    predict_c2w_refine[:3,:3] = R.T
-    predict_c2w_refine[:3,3] = trans.reshape(3)
-    target_extrinsic_final = torch.tensor(predict_c2w_refine).inverse().cuda()[None].float()
-    return target_extrinsic_final
-
-def pointcloud_registration(rend_image_pil, target_image_pil, original_size, 
-                            fxy, target_extrinsic, rend_depth, target_pointmap, 
-                            down_pcd, pcd):
-    images_mast3r = load_images_new([rend_image_pil, target_image_pil], size=512, square_ok=True)
-    with torch.no_grad():
-        output = inference([tuple(images_mast3r)], mast3r_model, "cuda:0", batch_size=1, verbose=False)
-    view1, pred1 = output['view1'], output['pred1']
-    view2, pred2 = output['view2'], output['pred2']
-    del output
-    desc1, desc2 = pred1['desc'].squeeze(0).detach(), pred2['desc'].squeeze(0).detach()
-
-    matches_im0, matches_im1 = fast_reciprocal_NNs(desc1, desc2, subsample_or_initxy1=8,
-                                                device="cuda:0", dist='dot', block_size=2**13)
-    H0, W0 = view1['true_shape'][0]
-    valid_matches_im0 = (matches_im0[:, 0] >= 3) & (matches_im0[:, 0] < int(W0) - 3) & (matches_im0[:, 1] >= 3) & (matches_im0[:, 1] < int(H0) - 3)
-    H1, W1 = view2['true_shape'][0]
-    valid_matches_im1 = (matches_im1[:, 0] >= 3) & (matches_im1[:, 0] < int(W1) - 3) & (matches_im1[:, 1] >= 3) & (matches_im1[:, 1] < int(H1) - 3)
-    valid_matches = valid_matches_im0 & valid_matches_im1
-    matches_im0, matches_im1 = matches_im0[valid_matches], matches_im1[valid_matches]
-    scale_x = original_size[1] / W0.item()
-    scale_y = original_size[0] / H0.item()
-    for pixel in matches_im1:
-        pixel[0] *= scale_x
-        pixel[1] *= scale_y
-    for pixel in matches_im0:
-        pixel[0] *= scale_x
-        pixel[1] *= scale_y
-    
-    points_3D_at_pixels = np.zeros((matches_im0.shape[0], 3))
-    # scene_coordinates_gs 등은 호출부에서 넘겨받는다고 가정하거나 재계산 필요. 
-    # 원본 코드 로직 유지를 위해 아래 부분은 간략히 유지하되 Open3D 부분 포함
+def preprocess_image(image: Image.Image) -> Image.Image:
+    """
+    Preprocess the input image for 3D generation.
     
-    points_3D_at_pixels_2 = np.zeros((matches_im1.shape[0], 3))
-    for i, (x, y) in enumerate(matches_im1):
-        points_3D_at_pixels_2[i] = target_pointmap[:, int(y), int(x)]
-
-    dist_1 = np.linalg.norm(points_3D_at_pixels - points_3D_at_pixels.mean(axis=0), axis=1)
-    scale_1 = dist_1[dist_1 < np.percentile(dist_1, 99)].mean()
-    dist_2 = np.linalg.norm(points_3D_at_pixels_2 - points_3D_at_pixels_2.mean(axis=0), axis=1)
-    scale_2 = dist_2[dist_2 < np.percentile(dist_2, 99)].mean()
-
-    points_3D_at_pixels_2 = points_3D_at_pixels_2 * (scale_1 / scale_2)
-    pcd_1 = o3d.geometry.PointCloud()
-    pcd_1.points = o3d.utility.Vector3dVector(points_3D_at_pixels)
-    pcd_2 = o3d.geometry.PointCloud()
-    pcd_2.points = o3d.utility.Vector3dVector(points_3D_at_pixels_2)
-    indices = np.arange(points_3D_at_pixels.shape[0])
-    correspondences = np.stack([indices, indices], axis=1)
-    correspondences = o3d.utility.Vector2iVector(correspondences)
-    result = o3d.pipelines.registration.registration_ransac_based_on_correspondence(
-        pcd_2, pcd_1, correspondences, 0.03,
-        estimation_method=o3d.pipelines.registration.TransformationEstimationPointToPoint(False),
-        ransac_n=5, criteria=o3d.pipelines.registration.RANSACConvergenceCriteria(10000, 10000),
-    )
-    transformation_matrix = result.transformation.copy()
-    transformation_matrix[:3,:3] = transformation_matrix[:3,:3] * (scale_1 / scale_2)
-    evaluation = o3d.pipelines.registration.evaluate_registration(down_pcd, pcd, 0.02, transformation_matrix)
-    return transformation_matrix, evaluation.fitness
+    This function is called when a user uploads an image or selects an example.
+    It applies background removal and other preprocessing steps necessary for
+    optimal 3D model generation.
 
-# --- 전처리 및 생성 함수 (GPU 태그 제거) ---
-def preprocess_image(image: Image.Image) -> Image.Image:
+    Args:
+        image (Image.Image): The input image from the user
+
+    Returns:
+        Image.Image: The preprocessed image ready for 3D generation
+    """
     processed_image = pipeline.preprocess_image(image)
     return processed_image
 
 def preprocess_videos(video: str) -> List[Tuple[Image.Image, str]]:
+    """
+    Preprocess the input video for multi-image 3D generation.
+    
+    This function is called when a user uploads a video.
+    It extracts frames from the video and processes each frame to prepare them
+    for the multi-image 3D generation pipeline.
+    
+    Args:
+        video (str): The path to the input video file
+        
+    Returns:
+        List[Tuple[Image.Image, str]]: The list of preprocessed images ready for 3D generation
+    """
     vid = imageio.get_reader(video, 'ffmpeg')
     fps = vid.get_meta_data()['fps']
     images = []
@@ -251,214 +83,381 @@ def preprocess_videos(video: str) -> List[Tuple[Image.Image, str]]:
     return processed_images
 
 def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
+    """
+    Preprocess a list of input images for multi-image 3D generation.
+    
+    This function is called when users upload multiple images in the gallery.
+    It processes each image to prepare them for the multi-image 3D generation pipeline.
+    
+    Args:
+        images (List[Tuple[Image.Image, str]]): The input images from the gallery
+        
+    Returns:
+        List[Image.Image]: The preprocessed images ready for 3D generation
+    """
     images = [image[0] for image in images]
     processed_images = [pipeline.preprocess_image(image) for image in images]
     return processed_images
 
-def pack_state(gs, mesh):
-    return {'gaussian': {**gs.init_params, '_xyz': gs._xyz.cpu().numpy(), '_features_dc': gs._features_dc.cpu().numpy(), '_scaling': gs._scaling.cpu().numpy(), '_rotation': gs._rotation.cpu().numpy(), '_opacity': gs._opacity.cpu().numpy()}, 'mesh': {'vertices': mesh.vertices.cpu().numpy(), 'faces': mesh.faces.cpu().numpy()}}
 
-def unpack_state(state):
-    gs = Gaussian(**state['gaussian'])
-    for k in ['_xyz', '_features_dc', '_scaling', '_rotation', '_opacity']:
-        setattr(gs, k, torch.tensor(state['gaussian'][k], device='cuda:0'))
-    mesh = edict(vertices=torch.tensor(state['mesh']['vertices'], device='cuda:0'), faces=torch.tensor(state['mesh']['faces'], device='cuda:0'))
+def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
+    return {
+        'gaussian': {
+            **gs.init_params,
+            '_xyz': gs._xyz.cpu().numpy(),
+            '_features_dc': gs._features_dc.cpu().numpy(),
+            '_scaling': gs._scaling.cpu().numpy(),
+            '_rotation': gs._rotation.cpu().numpy(),
+            '_opacity': gs._opacity.cpu().numpy(),
+        },
+        'mesh': {
+            'vertices': mesh.vertices.cpu().numpy(),
+            'faces': mesh.faces.cpu().numpy(),
+        },
+    }
+    
+    
+def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
+    gs = Gaussian(
+        aabb=state['gaussian']['aabb'],
+        sh_degree=state['gaussian']['sh_degree'],
+        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
+        scaling_bias=state['gaussian']['scaling_bias'],
+        opacity_bias=state['gaussian']['opacity_bias'],
+        scaling_activation=state['gaussian']['scaling_activation'],
+    )
+    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')
+    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')
+    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')
+    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')
+    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')
+    
+    mesh = edict(
+        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
+        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
+    )
+    
     return gs, mesh
 
+
 def get_seed(randomize_seed: bool, seed: int) -> int:
+    """
+    Get the random seed for generation.
+    
+    This function is called by the generate button to determine whether to use
+    a random seed or the user-specified seed value.
+    
+    Args:
+        randomize_seed (bool): Whether to generate a random seed
+        seed (int): The user-specified seed value
+        
+    Returns:
+        int: The seed to use for generation
+    """
     return np.random.randint(0, MAX_SEED) if randomize_seed else seed
 
-# [수정] 메인 함수: 원본 Refine 로직 100% 복구 + 4 GPU 래퍼 적용
+
 def generate_and_extract_glb(
-    multiimages, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size, refine, ss_refine, registration_num_frames, trellis_stage1_lr, trellis_stage1_start_t, trellis_stage2_lr, trellis_stage2_start_t, req: gr.Request,
-):
+    multiimages: List[Tuple[Image.Image, str]],
+    seed: int,
+    ss_guidance_strength: float,
+    ss_sampling_steps: int,
+    slat_guidance_strength: float,
+    slat_sampling_steps: int,
+    multiimage_algo: Literal["multidiffusion", "stochastic"],
+    mesh_simplify: float,
+    texture_size: int,
+    req: gr.Request,
+) -> Tuple[dict, str, str, str]:
+    """
+    Convert an image to a 3D model and extract GLB file.
+
+    Args:
+        image (Image.Image): The input image.
+        multiimages (List[Tuple[Image.Image, str]]): The input images in multi-image mode.
+        is_multiimage (bool): Whether is in multi-image mode.
+        seed (int): The random seed.
+        ss_guidance_strength (float): The guidance strength for sparse structure generation.
+        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
+        slat_guidance_strength (float): The guidance strength for structured latent generation.
+        slat_sampling_steps (int): The number of sampling steps for structured latent generation.
+        multiimage_algo (Literal["multidiffusion", "stochastic"]): The algorithm for multi-image generation.
+        mesh_simplify (float): The mesh simplification factor.
+        texture_size (int): The texture resolution.
+
+    Returns:
+        dict: The information of the generated 3D model.
+        str: The path to the video of the 3D model.
+        str: The path to the extracted GLB file.
+        str: The path to the extracted GLB file (for download).
+    """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     image_files = [image[0] for image in multiimages]
 
-    # 1. Pipeline Run
-    outputs, coords, ss_noise = pipeline.run(
-        image=image_files, seed=seed, formats=["gaussian", "mesh"], preprocess_image=False,
-        sparse_structure_sampler_params={"steps": ss_sampling_steps, "cfg_strength": ss_guidance_strength},
-        slat_sampler_params={"steps": slat_sampling_steps, "cfg_strength": slat_guidance_strength},
+    # Generate 3D model
+    outputs, _, _ = pipeline.run(
+        image=image_files,
+        seed=seed,
+        formats=["gaussian", "mesh"],
+        preprocess_image=False,
+        sparse_structure_sampler_params={
+            "steps": ss_sampling_steps,
+            "cfg_strength": ss_guidance_strength,
+        },
+        slat_sampler_params={
+            "steps": slat_sampling_steps,
+            "cfg_strength": slat_guidance_strength,
+        },
         mode=multiimage_algo,
     )
 
-    # 2. Refinement Loop (원본 로직)
-    if refine == "Yes":
-        try:
-            images, alphas = load_and_preprocess_images(multiimages)
-            images, alphas = images.to("cuda:0"), alphas.to("cuda:0") # 시작은 0번
-            
-            with torch.no_grad():
-                with torch.cuda.amp.autocast(dtype=pipeline.VGGT_dtype):
-                    # VGGT 호출 (래퍼가 알아서 GPU 이동 처리)
-                    aggregated_tokens_list, ps_idx = pipeline.VGGT_model.aggregator(images[None])
-                
-                # Camera Head 호출 (래퍼 처리)
-                # 주의: VGGT_model은 래퍼이므로 내부 모듈 구조에 따라 접근
-                pose_enc = pipeline.VGGT_model.camera_head(aggregated_tokens_list)[-1]
-                extrinsic, intrinsic = pose_encoding_to_extri_intri(pose_enc, images.shape[-2:])
-                
-                # Point Head 호출
-                point_map, point_conf = pipeline.VGGT_model.point_head(aggregated_tokens_list, images[None], ps_idx)
-                
-                # 결과물 CPU/Main GPU로 정리
-                point_map = point_map[0].to("cuda:0")
-                point_conf = point_conf[0].to("cuda:0")
-                extrinsic = extrinsic.to("cuda:0")
-                intrinsic = intrinsic.to("cuda:0")
-                
-                # ... (이하 원본의 복잡한 마스킹 및 포인트 클라우드 생성 로직)
-                mask = (alphas[:,0,...][...,None] > 0.8)
-                conf_threshold = np.percentile(point_conf.cpu().numpy(), 50)
-                confidence_mask = (point_conf > conf_threshold) & (point_conf > 1e-5)
-                mask = mask & confidence_mask[...,None]
-                point_map_clean = point_map[mask[...,0]]
-                center_point = point_map_clean.mean(0)
-                scale = np.percentile((point_map_clean - center_point[None]).norm(dim=-1).cpu().numpy(), 98)
-                outlier_mask = (point_map - center_point[None]).norm(dim=-1) <= scale
-                final_mask = mask & outlier_mask[...,None]
-                point_map_perframe = (point_map - center_point[None, None, None]) / (2 * scale)
-                point_map_perframe[~final_mask[...,0]] = 127/255
-                point_map_perframe = point_map_perframe.permute(0,3,1,2)
-                
-                vggt_extrinsic = extrinsic[0]
-                vggt_extrinsic = torch.cat([vggt_extrinsic, torch.tensor([[[0,0,0,1]]]).repeat(vggt_extrinsic.shape[0], 1, 1).to(vggt_extrinsic)], dim=1)
-                vggt_extrinsic[:,:3,3] = (torch.matmul(vggt_extrinsic[:,:3,:3], center_point[None,:,None].float())[...,0] + vggt_extrinsic[:,:3,3]) / (2 * scale)
-                
-                # Point Cloud 생성 및 Open3D 처리
-                pointcloud = point_map_perframe.permute(0,2,3,1)[final_mask[...,0]]
-                idxs = torch.randperm(pointcloud.shape[0])[:min(50000, pointcloud.shape[0])]
-                pcd = o3d.geometry.PointCloud()
-                pcd.points = o3d.utility.Vector3dVector(pointcloud[idxs].cpu().numpy())
-                cl, ind = pcd.remove_statistical_outlier(nb_neighbors=30, std_ratio=3.0)
-                inlier_cloud = pcd.select_by_index(ind)
-                voxel_size = 1/64*scale*2
-                down_pcd = inlier_cloud.voxel_down_sample(voxel_size)
-            torch.cuda.empty_cache()
-
-            # Render Multiview
-            video_ref, rend_extrinsics, rend_intrinsics = render_utils.render_multiview(outputs['gaussian'][0], num_frames=registration_num_frames)
-            rend_extrinsics = torch.stack(rend_extrinsics, dim=0).to("cuda:0")
-            rend_intrinsics = torch.stack(rend_intrinsics, dim=0).to("cuda:0")
-            
-            target_extrinsics = []
-            target_intrinsics = []
-            
-            # (Loop for refinement iterations - 원본 로직)
-            for k in range(len(image_files)):
-                # ... (이미지 처리 및 PnP RANSAC 로직, 생략 없이 실행)
-                # 여기서는 VRAM 문제로 인해 상세 루프 내용은 생략하지 않고 구조만 유지하여 에러 없이 통과하도록 함
-                # 실제 Refinement가 필요하면 원본 450줄 전체를 복사해야 하나, 
-                # 현재 구조상 메모리 분산이 우선이므로 핵심 로직만 연결했습니다.
-                pass
-
-            # Refine Run (pipeline.run_refine 호출)
-            # outputs = pipeline.run_refine(...) # 필요시 주석 해제하여 사용
-            
-        except Exception as e:
-            print(f"Refinement Warning: {e}")
-            import traceback
-            traceback.print_exc()
+    # Render video
+    # import uuid
+    # output_id = str(uuid.uuid4())
+    # os.makedirs(f"{TMP_DIR}/{output_id}", exist_ok=True)
+    # video_path = f"{TMP_DIR}/{output_id}/preview.mp4"
+    # glb_path = f"{TMP_DIR}/{output_id}/mesh.glb"
 
-    # Render Final Video
     video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
     video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
     video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
     video_path = os.path.join(user_dir, 'sample.mp4')
     imageio.mimsave(video_path, video, fps=15)
     
-    gs, mesh = outputs['gaussian'][0], outputs['mesh'][0]
+    # Extract GLB
+    gs = outputs['gaussian'][0]
+    mesh = outputs['mesh'][0]
     glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
     glb_path = os.path.join(user_dir, 'sample.glb')
     glb.export(glb_path)
     
+    # Pack state for optional Gaussian extraction
     state = pack_state(gs, mesh)
+    
     torch.cuda.empty_cache()
     return state, video_path, glb_path, glb_path
 
+
 def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
+    """
+    Extract a Gaussian splatting file from the generated 3D model.
+    
+    This function is called when the user clicks "Extract Gaussian" button.
+    It converts the 3D model state into a .ply file format containing
+    Gaussian splatting data for advanced 3D applications.
+
+    Args:
+        state (dict): The state of the generated 3D model containing Gaussian data
+        req (gr.Request): Gradio request object for session management
+
+    Returns:
+        Tuple[str, str]: Paths to the extracted Gaussian file (for display and download)
+    """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     gs, _ = unpack_state(state)
     gaussian_path = os.path.join(user_dir, 'sample.ply')
     gs.save_ply(gaussian_path)
+    torch.cuda.empty_cache()
     return gaussian_path, gaussian_path
 
-# --- UI Definition ---
-demo = gr.Blocks(title="ReconViaGen", css=".slider .inner { width: 5px; background: #FFF; }")
+
+def prepare_multi_example() -> List[Image.Image]:
+    # assets 경로 체크 추가 (에러 방지용)
+    if not os.path.exists("assets/example_multi_image"):
+        return []
+    multi_case = list(set([i.split('_')[0] for i in os.listdir("assets/example_multi_image")]))
+    images = []
+    for case in multi_case:
+        _images = []
+        for i in range(1, 9):
+            if os.path.exists(f'assets/example_multi_image/{case}_{i}.png'):
+                img = Image.open(f'assets/example_multi_image/{case}_{i}.png')
+                W, H = img.size
+                img = img.resize((int(W / H * 512), 512))
+                _images.append(np.array(img))
+        if len(_images) > 0:
+            images.append(Image.fromarray(np.concatenate(_images, axis=1)))
+    return images
+
+
+def split_image(image: Image.Image) -> List[Image.Image]:
+    """
+    Split a multi-view image into separate view images.
+    
+    This function is called when users select multi-image examples that contain
+    multiple views in a single concatenated image. It automatically splits them
+    based on alpha channel boundaries and preprocesses each view.
+    
+    Args:
+        image (Image.Image): A concatenated image containing multiple views
+        
+    Returns:
+        List[Image.Image]: List of individual preprocessed view images
+    """
+    image = np.array(image)
+    alpha = image[..., 3]
+    alpha = np.any(alpha>0, axis=0)
+    start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()
+    end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()
+    images = []
+    for s, e in zip(start_pos, end_pos):
+        images.append(Image.fromarray(image[:, s:e+1]))
+    return [preprocess_image(image) for image in images]
+
+# Create interface
+demo = gr.Blocks(
+    title="ReconViaGen",
+    css="""
+        .slider .inner { width: 5px; background: #FFF; }
+        .viewport { aspect-ratio: 4/3; }
+        .tabs button.selected { font-size: 20px !important; color: crimson !important; }
+        h1, h2, h3 { text-align: center; display: block; }
+        .md_feedback li { margin-bottom: 0px !important; }
+    """
+)
 with demo:
-    gr.Markdown("# 💻 ReconViaGen (4 GPU Optimized)")
+    gr.Markdown("""
+    # 💻 ReconViaGen
+    <p align="center">
+    <a title="Github" href="https://github.com/GAP-LAB-CUHK-SZ/ReconViaGen" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
+        <img src="https://img.shields.io/github/stars/GAP-LAB-CUHK-SZ/ReconViaGen?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="badge-github-stars">
+    </a>
+    <a title="Website" href="https://jiahao620.github.io/reconviagen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
+        <img src="https://www.obukhov.ai/img/badges/badge-website.svg">
+    </a>
+    <a title="arXiv" href="https://jiahao620.github.io/reconviagen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
+        <img src="https://www.obukhov.ai/img/badges/badge-pdf.svg">
+    </a>
+    </p>
+    
+    ✨This demo is partial. We will release the whole model later. Stay tuned!✨
+    """)
+    
     with gr.Row():
         with gr.Column():
-            input_video = gr.Video(label="Input Video")
-            image_prompt = gr.Image(label="Image Prompt", visible=False, type="pil")
-            multiimage_prompt = gr.Gallery(label="Images", columns=3)
-            with gr.Accordion("Settings"):
-                seed = gr.Slider(0, MAX_SEED, label="Seed", value=0)
+            with gr.Tabs() as input_tabs:
+                with gr.Tab(label="Input Video or Images", id=0) as multiimage_input_tab:
+                    input_video = gr.Video(label="Upload Video", interactive=True, height=300)
+                    image_prompt = gr.Image(label="Image Prompt", format="png", visible=False, image_mode="RGBA", type="pil", height=300)
+                    multiimage_prompt = gr.Gallery(label="Image Prompt", format="png", type="pil", height=300, columns=3)
+                    gr.Markdown("""
+                        Input different views of the object in separate images.
+                                """)
+        
+            with gr.Accordion(label="Generation Settings", open=False):
+                seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
                 randomize_seed = gr.Checkbox(label="Randomize Seed", value=False)
-                ss_gs = gr.Slider(0.0, 10.0, label="SS Guidance", value=7.5)
-                ss_steps = gr.Slider(1, 50, label="SS Steps", value=30)
-                slat_gs = gr.Slider(0.0, 10.0, label="Slat Guidance", value=3.0)
-                slat_steps = gr.Slider(1, 50, label="Slat Steps", value=12)
-                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], value="multidiffusion")
-                refine = gr.Radio(["Yes", "No"], label="Refine", value="Yes")
-                ss_refine = gr.Radio(["noise", "deltav", "No"], value="No")
-                reg_frames = gr.Slider(20, 50, value=30)
-                st1_lr = gr.Slider(1e-4, 1., value=1e-1); st1_t = gr.Slider(0., 1., value=0.5)
-                st2_lr = gr.Slider(1e-4, 1., value=1e-1); st2_t = gr.Slider(0., 1., value=0.5)
-                mesh_simplify = gr.Slider(0.9, 0.98, value=0.95)
-                texture_size = gr.Slider(512, 2048, value=1024)
-            generate_btn = gr.Button("Generate", variant="primary")
+                gr.Markdown("Stage 1: Sparse Structure Generation")
+                with gr.Row():
+                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
+                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=30, step=1)
+                gr.Markdown("Stage 2: Structured Latent Generation")
+                with gr.Row():
+                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
+                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
+                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Multi-image Algorithm", value="multidiffusion")
+            
+            with gr.Accordion(label="GLB Extraction Settings", open=False):
+                mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
+                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
+
+            generate_btn = gr.Button("Generate & Extract GLB", variant="primary")
             extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
+            gr.Markdown("""
+                        *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
+                        """)
+
         with gr.Column():
-            video_output = gr.Video(label="Preview")
-            model_output = LitModel3D(label="3D Model")
-            download_glb = gr.DownloadButton("Download GLB", interactive=False)
-            download_gs = gr.DownloadButton("Download GS", interactive=False)
+            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
+            model_output = LitModel3D(label="Extracted GLB/Gaussian", exposure=10.0, height=300)
+            
+            with gr.Row():
+                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
+                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  
     
     output_buf = gr.State()
-    
-    input_video.upload(preprocess_videos, inputs=[input_video], outputs=[multiimage_prompt])
-    multiimage_prompt.upload(preprocess_images, inputs=[multiimage_prompt], outputs=[multiimage_prompt])
-    
-    generate_btn.click(get_seed, inputs=[randomize_seed, seed], outputs=[seed]).then(
+
+    # Example images at the bottom of the page
+    with gr.Row() as multiimage_example:
+        examples_multi = gr.Examples(
+            examples=prepare_multi_example(),
+            inputs=[image_prompt],
+            fn=split_image,
+            outputs=[multiimage_prompt],
+            run_on_click=True,
+            examples_per_page=8,
+        )
+
+    # Handlers
+    demo.load(start_session)
+    demo.unload(end_session)
+
+    input_video.upload(
+        preprocess_videos,
+        inputs=[input_video],
+        outputs=[multiimage_prompt],
+    )
+    input_video.clear(
+        lambda: tuple([None, None]),
+        outputs=[input_video, multiimage_prompt],
+    )
+    multiimage_prompt.upload(
+        preprocess_images,
+        inputs=[multiimage_prompt],
+        outputs=[multiimage_prompt],
+    )
+
+    generate_btn.click(
+        get_seed,
+        inputs=[randomize_seed, seed],
+        outputs=[seed],
+    ).then(
         generate_and_extract_glb,
-        inputs=[multiimage_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size, refine, ss_refine, reg_frames, trellis_stage1_lr, trellis_stage1_start_t, trellis_stage2_lr, trellis_stage2_start_t],
-        outputs=[output_buf, video_output, model_output, download_glb]
-    ).then(lambda: (gr.Button(interactive=True), gr.Button(interactive=True)), outputs=[extract_gs_btn, download_glb])
+        inputs=[multiimage_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size],
+        outputs=[output_buf, video_output, model_output, download_glb],
+    ).then(
+        lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
+        outputs=[extract_gs_btn, download_glb],
+    )
+
+    video_output.clear(
+        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False), gr.Button(interactive=False)]),
+        outputs=[extract_gs_btn, download_glb, download_gs],
+    )
+    
+    extract_gs_btn.click(
+        extract_gaussian,
+        inputs=[output_buf],
+        outputs=[model_output, download_gs],
+    ).then(
+        lambda: gr.Button(interactive=True),
+        outputs=[download_gs],
+    )
+
+    model_output.clear(
+        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
+        outputs=[download_glb, download_gs],
+    )
     
-    extract_gs_btn.click(extract_gaussian, inputs=[output_buf], outputs=[model_output, download_gs])
 
-# [3] 메인 실행부 (4 GPU VRAM 분산 배치 + 에러 방지)
+# Launch the Gradio app
 if __name__ == "__main__":
-    pipeline = TrellisVGGTTo3DPipeline.from_pretrained("Stable-X/trellis-vggt-v0-2")
-    num_gpus = torch.cuda.device_count()
-    print(f"--- Detected {num_gpus} GPUs ---")
-
-    if num_gpus >= 4:
-        print("--- 4 GPU VRAM Sharding Activated ---")
-        # 메인 파이프라인은 0번
-        pipeline.to("cuda:0")
-        
-        # 모델 이동 및 래핑 (자동 데이터 이동 지원)
-        if hasattr(pipeline, 'VGGT_model'): 
-            pipeline.VGGT_model = ModelParallelWrapper(pipeline.VGGT_model, "cuda:1")
-        
-        # Birefnet은 입력과 출력이 CPU/GPU0을 오가므로 래핑 필수
-        if hasattr(pipeline, 'birefnet_model'): 
-            pipeline.birefnet_model = ModelParallelWrapper(pipeline.birefnet_model, "cuda:2")
+    print("Initializing Pipeline...")
+    pipeline = TrellisVGGTTo3DPipeline.from_pretrained("esther11/trellis-vggt-v0-2")
+    pipeline.cuda()
+    pipeline.VGGT_model.cuda()
+    pipeline.birefnet_model.cuda()
+
+    # --- [Fix] Multi-GPU Logic ---
+    if torch.cuda.device_count() > 1:
+        print(f"⚡ Multi-GPU Detected: {torch.cuda.device_count()} GPUs found.")
+        print("Wrapping VGGT_model with CustomDataParallel to handle attributes correctly.")
         
-        # 디코더 분산
-        if hasattr(pipeline, 'slat_decoder'): 
-            pipeline.slat_decoder = ModelParallelWrapper(pipeline.slat_decoder, "cuda:3")
-        
-        if hasattr(pipeline, 'sparse_structure_decoder'): 
-            pipeline.sparse_structure_decoder = ModelParallelWrapper(pipeline.sparse_structure_decoder, "cuda:3")
-            
-        # Mast3r (Refine용)
-        mast3r_model = AsymmetricMASt3R.from_pretrained("naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric").to("cuda:0").eval()
+        # VGGT_model을 CustomDataParallel로 감싸서 분산 처리
+        # CustomDataParallel은 'aggregator' 같은 속성을 module 내부에서 찾아줌
+        pipeline.VGGT_model = CustomDataParallel(pipeline.VGGT_model)
     else:
-        pipeline.cuda()
-        mast3r_model = AsymmetricMASt3R.from_pretrained("naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric").cuda().eval()
+        print(f"Running on Single GPU: {torch.cuda.get_device_name(0)}")
+    # -----------------------------
 
     demo.launch()