Upload 63 files

.gitattributes

@@ -35,3 +35,65 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 wheels/diff_gaussian_rasterization-0.0.0-cp310-cp310-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
 wheels/nvdiffrast-0.3.3-cp310-cp310-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+assets/example_image/T.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_building_building.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_building_castle.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_building_colorful_cottage.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_building_maya_pyramid.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_building_mushroom.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_building_space_station.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_dragon.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_elephant.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_furry.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_quadruped.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_robot_crab.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_robot_dinosour.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_creature_rock_monster.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_humanoid_block_robot.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_humanoid_dragonborn.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_humanoid_dwarf.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_humanoid_goblin.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_humanoid_mech.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_crate.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_fireplace.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_gate.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_lantern.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_magicbook.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_mailbox.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_monster_chest.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_paper_machine.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_phonograph.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_portal2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_storage_chest.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_telephone.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_television.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_misc_workbench.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_biplane.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_bulldozer.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_cart.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_excavator.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_helicopter.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_locomotive.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/typical_vehicle_pirate_ship.png filter=lfs diff=lfs merge=lfs -text
+assets/example_image/weatherworn_misc_paper_machine3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/character_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/character_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/character_3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/mushroom_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/mushroom_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/mushroom_3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/orangeguy_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/orangeguy_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/orangeguy_3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/popmart_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/popmart_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/popmart_3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/rabbit_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/rabbit_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/rabbit_3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/tiger_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/tiger_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/tiger_3.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/yoimiya_1.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/yoimiya_2.png filter=lfs diff=lfs merge=lfs -text
+assets/example_multi_image/yoimiya_3.png filter=lfs diff=lfs merge=lfs -text

app.py

@@ -1,9 +1,8 @@
 import gradio as gr
-from gradio_litmodel3d import LitModel3D
 import spaces
+from gradio_litmodel3d import LitModel3D
 
 import os
-
 import shutil
 os.environ['SPCONV_ALGO'] = 'native'
 from typing import *
@@ -12,97 +11,111 @@ import numpy as np
 import imageio
 from easydict import EasyDict as edict
 from PIL import Image
-from Amodal3R.pipelines import Amodal3RImageTo3DPipeline
 from trellis.pipelines import TrellisImageTo3DPipeline
-from Amodal3R.representations import Gaussian, MeshExtractResult
-from Amodal3R.utils import render_utils, postprocessing_utils
-from segment_anything import sam_model_registry, SamPredictor
-from huggingface_hub import hf_hub_download
-import cv2
+from trellis.representations import Gaussian, MeshExtractResult
+from trellis.utils import render_utils, postprocessing_utils
 
 
 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)
 
+
 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))
     shutil.rmtree(user_dir)
 
-def reset_image(predictor, img):
+
+def preprocess_image(image: Image.Image) -> Image.Image:
     """
-    上传图像后调用：
-      - 重置 predictor，
-      - 设置 predictor 的输入图像，
-      - 返回原图
+    Preprocess the input image.
+    Args:
+        image (Image.Image): The input image.
+    Returns:
+        Image.Image: The preprocessed image.
     """
-    predictor.set_image(img)
-    original_img = img.copy()
-    # 返回predictor，visible occlusion mask初始化, 原始图像
-    return predictor, original_img, "The models are ready."
-
-def button_clickable(selected_points):
-    if len(selected_points) > 0:
-        return gr.Button.update(interactive=True)
-    else:
-        return gr.Button.update(interactive=False)
+    processed_image = pipeline.preprocess_image(image)
+    return processed_image
+
 
-def run_sam(predictor, selected_points):
+def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
     """
-    调用 SAM 模型进行分割。
+    Preprocess a list of input images.
+    
+    Args:
+        images (List[Tuple[Image.Image, str]]): The input images.
+        
+    Returns:
+        List[Image.Image]: The preprocessed images.
     """
-    # predictor.set_image(image)
-    if len(selected_points) == 0:
-        return [], None
-    input_points = [p for p in selected_points]
-    input_labels = [1 for _ in range(len(selected_points))]
-    masks, _, _ = predictor.predict(
-        point_coords=np.array(input_points),
-        point_labels=np.array(input_labels),
-        multimask_output=False,  # 单对象输出
+    images = [image[0] for image in images]
+    processed_images = [pipeline.preprocess_image(image) for image in images]
+    return processed_images
+
+
+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'],
     )
-    best_mask = masks[0].astype(np.uint8)
-    # dilate
-    if len(selected_points) > 1:
-        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
-        best_mask = cv2.dilate(best_mask, kernel, iterations=1)
-        best_mask = cv2.erode(best_mask, kernel, iterations=1)
-    return best_mask
-
-def apply_mask_overlay(image, mask, color=(255, 0, 0)):
-    """
-    在原图上叠加 mask：使用红色绘制 mask 的轮廓，非 mask 区域叠加浅灰色半透明遮罩。
-    """
-    img_arr = image
-    overlay = img_arr.copy()
-    gray_color = np.array([200, 200, 200], dtype=np.uint8)
-    non_mask = mask == 0
-    overlay[non_mask] = (0.5 * overlay[non_mask] + 0.5 * gray_color).astype(np.uint8)
-    contours, _ = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
-    cv2.drawContours(overlay, contours, -1, color, 2)
-    return overlay
-
-def segment_and_overlay(image, points, sam_predictor):
+    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:
     """
-    调用 run_sam 获得 mask，然后叠加显示分割结果。
+    Get the random seed.
     """
-    visible_mask = run_sam(sam_predictor, points)
-    overlaid = apply_mask_overlay(image, visible_mask * 255)
-    return overlaid, visible_mask
+    return np.random.randint(0, MAX_SEED) if randomize_seed else seed
 
 
 @spaces.GPU
 def image_to_3d(
-    images: np.ndarray,
+    image: Image.Image,
+    multiimages: List[Tuple[Image.Image, str]],
+    is_multiimage: bool,
     seed: int,
     ss_guidance_strength: float,
     ss_sampling_steps: int,
     slat_guidance_strength: float,
     slat_sampling_steps: int,
+    multiimage_algo: Literal["multidiffusion", "stochastic"],
     req: gr.Request,
 ) -> Tuple[dict, str]:
     """
@@ -122,21 +135,37 @@ def image_to_3d(
         str: The path to the video of the 3D model.
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
-    outputs = pipeline.run_multi_image(
-        [Image.fromarray(images)],
-        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="stochastic",
-    )
+    if not is_multiimage:
+        outputs = pipeline.run(
+            image,
+            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,
+            },
+        )
+    else:
+        outputs = pipeline.run_multi_image(
+            [image[0] for image in multiimages],
+            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,
+        )
     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))]
@@ -153,9 +182,15 @@ def extract_glb(
     mesh_simplify: float,
     texture_size: int,
     req: gr.Request,
-) -> tuple:
+) -> Tuple[str, str]:
     """
-    从生成的 3D 模型中提取 GLB 文件。
+    Extract a GLB file from the 3D model.
+    Args:
+        state (dict): The state of the generated 3D model.
+        mesh_simplify (float): The mesh simplification factor.
+        texture_size (int): The texture resolution.
+    Returns:
+        str: The path to the extracted GLB file.
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     gs, mesh = unpack_state(state)
@@ -167,9 +202,13 @@ def extract_glb(
 
 
 @spaces.GPU
-def extract_gaussian(state: dict, req: gr.Request) -> tuple:
+def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
     """
-    从生成的 3D 模型中提取 Gaussian 文件。
+    Extract a Gaussian file from the 3D model.
+    Args:
+        state (dict): The state of the generated 3D model.
+    Returns:
+        str: The path to the extracted Gaussian file.
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     gs, _ = unpack_state(state)
@@ -179,273 +218,59 @@ def extract_gaussian(state: dict, req: gr.Request) -> tuple:
     return gaussian_path, gaussian_path
 
 
-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:
-    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_sam_predictor():
-    # sam_checkpoint = hf_hub_download("ybelkada/segment-anything", "checkpoints/sam_vit_h_4b8939.pth")
-    # model_type = "vit_h"
-    # sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
-    # sam_predictor = SamPredictor(sam)
-    # return sam_predictor
-    return predictor
-
-
-def draw_points_on_image(image, point):
-    """在图像上绘制所有点，points 为 [(x, y, point_type), ...]"""
-    image_with_points = image.copy()
-    x, y = point
-    color = (255, 0, 0)
-    cv2.circle(image_with_points, (int(x), int(y)), radius=10, color=color, thickness=-1)
-    return image_with_points
-
-
-def see_point(image, x, y):
-    """
-    see操作：不修改 points 列表，仅在图像上临时显示这个点，
-    并返回更新后的图像和当前列表（不更新）。
-    """
-    # 复制当前列表，并在副本中加上新点（仅用于显示）
-    updated_image = draw_points_on_image(image, [x,y])
-    return updated_image
-
-def add_point(x, y, visible_points):
-    """
-    add操作：将新点添加到 points 列表中，
-    并返回更新后的图像和新的点列表。
-    """
-    if [x, y] not in visible_points:
-        visible_points.append([x, y])
-    return visible_points
-
-def delete_point(visible_points):
-    """
-    delete操作：删除 points 列表中的最后一个点，
-    并返回更新后的图像和新的点列表。
-    """
-    visible_points.pop()
-    return visible_points
-
+def prepare_multi_example() -> List[Image.Image]:
+    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, 4):
+            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))
+        images.append(Image.fromarray(np.concatenate(_images, axis=1)))
+    return images
 
-def clear_all_points(image):
-    """
-    清除所有点：返回原图、空的 visible 和 occlusion 列表，
-    以及更新后的点文本信息和空下拉菜单列表。
-    """
-    updated_image = image.copy()
-    return updated_image
 
-def see_visible_points(image, visible_points):
+def split_image(image: Image.Image) -> List[Image.Image]:
     """
-    在图像上绘制所有 visible 点（红色）。
+    Split an image into multiple views.
     """
-    updated_image = image.copy()
-    for p in visible_points:
-        cv2.circle(updated_image, (int(p[0]), int(p[1])), radius=10, color=(255, 0, 0), thickness=-1)
-    return updated_image
-
-def update_all_points(visible_points):
-    text = f"Points: {visible_points}"
-    visible_dropdown_choices = [f"({p[0]}, {p[1]})" for p in visible_points]
-    # 返回更新字典来明确设置 choices 和 value
-    return text, gr.Dropdown(label="Select Point to Delete", choices=visible_dropdown_choices, value=None, interactive=True)
-
-def delete_selected_visible(image, visible_points, selected_value):
-    # selected_value 是类似 "(x, y)" 的字符串
-    try:
-        selected_index = [f"({p[0]}, {p[1]})" for p in visible_points].index(selected_value)
-    except ValueError:
-        selected_index = None
-    if selected_index is not None and 0 <= selected_index < len(visible_points):
-        visible_points.pop(selected_index)
-    updated_image = image.copy()
-    # 重新绘制所有 visible 点（红色）
-    for p in visible_points:
-        cv2.circle(updated_image, (int(p[0]), int(p[1])), radius=10, color=(255, 0, 0), thickness=-1)
-    updated_text, vis_dropdown = update_all_points(visible_points)
-    return updated_image, visible_points, updated_text, vis_dropdown
-
-def add_mask(mask, mask_list):
-    # check if the mask if same as the last mask in the list 
-    if len(mask_list) > 0:
-        if np.array_equal(mask, mask_list[-1]):
-            return mask_list
-    mask_list.append(mask)
-    return mask_list
-
-def vis_mask(image, mask_list):
-    updated_image = image.copy()
-    # combine all the mask:
-    combined_mask = np.zeros_like(updated_image[:, :, 0])
-    for mask in mask_list:
-        combined_mask = cv2.bitwise_or(combined_mask, mask)
-    # overlay the mask on the image
-    updated_image = apply_mask_overlay(updated_image, combined_mask)
-    return updated_image
-
-def delete_mask(mask_list):
-    if len(mask_list) > 0:
-        mask_list.pop()
-    return mask_list
-
-def check_combined_mask(image, visibility_mask, mask_list, scale=0.6):
-    updated_image = image.copy()
-    # combine all the mask:
-    combined_mask = np.zeros_like(updated_image[:, :, 0])
-    occluded_mask = np.zeros_like(updated_image[:, :, 0])
-    if len(mask_list) == 0:
-        combined_mask = visibility_mask
-    else:
-        for mask in mask_list:
-            combined_mask = cv2.bitwise_or(combined_mask, mask)
-
-    if len(mask_list) > 1:
-        kernel = np.ones((5, 5), np.uint8)
-        dilate_iterations = 1
-        combined_mask = cv2.dilate(combined_mask, kernel, iterations=dilate_iterations)
-        combined_mask = cv2.erode(combined_mask, kernel, iterations=dilate_iterations)
-    
-    masked_img = updated_image * combined_mask[:, :, None]
-    occluded_mask[combined_mask == 1] = 127
-
-    # move the visible part to the center of the image
-    x, y, w, h = cv2.boundingRect(combined_mask.astype(np.uint8))
-    cropped_occluded_mask = (occluded_mask[y:y+h, x:x+w]).astype(np.uint8)
-    cropped_img = masked_img[y:y+h, x:x+w]
-
-    target_size = 512
-    scale_factor = target_size / max(w, h)
-    new_w = int(round(w * scale_factor * scale))
-    new_h = int(round(h * scale_factor * scale))
-
-    resized_occluded_mask = cv2.resize(cropped_occluded_mask.astype(np.uint8), (new_w, new_h), cv2.INTER_NEAREST)
-    resized_img = cv2.resize(cropped_img, (new_w, new_h), cv2.INTER_NEAREST)
-
-    final_img = np.zeros((target_size, target_size, 3), dtype=updated_image.dtype)
-    final_occluded_mask = np.zeros((target_size, target_size), dtype=np.uint8)
-
-    x_offset = (target_size - new_w) // 2
-    y_offset = (target_size - new_h) // 2
-
-    final_img[y_offset:y_offset+new_h, x_offset:x_offset+new_w] = resized_img
-    final_occluded_mask[y_offset:y_offset+new_h, x_offset:x_offset+new_w] = resized_occluded_mask
-
-    return final_img, occluded_mask
-
-
-
-def get_seed(randomize_seed: bool, seed: int) -> int:
-    """
-    Get the random seed.
-    """
-    return np.random.randint(0, MAX_SEED) if randomize_seed else seed
+    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]
 
 
 with gr.Blocks(delete_cache=(600, 600)) as demo:
     gr.Markdown("""
-    ## 3D Amodal Reconstruction with [Amodal3R](https://sm0kywu.github.io/Amodal3R/)          
+    ## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
+    * Upload an image and click "Generate" to create a 3D asset. If the image has alpha channel, it be used as the mask. Otherwise, we use `rembg` to remove the background.
+    * If you find the generated 3D asset satisfactory, click "Extract GLB" to extract the GLB file and download it.
+    
+    ✨New: 1) Experimental multi-image support. 2) Gaussian file extraction.
     """)
-
-     # 定义各状态变量
-    predictor = gr.State(value=get_sam_predictor())
-    visible_points_state = gr.State(value=[])
-    occlusion_points_state = gr.State(value=[])
-    original_image = gr.State(value=None)
-    visibility_mask = gr.State(value=None)
-    visibility_mask_list = gr.State(value=[])
-
-    occluded_mask = gr.State(value=None)
-    output_buf = gr.State()
-
-
-    with gr.Row():
-        gr.Markdown("""* Step 1 - Generate Visibility Mask and Occlusion Mask.
-        * Please wait for a few seconds after uploading the image. The 2D segmenter is getting ready.
-        * Add the point prompts to indicate the target object and occluders separately.
-        * "Render Point", see the position of the point to be added.
-        * "Add Point", the point will be added to the list.
-        * "Generate mask", see the segmented area corresponding to current point list.
-        * "Add mask", current mask will be added for 3D amodal completion.
-        """)
-    with gr.Row():
-        with gr.Column():
-            input_image = gr.Image(type="numpy", label='Input Occlusion Image', sources="upload", height=300)
-            with gr.Row():
-                message = gr.Markdown("Please wait a few seconds after uploading the image.", label="Message")  # 用于显示提示信息
-            with gr.Row():
-                x_input = gr.Number(label="X Coordinate", value=0)
-                y_input = gr.Number(label="Y Coordinate", value=0)
-            with gr.Row():
-                see_button = gr.Button("Render Point")
-                add_button = gr.Button("Add Point")
-            with gr.Row():
-                clear_button = gr.Button("Clear Points")
-                see_visible_button = gr.Button("Render Added Points")
-            with gr.Row():
-                # 新增文本框实时显示点列表
-                points_text = gr.Textbox(label="Points List", interactive=False)
-            with gr.Row():
-                # 新增下拉菜单，用户可选择需要删除的点
-                visible_points_dropdown = gr.Dropdown(label="Select Point to Delete", choices=[], value=None, interactive=True)
-                delete_visible_button = gr.Button("Delete Selected Visible")
-        with gr.Column():
-            # 用于显示 SAM 分割结果
-            visible_mask = gr.Image(label='Visible Mask', interactive=False, height=300)
-            with gr.Row():
-                gen_vis_mask = gr.Button("Generate Mask")
-                add_vis_mask = gr.Button("Add Mask")
-            with gr.Row():
-                render_vis_mask = gr.Button("Render Mask")
-                undo_vis_mask = gr.Button("Undo Last Mask")
-            vis_input = gr.Image(label='Visible Input', interactive=False, height=300)
-            with gr.Row():
-                zoom_scale = gr.Slider(0.3, 1.0, label="Target Object Scale", value=0.6, step=0.1)
-                check_visible_input = gr.Button("Generate Occluded Input")
-    with gr.Row():
-        gr.Markdown("""* Step 2 - 3D Amodal Completion.
-        * Different random seeds can be tried in "Generation Settings", if you think the results are not ideal.
-        * If the reconstruction 3D asset is satisfactory, you can extract the GLB file and download it.
-        """)
+    
     with gr.Row():
         with gr.Column():
-            with gr.Accordion(label="Generation Settings", open=True):
-                seed = gr.Slider(0, MAX_SEED, label="Seed", value=1, step=1)
+            with gr.Tabs() as input_tabs:
+                with gr.Tab(label="Single Image", id=0) as single_image_input_tab:
+                    image_prompt = gr.Image(label="Image Prompt", format="png", image_mode="RGBA", type="pil", height=300)
+                with gr.Tab(label="Multiple Images", id=1) as multiimage_input_tab:
+                    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. 
+                        
+                        *NOTE: this is an experimental algorithm without training a specialized model. It may not produce the best results for all images, especially those having different poses or inconsistent details.*
+                    """)
+            
+            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=True)
                 gr.Markdown("Stage 1: Sparse Structure Generation")
                 with gr.Row():
@@ -455,114 +280,127 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
                 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="stochastic")
+
             generate_btn = gr.Button("Generate")
+            
+            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)
+            
+            with gr.Row():
+                extract_glb_btn = gr.Button("Extract GLB", interactive=False)
+                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="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)  
     
-    # # Handlers
+    is_multiimage = gr.State(False)
+    output_buf = gr.State()
+
+    # Example images at the bottom of the page
+    with gr.Row() as single_image_example:
+        examples = gr.Examples(
+            examples=[
+                f'assets/example_image/{image}'
+                for image in os.listdir("assets/example_image")
+            ],
+            inputs=[image_prompt],
+            fn=preprocess_image,
+            outputs=[image_prompt],
+            run_on_click=True,
+            examples_per_page=64,
+        )
+    with gr.Row(visible=False) 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)
-
-    # ---------------------------
-    # 原有交互逻辑（略）
-    # ---------------------------
-    input_image.upload(
-        reset_image,
-        [predictor, input_image],
-        [predictor, original_image, message],
-    )
-    see_button.click(
-        see_point, 
-        inputs=[original_image, x_input, y_input], 
-        outputs=[input_image]
-    )
-    add_button.click(
-        add_point, 
-        inputs=[x_input, y_input, visible_points_state], 
-        outputs=[visible_points_state]
-    )
     
-    # ---------------------------
-    # 新增的交互逻辑
-    # ---------------------------
-    clear_button.click(
-        clear_all_points,
-        inputs=[original_image],
-        outputs=[input_image]
+    single_image_input_tab.select(
+        lambda: tuple([False, gr.Row.update(visible=True), gr.Row.update(visible=False)]),
+        outputs=[is_multiimage, single_image_example, multiimage_example]
     )
-    see_visible_button.click(
-        see_visible_points,
-        inputs=[input_image, visible_points_state],
-        outputs=input_image
+    multiimage_input_tab.select(
+        lambda: tuple([True, gr.Row.update(visible=False), gr.Row.update(visible=True)]),
+        outputs=[is_multiimage, single_image_example, multiimage_example]
     )
-    # 当 visible_points_state 或 occlusion_points_state 变化时，更新文本框和下拉菜单
-    visible_points_state.change(
-        update_all_points,
-        inputs=[visible_points_state],
-        outputs=[points_text, visible_points_dropdown]
+    
+    image_prompt.upload(
+        preprocess_image,
+        inputs=[image_prompt],
+        outputs=[image_prompt],
     )
-    delete_visible_button.click(
-        delete_selected_visible,
-        inputs=[input_image, visible_points_state, visible_points_dropdown],
-        outputs=[input_image, visible_points_state, points_text, visible_points_dropdown]
+    multiimage_prompt.upload(
+        preprocess_images,
+        inputs=[multiimage_prompt],
+        outputs=[multiimage_prompt],
     )
 
-    # 生成mask的逻辑
-    gen_vis_mask.click(
-        segment_and_overlay,
-        inputs=[original_image, visible_points_state, predictor],
-        outputs=[visible_mask, visibility_mask]
-    )
-    add_vis_mask.click(
-        add_mask,
-        inputs=[visibility_mask, visibility_mask_list],
-        outputs=[visibility_mask_list]
-    )
-    render_vis_mask.click(
-        vis_mask,
-        inputs=[original_image, visibility_mask_list],
-        outputs=[visible_mask]
+    generate_btn.click(
+        get_seed,
+        inputs=[randomize_seed, seed],
+        outputs=[seed],
+    ).then(
+        image_to_3d,
+        inputs=[image_prompt, multiimage_prompt, is_multiimage, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo],
+        outputs=[output_buf, video_output],
+    ).then(
+        lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
+        outputs=[extract_glb_btn, extract_gs_btn],
     )
-    undo_vis_mask.click(
-        delete_mask,
-        inputs=[visibility_mask_list],
-        outputs=[visibility_mask_list]
+
+    video_output.clear(
+        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
+        outputs=[extract_glb_btn, extract_gs_btn],
     )
 
-    check_visible_input.click(
-        check_combined_mask,
-        inputs=[original_image, visibility_mask, visibility_mask_list, zoom_scale],
-        outputs=[vis_input, occluded_mask]
+    extract_glb_btn.click(
+        extract_glb,
+        inputs=[output_buf, mesh_simplify, texture_size],
+        outputs=[model_output, download_glb],
+    ).then(
+        lambda: gr.Button(interactive=True),
+        outputs=[download_glb],
     )
     
+    extract_gs_btn.click(
+        extract_gaussian,
+        inputs=[output_buf],
+        outputs=[model_output, download_gs],
+    ).then(
+        lambda: gr.Button(interactive=True),
+        outputs=[download_gs],
+    )
 
-    # 3D Amodal Reconstruction
-    # generate_btn.click(
-    #     get_seed,
-    #     inputs=[randomize_seed, seed],
-    #     outputs=[seed],
-    # ).then(
-    #     image_to_3d,
-    #     inputs=[vis_input, occluded_mask, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
-    #     outputs=[output_buf, video_output],
-    # )
-
-    generate_btn.click(
-        image_to_3d,
-        inputs=[vis_input, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
-        outputs=[output_buf, video_output],
+    model_output.clear(
+        lambda: gr.Button(interactive=False),
+        outputs=[download_glb],
     )
     
 
-# 启动 Gradio App
+# Launch the Gradio app
 if __name__ == "__main__":
-    # pipeline = Amodal3RImageTo3DPipeline.from_pretrained("Sm0kyWu/Amodal3R")
     pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")
     pipeline.cuda()
-    predictor = get_sam_predictor()
-    predictor = predictor.cuda()
     try:
-        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))
+        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))    # Preload rembg
     except:
         pass
-    demo.launch()
+    demo.launch()