Update app.py

app.py

@@ -1,6 +1,4 @@
 import gradio as gr
-import spaces
-from gradio_litmodel3d import LitModel3D
 import os
 import shutil
 
@@ -14,6 +12,7 @@ from PIL import Image
 from trellis.pipelines import TrellisVGGTTo3DPipeline
 from trellis.representations import Gaussian, MeshExtractResult
 from trellis.utils import render_utils, postprocessing_utils
+from gradio_litmodel3d import LitModel3D
 
 MAX_SEED = np.iinfo(np.int32).max
 TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
@@ -25,43 +24,14 @@ def start_session(req: gr.Request):
 
 def end_session(req: gr.Request):
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
-    # 폴더가 존재할 때만 삭제하도록 수정 (FileNotFoundError 방지)
     if os.path.exists(user_dir):
         shutil.rmtree(user_dir)
 
-@spaces.GPU
 def preprocess_image(image: Image.Image) -> Image.Image:
-    """
-    Preprocess the input image for 3D generation.
-    
-    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.
-
-    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
 
-@spaces.GPU
 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 = []
@@ -75,25 +45,11 @@ def preprocess_videos(video: str) -> List[Tuple[Image.Image, str]]:
     processed_images = [pipeline.preprocess_image(image) for image in images]
     return processed_images
 
-@spaces.GPU
 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: Gaussian, mesh: MeshExtractResult) -> dict:
     return {
         'gaussian': {
@@ -109,8 +65,7 @@ def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
             'faces': mesh.faces.cpu().numpy(),
         },
     }
-    
-    
+
 def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
     gs = Gaussian(
         aabb=state['gaussian']['aabb'],
@@ -120,38 +75,22 @@ def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
         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')
+    # 추론 시 메인 장치인 cuda:0으로 데이터를 보냅니다.
+    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda:0')
+    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda:0')
+    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda:0')
+    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda:0')
+    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda:0')
     
     mesh = edict(
-        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
-        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
+        vertices=torch.tensor(state['mesh']['vertices'], device='cuda:0'),
+        faces=torch.tensor(state['mesh']['faces'], device='cuda:0'),
     )
-    
     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
 
-
-@spaces.GPU(duration=120)
 def generate_and_extract_glb(
     multiimages: List[Tuple[Image.Image, str]],
     seed: int,
@@ -164,32 +103,9 @@ def generate_and_extract_glb(
     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]
 
-    # Generate 3D model
     outputs, _, _ = pipeline.run(
         image=image_files,
         seed=seed,
@@ -206,49 +122,23 @@ def generate_and_extract_glb(
         mode=multiimage_algo,
     )
 
-    # 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"
-
     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)
     
-    # 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
 
-
-@spaces.GPU
 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')
@@ -256,15 +146,16 @@ def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
     torch.cuda.empty_cache()
     return gaussian_path, gaussian_path
 
-
 def prepare_multi_example() -> List[Image.Image]:
+    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')
+            path = f'assets/example_multi_image/{case}_{i}.png'
+            if os.path.exists(path):
+                img = Image.open(path)
                 W, H = img.size
                 img = img.resize((int(W / H * 512), 512))
                 _images.append(np.array(img))
@@ -272,21 +163,7 @@ def prepare_multi_example() -> List[Image.Image]:
             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)
@@ -297,168 +174,72 @@ def split_image(image: Image.Image) -> List[Image.Image]:
         images.append(Image.fromarray(image[:, s:e+1]))
     return [preprocess_image(image) for image in images]
 
-# Create interface
+# --- Gradio UI ---
 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; }
-    """
+    css=".slider .inner { width: 5px; background: #FFF; } .viewport { aspect-ratio: 4/3; }"
 )
+
 with demo:
-    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!✨
-    """)
-    
+    gr.Markdown("# 💻 ReconViaGen\n✨This demo is partial. Stay tuned!✨")
     with gr.Row():
         with gr.Column():
             with gr.Tabs() as input_tabs:
-                with gr.Tab(label="Input Video or Images", id=0) as multiimage_input_tab:
+                with gr.Tab(label="Input Video or Images", id=0):
                     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):
+                    image_prompt = gr.Image(label="Image Prompt", visible=False, type="pil", height=300)
+                    multiimage_prompt = gr.Gallery(label="Image Prompt", columns=3)
+            with gr.Accordion(label="Settings", open=False):
                 seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
                 randomize_seed = gr.Checkbox(label="Randomize Seed", value=False)
-                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)
-
+                ss_guidance_strength = gr.Slider(0.0, 10.0, label="SS Guidance", value=7.5)
+                ss_sampling_steps = gr.Slider(1, 50, label="SS Steps", value=30)
+                slat_guidance_strength = gr.Slider(0.0, 10.0, label="Slat Guidance", value=3.0)
+                slat_sampling_steps = gr.Slider(1, 50, label="Slat Steps", value=12)
+                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], value="multidiffusion")
+                mesh_simplify = gr.Slider(0.9, 0.98, value=0.95)
+                texture_size = gr.Slider(512, 2048, 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="Generated 3D Asset", autoplay=True, loop=True, height=300)
-            model_output = LitModel3D(label="Extracted GLB/Gaussian", exposure=10.0, height=300)
-            
+            video_output = gr.Video(label="Generated 3D Asset")
+            model_output = LitModel3D(label="Extracted GLB/Gaussian")
             with gr.Row():
-                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
-                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  
-    
+                download_glb = gr.DownloadButton("Download GLB", interactive=False)
+                download_gs = gr.DownloadButton("Download Gaussian", interactive=False)
     output_buf = gr.State()
+    
+    with gr.Row():
+        gr.Examples(examples=prepare_multi_example(), inputs=[image_prompt], fn=split_image, outputs=[multiimage_prompt], run_on_click=True)
 
-    # 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(
+    input_video.upload(preprocess_videos, inputs=[input_video], outputs=[multiimage_prompt])
+    input_video.clear(lambda: (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],
         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],
-    )
-    
+    ).then(lambda: (gr.Button(interactive=True), gr.Button(interactive=True)), outputs=[extract_gs_btn, download_glb])
+    extract_gs_btn.click(extract_gaussian, inputs=[output_buf], outputs=[model_output, download_gs])
 
+# --- 메인 실행부 (VRAM 분산 최적화) ---
 if __name__ == "__main__":
-    # 1. 모델 로드
+    # 모델 로드
     pipeline = TrellisVGGTTo3DPipeline.from_pretrained("Stable-X/trellis-vggt-v0-2")
     
-    # 2. 멀티 GPU 강제 설정
-    if torch.cuda.is_available():
-        device = torch.device("cuda:0")
-        # 모든 하위 모델들을 먼저 0번 GPU로 확실히 보냅니다
-        pipeline.to(device)
-        
-        num_gpus = torch.cuda.device_count()
-        if num_gpus > 1:
-            print(f"--- 멀티 GPU 활성화: {num_gpus}개의 GPU를 사용합니다 ---")
-            # 에러가 났던 birefnet과 주요 모델들을 DataParallel로 래핑
-            try:
-                if hasattr(pipeline, 'VGGT_model'):
-                    pipeline.VGGT_model = torch.nn.DataParallel(pipeline.VGGT_model).cuda()
-                if hasattr(pipeline, 'birefnet_model'):
-                    pipeline.birefnet_model = torch.nn.DataParallel(pipeline.birefnet_model).cuda()
-                if hasattr(pipeline, 'sparse_structure_decoder'):
-                    pipeline.sparse_structure_decoder = torch.nn.DataParallel(pipeline.sparse_structure_decoder).cuda()
-                if hasattr(pipeline, 'slat_decoder'):
-                    pipeline.slat_decoder = torch.nn.DataParallel(pipeline.slat_decoder).cuda()
-            except Exception as e:
-                print(f"멀티 GPU 설정 중 경고 발생(단일 GPU로 전환): {e}")
-                pipeline.to(device)
-    
-    # 3. 앱 실행
+    num_gpus = torch.cuda.device_count()
+    if num_gpus >= 4:
+        print(f"--- 4 GPUs Detected: Splitting Models to prevent VRAM Error ---")
+        # 모델의 각 부분을 서로 다른 GPU 메모리에 적재하여 1개 GPU의 부담을 줄임
+        pipeline.to("cuda:0") 
+        if hasattr(pipeline, 'VGGT_model'): pipeline.VGGT_model.to("cuda:1")
+        if hasattr(pipeline, 'birefnet_model'): pipeline.birefnet_model.to("cuda:2")
+        # 가장 무거운 디코더들은 3번 GPU로 격리
+        if hasattr(pipeline, 'slat_decoder'): pipeline.slat_decoder.to("cuda:3")
+        if hasattr(pipeline, 'sparse_structure_decoder'): pipeline.sparse_structure_decoder.to("cuda:3")
+    else:
+        pipeline.cuda()
+
     demo.launch()