souging/TRELLIS_TextTo3D

1import gradio as gr2import spaces3from gradio_litmodel3d import LitModel3D4 5import os6import shutil7os.environ['SPCONV_ALGO'] = 'native'8from typing import *9import torch10import numpy as np11import imageio12from easydict import EasyDict as edict13from PIL import Image14from trellis.pipelines import TrellisImageTo3DPipeline15from trellis.representations import Gaussian, MeshExtractResult16from trellis.utils import render_utils, postprocessing_utils17 18 19MAX_SEED = np.iinfo(np.int32).max20TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')21os.makedirs(TMP_DIR, exist_ok=True)22 23 24def start_session(req: gr.Request):25    user_dir = os.path.join(TMP_DIR, str(req.session_hash))26    os.makedirs(user_dir, exist_ok=True)27    28    29def end_session(req: gr.Request):30    user_dir = os.path.join(TMP_DIR, str(req.session_hash))31    shutil.rmtree(user_dir)32 33 34def preprocess_image(image: Image.Image) -> Image.Image:35    """36    Preprocess the input image.37 38    Args:39        image (Image.Image): The input image.40 41    Returns:42        Image.Image: The preprocessed image.43    """44    processed_image = pipeline.preprocess_image(image)45    return processed_image46 47 48def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:49    """50    Preprocess a list of input images.51    52    Args:53        images (List[Tuple[Image.Image, str]]): The input images.54        55    Returns:56        List[Image.Image]: The preprocessed images.57    """58    images = [image[0] for image in images]59    processed_images = [pipeline.preprocess_image(image) for image in images]60    return processed_images61 62 63def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:64    return {65        'gaussian': {66            **gs.init_params,67            '_xyz': gs._xyz.cpu().numpy(),68            '_features_dc': gs._features_dc.cpu().numpy(),69            '_scaling': gs._scaling.cpu().numpy(),70            '_rotation': gs._rotation.cpu().numpy(),71            '_opacity': gs._opacity.cpu().numpy(),72        },73        'mesh': {74            'vertices': mesh.vertices.cpu().numpy(),75            'faces': mesh.faces.cpu().numpy(),76        },77    }78    79    80def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:81    gs = Gaussian(82        aabb=state['gaussian']['aabb'],83        sh_degree=state['gaussian']['sh_degree'],84        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],85        scaling_bias=state['gaussian']['scaling_bias'],86        opacity_bias=state['gaussian']['opacity_bias'],87        scaling_activation=state['gaussian']['scaling_activation'],88    )89    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')90    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')91    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')92    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')93    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')94    95    mesh = edict(96        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),97        faces=torch.tensor(state['mesh']['faces'], device='cuda'),98    )99    100    return gs, mesh101 102 103def get_seed(randomize_seed: bool, seed: int) -> int:104    """105    Get the random seed.106    """107    return np.random.randint(0, MAX_SEED) if randomize_seed else seed108 109 110@spaces.GPU111def image_to_3d(112    image: Image.Image,113    multiimages: List[Tuple[Image.Image, str]],114    is_multiimage: bool,115    seed: int,116    ss_guidance_strength: float,117    ss_sampling_steps: int,118    slat_guidance_strength: float,119    slat_sampling_steps: int,120    multiimage_algo: Literal["multidiffusion", "stochastic"],121    req: gr.Request,122) -> Tuple[dict, str]:123    """124    Convert an image to a 3D model.125 126    Args:127        image (Image.Image): The input image.128        multiimages (List[Tuple[Image.Image, str]]): The input images in multi-image mode.129        is_multiimage (bool): Whether is in multi-image mode.130        seed (int): The random seed.131        ss_guidance_strength (float): The guidance strength for sparse structure generation.132        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.133        slat_guidance_strength (float): The guidance strength for structured latent generation.134        slat_sampling_steps (int): The number of sampling steps for structured latent generation.135        multiimage_algo (Literal["multidiffusion", "stochastic"]): The algorithm for multi-image generation.136 137    Returns:138        dict: The information of the generated 3D model.139        str: The path to the video of the 3D model.140    """141    user_dir = os.path.join(TMP_DIR, str(req.session_hash))142    if not is_multiimage:143        outputs = pipeline.run(144            image,145            seed=seed,146            formats=["gaussian", "mesh"],147            preprocess_image=False,148            sparse_structure_sampler_params={149                "steps": ss_sampling_steps,150                "cfg_strength": ss_guidance_strength,151            },152            slat_sampler_params={153                "steps": slat_sampling_steps,154                "cfg_strength": slat_guidance_strength,155            },156        )157    else:158        outputs = pipeline.run_multi_image(159            [image[0] for image in multiimages],160            seed=seed,161            formats=["gaussian", "mesh"],162            preprocess_image=False,163            sparse_structure_sampler_params={164                "steps": ss_sampling_steps,165                "cfg_strength": ss_guidance_strength,166            },167            slat_sampler_params={168                "steps": slat_sampling_steps,169                "cfg_strength": slat_guidance_strength,170            },171            mode=multiimage_algo,172        )173    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']174    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']175    video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]176    video_path = os.path.join(user_dir, 'sample.mp4')177    imageio.mimsave(video_path, video, fps=15)178    state = pack_state(outputs['gaussian'][0], outputs['mesh'][0])179    torch.cuda.empty_cache()180    return state, video_path181 182 183@spaces.GPU(duration=90)184def extract_glb(185    state: dict,186    mesh_simplify: float,187    texture_size: int,188    req: gr.Request,189) -> Tuple[str, str]:190    """191    Extract a GLB file from the 3D model.192 193    Args:194        state (dict): The state of the generated 3D model.195        mesh_simplify (float): The mesh simplification factor.196        texture_size (int): The texture resolution.197 198    Returns:199        str: The path to the extracted GLB file.200    """201    user_dir = os.path.join(TMP_DIR, str(req.session_hash))202    gs, mesh = unpack_state(state)203    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)204    glb_path = os.path.join(user_dir, 'sample.glb')205    glb.export(glb_path)206    torch.cuda.empty_cache()207    return glb_path, glb_path208 209 210@spaces.GPU211def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:212    """213    Extract a Gaussian file from the 3D model.214 215    Args:216        state (dict): The state of the generated 3D model.217 218    Returns:219        str: The path to the extracted Gaussian file.220    """221    user_dir = os.path.join(TMP_DIR, str(req.session_hash))222    gs, _ = unpack_state(state)223    gaussian_path = os.path.join(user_dir, 'sample.ply')224    gs.save_ply(gaussian_path)225    torch.cuda.empty_cache()226    return gaussian_path, gaussian_path227 228 229def prepare_multi_example() -> List[Image.Image]:230    multi_case = list(set([i.split('_')[0] for i in os.listdir("assets/example_multi_image")]))231    images = []232    for case in multi_case:233        _images = []234        for i in range(1, 4):235            img = Image.open(f'assets/example_multi_image/{case}_{i}.png')236            W, H = img.size237            img = img.resize((int(W / H * 512), 512))238            _images.append(np.array(img))239        images.append(Image.fromarray(np.concatenate(_images, axis=1)))240    return images241 242 243def split_image(image: Image.Image) -> List[Image.Image]:244    """245    Split an image into multiple views.246    """247    image = np.array(image)248    alpha = image[..., 3]249    alpha = np.any(alpha>0, axis=0)250    start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()251    end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()252    images = []253    for s, e in zip(start_pos, end_pos):254        images.append(Image.fromarray(image[:, s:e+1]))255    return [preprocess_image(image) for image in images]256 257 258with gr.Blocks(delete_cache=(600, 600)) as demo:259    gr.Markdown("""260    ## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)261    * 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.262    * If you find the generated 3D asset satisfactory, click "Extract GLB" to extract the GLB file and download it.263    264    ✨New: 1) Experimental multi-image support. 2) Gaussian file extraction.265    """)266    267    with gr.Row():268        with gr.Column():269            with gr.Tabs() as input_tabs:270                with gr.Tab(label="Single Image", id=0) as single_image_input_tab:271                    image_prompt = gr.Image(label="Image Prompt", format="png", image_mode="RGBA", type="pil", height=300)272                with gr.Tab(label="Multiple Images", id=1) as multiimage_input_tab:273                    multiimage_prompt = gr.Gallery(label="Image Prompt", format="png", type="pil", height=300, columns=3)274                    gr.Markdown("""275                        Input different views of the object in separate images. 276                        277                        *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.*278                    """)279            280            with gr.Accordion(label="Generation Settings", open=False):281                seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)282                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)283                gr.Markdown("Stage 1: Sparse Structure Generation")284                with gr.Row():285                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)286                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)287                gr.Markdown("Stage 2: Structured Latent Generation")288                with gr.Row():289                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)290                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)291                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Multi-image Algorithm", value="stochastic")292 293            generate_btn = gr.Button("Generate")294            295            with gr.Accordion(label="GLB Extraction Settings", open=False):296                mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)297                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)298            299            with gr.Row():300                extract_glb_btn = gr.Button("Extract GLB", interactive=False)301                extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)302            gr.Markdown("""303                        *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*304                        """)305 306        with gr.Column():307            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)308            model_output = LitModel3D(label="Extracted GLB/Gaussian", exposure=10.0, height=300)309            310            with gr.Row():311                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)312                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  313    314    is_multiimage = gr.State(False)315    output_buf = gr.State()316 317    # Example images at the bottom of the page318    with gr.Row() as single_image_example:319        examples = gr.Examples(320            examples=[321                f'assets/example_image/{image}'322                for image in os.listdir("assets/example_image")323            ],324            inputs=[image_prompt],325            fn=preprocess_image,326            outputs=[image_prompt],327            run_on_click=True,328            examples_per_page=64,329        )330    with gr.Row(visible=False) as multiimage_example:331        examples_multi = gr.Examples(332            examples=prepare_multi_example(),333            inputs=[image_prompt],334            fn=split_image,335            outputs=[multiimage_prompt],336            run_on_click=True,337            examples_per_page=8,338        )339 340    # Handlers341    demo.load(start_session)342    demo.unload(end_session)343    344    single_image_input_tab.select(345        lambda: tuple([False, gr.Row.update(visible=True), gr.Row.update(visible=False)]),346        outputs=[is_multiimage, single_image_example, multiimage_example]347    )348    multiimage_input_tab.select(349        lambda: tuple([True, gr.Row.update(visible=False), gr.Row.update(visible=True)]),350        outputs=[is_multiimage, single_image_example, multiimage_example]351    )352    353    image_prompt.upload(354        preprocess_image,355        inputs=[image_prompt],356        outputs=[image_prompt],357    )358    multiimage_prompt.upload(359        preprocess_images,360        inputs=[multiimage_prompt],361        outputs=[multiimage_prompt],362    )363 364    generate_btn.click(365        get_seed,366        inputs=[randomize_seed, seed],367        outputs=[seed],368    ).then(369        image_to_3d,370        inputs=[image_prompt, multiimage_prompt, is_multiimage, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo],371        outputs=[output_buf, video_output],372    ).then(373        lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),374        outputs=[extract_glb_btn, extract_gs_btn],375    )376 377    video_output.clear(378        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),379        outputs=[extract_glb_btn, extract_gs_btn],380    )381 382    extract_glb_btn.click(383        extract_glb,384        inputs=[output_buf, mesh_simplify, texture_size],385        outputs=[model_output, download_glb],386    ).then(387        lambda: gr.Button(interactive=True),388        outputs=[download_glb],389    )390    391    extract_gs_btn.click(392        extract_gaussian,393        inputs=[output_buf],394        outputs=[model_output, download_gs],395    ).then(396        lambda: gr.Button(interactive=True),397        outputs=[download_gs],398    )399 400    model_output.clear(401        lambda: gr.Button(interactive=False),402        outputs=[download_glb],403    )404    405 406# Launch the Gradio app407if __name__ == "__main__":408    pipeline = TrellisImageTo3DPipeline.from_pretrained("JeffreyXiang/TRELLIS-image-large")409    pipeline.cuda()410    try:411        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))    # Preload rembg412    except:413        pass414    demo.launch()415