app.py

import spaces

import os
import imageio
import numpy as np
import torch
import rembg
from PIL import Image
from torchvision.transforms import v2
from pytorch_lightning import seed_everything
from omegaconf import OmegaConf
from einops import rearrange, repeat
from tqdm import tqdm
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler

from src.utils.train_util import instantiate_from_config
from src.utils.camera_util import (
    FOV_to_intrinsics, 
    get_zero123plus_input_cameras,
    get_circular_camera_poses,
)
from src.utils.mesh_util import save_obj, save_glb
from src.utils.infer_util import remove_background, resize_foreground, images_to_video

import tempfile
from functools import partial

from huggingface_hub import hf_hub_download

import gradio as gr



def get_render_cameras(batch_size=1, M=120, radius=2.5, elevation=10.0, is_flexicubes=False):
    """
    Get the rendering camera parameters.
    """
    c2ws = get_circular_camera_poses(M=M, radius=radius, elevation=elevation)
    if is_flexicubes:
        cameras = torch.linalg.inv(c2ws)
        cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1, 1)
    else:
        extrinsics = c2ws.flatten(-2)
        intrinsics = FOV_to_intrinsics(50.0).unsqueeze(0).repeat(M, 1, 1).float().flatten(-2)
        cameras = torch.cat([extrinsics, intrinsics], dim=-1)
        cameras = cameras.unsqueeze(0).repeat(batch_size, 1, 1)
    return cameras


def images_to_video(images, output_path, fps=30):
    # images: (N, C, H, W)
    os.makedirs(os.path.dirname(output_path), exist_ok=True)
    frames = []
    for i in range(images.shape[0]):
        frame = (images[i].permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8).clip(0, 255)
        assert frame.shape[0] == images.shape[2] and frame.shape[1] == images.shape[3], \
            f"Frame shape mismatch: {frame.shape} vs {images.shape}"
        assert frame.min() >= 0 and frame.max() <= 255, \
            f"Frame value out of range: {frame.min()} ~ {frame.max()}"
        frames.append(frame)
    imageio.mimwrite(output_path, np.stack(frames), fps=fps, codec='h264')


###############################################################################
# Configuration.
###############################################################################

import shutil

def find_cuda():
    # Check if CUDA_HOME or CUDA_PATH environment variables are set
    cuda_home = os.environ.get('CUDA_HOME') or os.environ.get('CUDA_PATH')

    if cuda_home and os.path.exists(cuda_home):
        return cuda_home

    # Search for the nvcc executable in the system's PATH
    nvcc_path = shutil.which('nvcc')

    if nvcc_path:
        # Remove the 'bin/nvcc' part to get the CUDA installation path
        cuda_path = os.path.dirname(os.path.dirname(nvcc_path))
        return cuda_path

    return None

cuda_path = find_cuda()

if cuda_path:
    print(f"CUDA installation found at: {cuda_path}")
else:
    print("CUDA installation not found")

config_path = 'configs/instant-nerf-large.yaml'
config = OmegaConf.load(config_path)
config_name = os.path.basename(config_path).replace('.yaml', '')
model_config = config.model_config
infer_config = config.infer_config

IS_FLEXICUBES = True if config_name.startswith('instant-mesh') else False

device = torch.device('cuda')

# load diffusion model
print('Loading diffusion model ...')
pipeline = DiffusionPipeline.from_pretrained(
    "sudo-ai/zero123plus-v1.2", 
    custom_pipeline="zero123plus",
    torch_dtype=torch.float16,
)
pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
    pipeline.scheduler.config, timestep_spacing='trailing'
)

# load custom white-background UNet
unet_ckpt_path = hf_hub_download(repo_id="TencentARC/InstantMesh", filename="diffusion_pytorch_model.bin", repo_type="model")
state_dict = torch.load(unet_ckpt_path, map_location='cpu')
pipeline.unet.load_state_dict(state_dict, strict=True)

pipeline = pipeline.to(device)

# load reconstruction model
print('Loading reconstruction model ...')
model_ckpt_path = hf_hub_download(repo_id="TencentARC/InstantMesh", filename="instant_nerf_large.ckpt", repo_type="model")
model = instantiate_from_config(model_config)
state_dict = torch.load(model_ckpt_path, map_location='cpu')['state_dict']
state_dict = {k[14:]: v for k, v in state_dict.items() if k.startswith('lrm_generator.') and 'source_camera' not in k}
model.load_state_dict(state_dict, strict=True)

model = model.to(device)

print('Loading Finished!')


def check_input_image(input_image):
    if input_image is None:
        raise gr.Error("No image uploaded!")
    
def check_style_image(style_image):
    if style_image is None:
        raise gr.Error("No style image uploaded!")


def preprocess(input_image, do_remove_background):

    rembg_session = rembg.new_session() if do_remove_background else None

    if do_remove_background:
        input_image = remove_background(input_image, rembg_session)
        input_image = resize_foreground(input_image, 0.85)

    return input_image


@spaces.GPU
def generate_mvs(input_image, sample_steps, sample_seed):

    seed_everything(sample_seed)
    
    # sampling
    z123_image = pipeline(
        input_image, 
        num_inference_steps=sample_steps
    ).images[0]

    show_image = np.asarray(z123_image, dtype=np.uint8).copy()
    show_image = torch.from_numpy(show_image)     # (960, 640, 3)
    show_image = rearrange(show_image, '(n h) (m w) c -> (n m) h w c', n=3, m=2)
    show_image = rearrange(show_image, '(n m) h w c -> (n h) (m w) c', n=2, m=3)
    show_image = Image.fromarray(show_image.numpy())

    return z123_image, show_image

@spaces.GPU
def make3d(mv_images, style_image, alpha, style_layers):
    """
    mv_images: single multi-view image (pil or numpy)
    style_image: PIL image
    alpha: float
    style_layers: int
    """
    global model

    # Save the uploaded style image to a temporary file, so the model can read it from disk
    style_path = tempfile.NamedTemporaryFile(suffix=".png", delete=False).name
    style_image.save(style_path)

    if IS_FLEXICUBES:
        model.init_flexicubes_geometry(device, use_renderer=False)

    images = np.asarray(mv_images, dtype=np.float32) / 255.0
    images = torch.from_numpy(images).permute(2, 0, 1).contiguous().float()  # (3, 960, 640)
    images = rearrange(images, 'c (n h) (m w) -> (n m) c h w', n=3, m=2)     # (6, 3, 320, 320)

    input_cameras = get_zero123plus_input_cameras(batch_size=1, radius=4.0).to(device)
    render_cameras = get_render_cameras(
        batch_size=1,
        radius=2.5,
        is_flexicubes=IS_FLEXICUBES
    ).to(device)

    images = images.unsqueeze(0).to(device)
    images = v2.functional.resize(images, (320, 320), interpolation=3, antialias=True).clamp(0, 1)

    mesh_fpath = tempfile.NamedTemporaryFile(suffix=".obj", delete=False).name
    mesh_basename = os.path.basename(mesh_fpath).split('.')[0]
    mesh_dirname = os.path.dirname(mesh_fpath)
    mesh_glb_fpath = os.path.join(mesh_dirname, f"{mesh_basename}.glb")

    with torch.no_grad():
        # get triplane, now passing style_path, alpha, style_layers
        planes = model.forward_planes(
            images,
            input_cameras,
            style_path,
            float(alpha),
            int(style_layers),
        )

        # extract mesh
        mesh_out = model.extract_mesh(
            planes,
            use_texture_map=False,
            **infer_config,
        )

        vertices, faces, vertex_colors = mesh_out
        vertices = vertices[:, [1, 2, 0]]

        save_glb(vertices, faces, vertex_colors, mesh_glb_fpath)
        save_obj(vertices, faces, vertex_colors, mesh_fpath)
        print(f"Mesh saved to {mesh_fpath}")
    return mesh_fpath, mesh_glb_fpath

_HEADER_ = '''
<h2><b>3DStylizationLRM</b></h2>
This demo lets you provide a content image, a style image, an alpha blending value, and the number of style layers to inject. It will generate 3D geometry stylized accordingly.

❗️❗️❗️ **Notes:**
- Content image background can be removed automatically.
- Adjust the **Alpha** slider to control style blending strength.
- Adjust **Style Layers** to choose how many layers of style to inject.
'''

_CITE_ = r"""
If 3D Stylization LRM is helpful, please help to ⭐ the <a href='https://github.com/ipekoztas/3D-Stylization-LRM' target='_blank'>Github Repo</a>. Thanks!
---
📝 **Citation**

If you find our work useful for your research or applications, please cite using this bibtex:
```bibtex
@article{oztas20253dstylizationlargereconstruction,
    title={3D Stylization via Large Reconstruction Model}, 
    author={Ipek Oztas and Duygu Ceylan and Aysegul Dundar},
    journal={https://arxiv.org/abs/2504.21836},
    year={2025}
}
```
📋 **License**  
Apache-2.0 LICENSE.  
"""

with gr.Blocks() as demo:
    gr.Markdown(_HEADER_)
    with gr.Row(variant="panel"):
        with gr.Column():
            with gr.Row():
                input_image = gr.Image(
                    label="Input Image",
                    image_mode="RGBA",
                    sources="upload",
                    #width=256,
                    #height=256,
                    type="pil",
                    elem_id="content_image",
                )
                # Style Image Upload
                style_image = gr.Image(
                    label="Style Image",
                    image_mode="RGB",
                    type="pil",
                    elem_id="style_image",
                )
                processed_image = gr.Image(
                    label="Processed Image", 
                    image_mode="RGBA", 
                    #width=256,
                    #height=256,
                    type="pil", 
                    interactive=False
                )
            with gr.Row():
                with gr.Group():
                    do_remove_background = gr.Checkbox(
                        label="Remove Background", value=True
                    )
                    sample_seed = gr.Number(value=42, label="Seed Value", precision=0)

                    sample_steps = gr.Slider(
                        label="Sample Steps",
                        minimum=30,
                        maximum=75,
                        value=75,
                        step=5
                    )

            with gr.Row():
                alpha = gr.Slider(
                    label="Alpha Value",
                    minimum=0.0,
                    maximum=1.0,
                    value=0.7,
                    step=0.01,
                )
                style_layers = gr.Slider(
                    label="Style Layers",
                    minimum=1,
                    maximum=10,
                    value=4,
                    step=1,
                )

            with gr.Row():
                submit = gr.Button("Generate", elem_id="generate", variant="primary")

            with gr.Row(variant="panel"):
                with gr.Column():
                    gr.Examples(
                        examples=[
                            os.path.join("examples", img_name) for img_name in sorted(os.listdir("examples"))
                        ],
                        inputs=[input_image],
                        label="Examples",
                        cache_examples=False,
                        examples_per_page=16
                    )
                with gr.Column():
                    gr.Examples(
                        examples=[
                            os.path.join("styles", img_name) for img_name in sorted(os.listdir("styles"))
                        ],
                        inputs=[style_image],
                        label="Styles",
                        cache_examples=False,
                        examples_per_page=16
                    )

        with gr.Column():

            with gr.Row():

                with gr.Column():
                    mv_show_images = gr.Image(
                        label="Generated Multi-views",
                        type="pil",
                        width=379,
                        interactive=False
                    )

                # with gr.Column():
                #     output_video = gr.Video(
                #         label="video", format="mp4",
                #         width=379,
                #         autoplay=True,
                #         interactive=False
                #     )

            with gr.Row():
                with gr.Tab("OBJ"):
                    output_model_obj = gr.Model3D(
                        label="Output Model (OBJ Format)",
                        interactive=False,
                    )
                    gr.Markdown("Note: Downloaded .obj model will be flipped. Export .glb instead or manually flip it before usage.")
                with gr.Tab("GLB"):
                    output_model_glb = gr.Model3D(
                        label="Output Model (GLB Format)",
                        interactive=False,
                    )
                    gr.Markdown("Note: The model shown here has a darker appearance. Download to get correct results.")

            with gr.Row():
                gr.Markdown('''Try a different <b>seed value</b> if the result is unsatisfying (Default: 42).''')

    gr.Markdown(_CITE_)

    mv_images = gr.State()

    # Chain of actions:
    submit.click(fn=check_input_image, inputs=[input_image]).success(
        fn=check_style_image, inputs=[style_image]
    ).success(
        fn=preprocess,
        inputs=[input_image, do_remove_background],
        outputs=[processed_image],
    ).success(
        fn=generate_mvs,
        inputs=[processed_image, sample_steps, sample_seed],
        outputs=[mv_images, mv_show_images],
    ).success(
        fn=make3d,
        inputs=[mv_images, style_image, alpha, style_layers],
        outputs=[output_model_obj, output_model_glb],
    )

demo.launch()