app.py

import spaces
import logging
import os
import shlex
import subprocess
import tempfile
import time

import gradio as gr
import numpy as np
import rembg
import torch
from PIL import Image
from functools import partial

subprocess.run(shlex.split('pip install wheel/torchmcubes-0.1.0-cp310-cp310-linux_x86_64.whl'))

from tsr.system import TSR
from tsr.utils import remove_background, resize_foreground, to_gradio_3d_orientation


HEADER = """
# Real3D Demo

**Real3D** is a single-view large reconstruction model with self-training on real-world images. Project page & Paper: https://hwjiang1510.github.io/Real3D/

**Tips:**

**Important** We noticed the running usually fails if it shows "waiting for GPU". If the result shows as a cube, please re-run it.
1. If you find the result is unsatisfactory, please try to change the foreground ratio. It might improve the results.
2. It's better to disable "Remove Background" for the provided examples since they have already been preprocessed.
3. Otherwise, please disable "Remove Background" option only if your input image is RGBA with a transparent background, image contents are centered and occupy more than 70% of image width or height.
"""


if torch.cuda.is_available():
    device = "cuda:0"
else:
    device = "cpu"

model = TSR.from_pretrained(
    "hwjiang/Real3D",
    config_name="config.yaml",
    weight_name="model_both_trained_v1.ckpt",
)
model.renderer.set_chunk_size(131072)
model.to(device)

rembg_session = rembg.new_session()


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


def preprocess(input_image, do_remove_background, foreground_ratio):
    def fill_background(image):
        image = np.array(image).astype(np.float32) / 255.0
        image = image[:, :, :3] * image[:, :, 3:4] + (1 - image[:, :, 3:4]) * 0.5
        image = Image.fromarray((image * 255.0).astype(np.uint8))
        return image

    if do_remove_background:
        image = input_image.convert("RGB")
        image = remove_background(image, rembg_session)
        image = resize_foreground(image, foreground_ratio)
        image = fill_background(image)
    else:
        image = input_image
        if image.mode == "RGBA":
            image = fill_background(image)
    return image


@spaces.GPU
def generate(image, mc_resolution, formats=["obj", "glb"]):
    scene_codes = model(image, device=device)
    mesh = model.extract_mesh(scene_codes, resolution=mc_resolution)[0]
    mesh = to_gradio_3d_orientation(mesh)

    mesh_path_glb = tempfile.NamedTemporaryFile(suffix=f".glb", delete=False)
    mesh.export(mesh_path_glb.name)

    mesh_path_obj = tempfile.NamedTemporaryFile(suffix=f".obj", delete=False)
    mesh.apply_scale([-1, 1, 1])  # Otherwise the visualized .obj will be flipped
    mesh.export(mesh_path_obj.name)
    
    return mesh_path_obj.name, mesh_path_glb.name

def run_example(image_pil):
    preprocessed = preprocess(image_pil, True, 0.65)
    mesh_name_obj, mesh_name_glb = generate(preprocessed, 256, ["obj", "glb"])
    return preprocessed, mesh_name_obj, mesh_name_glb

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",
                    type="pil",
                    elem_id="content_image",
                )
                processed_image = gr.Image(label="Processed Image", interactive=False)
            with gr.Row():
                with gr.Group():
                    do_remove_background = gr.Checkbox(
                        label="Remove Background", value=True
                    )
                    foreground_ratio = gr.Slider(
                        label="Foreground Ratio",
                        minimum=0.5,
                        maximum=1.0,
                        value=0.65,
                        step=0.05,
                    )
                    mc_resolution = gr.Slider(
                        label="Marching Cubes Resolution",
                        minimum=32,
                        maximum=320,
                        value=256,
                        step=32
                     )
            with gr.Row():
                submit = gr.Button("Generate", elem_id="generate", variant="primary")
        with gr.Column():
            with gr.Tab("OBJ"):
                output_model_obj = gr.Model3D(
                    label="Output Model (OBJ Format)",
                    interactive=False,
                )
                gr.Markdown("Note: Downloaded object will be flipped in case of .obj export. 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(variant="panel"):
        gr.Examples(
            examples=[
                os.path.join("examples", img_name) for img_name in sorted(os.listdir("examples"))
            ],
            inputs=[input_image],
            outputs=[processed_image, output_model_obj, output_model_glb],
            cache_examples="lazy",
            fn=partial(run_example),
            label="Examples",
            examples_per_page=20
        )
    submit.click(fn=check_input_image, inputs=[input_image]).success(
        fn=preprocess,
        inputs=[input_image, do_remove_background, foreground_ratio],
        outputs=[processed_image],
    ).success(
        fn=generate,
        inputs=[processed_image, mc_resolution],
        outputs=[output_model_obj, output_model_glb],
    )

demo.queue(max_size=10)
demo.launch()