app.py

import spaces
import subprocess
# Install flash attention, skipping CUDA build if necessary
subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
)
import os
import torch
import trimesh
from accelerate.utils import set_seed
from accelerate import Accelerator
import numpy as np
import gradio as gr
from main import load_v2
from mesh_to_pc import process_mesh_to_pc
import time
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d.art3d import Poly3DCollection
from PIL import Image
import io

model = load_v2()

device = torch.device('cuda')
accelerator = Accelerator(
    mixed_precision="fp16",
)
model = accelerator.prepare(model)
model.eval()
print("Model loaded to device")

def wireframe_render(mesh):
    views = [
        (90, 20), (270, 20)
    ]
    mesh.vertices = mesh.vertices[:, [0, 2, 1]]

    bounding_box = mesh.bounds
    center = mesh.centroid
    scale = np.ptp(bounding_box, axis=0).max()

    fig = plt.figure(figsize=(10, 10))

    # Function to render and return each view as an image
    def render_view(mesh, azimuth, elevation):
        ax = fig.add_subplot(111, projection='3d')
        ax.set_axis_off()

        # Extract vertices and faces for plotting
        vertices = mesh.vertices
        faces = mesh.faces

        # Plot faces
        ax.add_collection3d(Poly3DCollection(
            vertices[faces],
            facecolors=(0.8, 0.5, 0.2, 1.0),  # Brownish yellow
            edgecolors='k',
            linewidths=0.5,
        ))

        # Set limits and center the view on the object
        ax.set_xlim(center[0] - scale / 2, center[0] + scale / 2)
        ax.set_ylim(center[1] - scale / 2, center[1] + scale / 2)
        ax.set_zlim(center[2] - scale / 2, center[2] + scale / 2)

        # Set view angle
        ax.view_init(elev=elevation, azim=azimuth)

        # Save the figure to a buffer
        buf = io.BytesIO()
        plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0, dpi=300)
        plt.clf()
        buf.seek(0)

        return Image.open(buf)

    # Render each view and store in a list
    images = [render_view(mesh, az, el) for az, el in views]

    # Combine images horizontally
    widths, heights = zip(*(i.size for i in images))
    total_width = sum(widths)
    max_height = max(heights)

    combined_image = Image.new('RGBA', (total_width, max_height))

    x_offset = 0
    for img in images:
        combined_image.paste(img, (x_offset, 0))
        x_offset += img.width

    # Save the combined image
    save_path = f"combined_mesh_view_{int(time.time())}.png"
    combined_image.save(save_path)

    plt.close(fig)
    return save_path

@spaces.GPU(duration=360)
def do_inference(input_3d, sample_seed=0, do_sampling=False, do_marching_cubes=False):
    set_seed(sample_seed)
    print("Seed value:", sample_seed)

    input_mesh = trimesh.load(input_3d)
    pc_list, mesh_list = process_mesh_to_pc([input_mesh], marching_cubes = do_marching_cubes)
    pc_normal = pc_list[0] # 4096, 6
    mesh = mesh_list[0]
    vertices = mesh.vertices

    pc_coor = pc_normal[:, :3]
    normals = pc_normal[:, 3:]

    bounds = np.array([vertices.min(axis=0), vertices.max(axis=0)])
    # scale mesh and pc
    vertices = vertices - (bounds[0] + bounds[1])[None, :] / 2
    vertices = vertices / (bounds[1] - bounds[0]).max()
    mesh.vertices = vertices
    pc_coor = pc_coor - (bounds[0] + bounds[1])[None, :] / 2
    pc_coor = pc_coor / (bounds[1] - bounds[0]).max()

    mesh.merge_vertices()
    mesh.update_faces(mesh.nondegenerate_faces())
    mesh.update_faces(mesh.unique_faces())
    mesh.remove_unreferenced_vertices()
    mesh.fix_normals()
    try:
        if mesh.visual.vertex_colors is not None:
            orange_color = np.array([255, 165, 0, 255], dtype=np.uint8)
    
            mesh.visual.vertex_colors = np.tile(orange_color, (mesh.vertices.shape[0], 1))
        else:
            orange_color = np.array([255, 165, 0, 255], dtype=np.uint8)
            mesh.visual.vertex_colors = np.tile(orange_color, (mesh.vertices.shape[0], 1))
    except Exception as e:
        print(e)
    input_save_name = f"processed_input_{int(time.time())}.obj"
    mesh.export(input_save_name)
    input_render_res = wireframe_render(mesh)

    pc_coor = pc_coor / np.abs(pc_coor).max() * 0.99 # input should be from -1 to 1

    assert (np.linalg.norm(normals, axis=-1) > 0.99).all(), "normals should be unit vectors, something wrong"
    normalized_pc_normal = np.concatenate([pc_coor, normals], axis=-1, dtype=np.float16)

    input = torch.tensor(normalized_pc_normal, dtype=torch.float16, device=device)[None]
    print("Data loaded")

    # with accelerator.autocast():
    with accelerator.autocast():
        outputs = model(input, do_sampling)
    print("Model inference done")
    recon_mesh = outputs[0]

    valid_mask = torch.all(~torch.isnan(recon_mesh.reshape((-1, 9))), dim=1)
    recon_mesh = recon_mesh[valid_mask]  # nvalid_face x 3 x 3
    vertices = recon_mesh.reshape(-1, 3).cpu()
    vertices_index = np.arange(len(vertices))  # 0, 1, ..., 3 x face
    triangles = vertices_index.reshape(-1, 3)

    artist_mesh = trimesh.Trimesh(vertices=vertices, faces=triangles, force="mesh",
                                 merge_primitives=True)

    artist_mesh.merge_vertices()
    artist_mesh.update_faces(artist_mesh.nondegenerate_faces())
    artist_mesh.update_faces(artist_mesh.unique_faces())
    artist_mesh.remove_unreferenced_vertices()
    artist_mesh.fix_normals()

    if do_smooth_shading:
        smooth_shaded(artist_mesh)

    if artist_mesh.visual.vertex_colors is not None:
        orange_color = np.array([255, 165, 0, 255], dtype=np.uint8)

        artist_mesh.visual.vertex_colors = np.tile(orange_color, (artist_mesh.vertices.shape[0], 1))
    else:
        orange_color = np.array([255, 165, 0, 255], dtype=np.uint8)
        artist_mesh.visual.vertex_colors = np.tile(orange_color, (artist_mesh.vertices.shape[0], 1))

    num_faces = len(artist_mesh.faces)

    brown_color = np.array([165, 42, 42, 255], dtype=np.uint8)
    face_colors = np.tile(brown_color, (num_faces, 1))

    artist_mesh.visual.face_colors = face_colors
    # add time stamp to avoid cache
    save_name = f"output_{int(time.time())}.obj"
    artist_mesh.export(save_name)
    output_render = wireframe_render(artist_mesh)
    return input_save_name, input_render_res, save_name, output_render


_HEADER_ = """
## Step 2: Simplify the generated 3D Mesh and Shader Smooth (optional)

ADD ILLUSTRATION

- The 3D Mesh Generated contains too much polygons, fortunately, we can use another AI model to help us optimize it.

- The model we use is called [MeshAnythingV2](https://huggingface.co/Yiwen-ntu/MeshAnythingV2).

- We can make the simplified mesh more smooth is to use Shader Smooth.

- You can usually do it in Blender, but we can do it directly here. Simply -> ✅ Shader Smooth.

## 💡 Tips

- We don't click on Preprocess with marching Cubes, because in the last step the input mesh was produced by it.
- Limited by computational resources, MeshAnything is trained on meshes with fewer than 1600 faces and cannot generate meshes with more than 1600 faces. The shape of the input mesh should be sharp enough; otherwise, it will be challenging to represent it with only 1600 faces. Thus, feed-forward image-to-3D methods may often produce bad results due to insufficient shape quality.

"""

output_model_obj = gr.Model3D(
    label="Generated Mesh (OBJ Format)",
    display_mode="wireframe",
    clear_color=[1, 1, 1, 1],
)
preprocess_model_obj = gr.Model3D(
    label="Processed Input Mesh (OBJ Format)",
    display_mode="wireframe",
    clear_color=[1, 1, 1, 1],
)
input_image_render = gr.Image(
    label="Wireframe Render of Processed Input Mesh",
)
output_image_render = gr.Image(
    label="Wireframe Render of Generated Mesh",
)
with (gr.Blocks() as demo):
    gr.Markdown(_HEADER_)
    with gr.Row(variant="panel"):
        with gr.Column():
            with gr.Row():
                input_3d = gr.Model3D(
                    label="Input Mesh",
                    display_mode="wireframe",
                    clear_color=[1,1,1,1],
                )

            with gr.Row():
                with gr.Group():
                    do_marching_cubes = gr.Checkbox(label="Preprocess with Marching Cubes", value=False)
                    do_smooth_shading = gr.Checkbox(label="Apply Smooth Shading", value=False)
                    do_sampling = gr.Checkbox(label="Random Sampling", value=False)
                    sample_seed = gr.Number(value=0, label="Seed Value", precision=0)

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

            with gr.Row(variant="panel"):
                mesh_examples = gr.Examples(
                    examples=[
                        os.path.join("examples", img_name) for img_name in sorted(os.listdir("examples"))
                    ],
                    inputs=input_3d,
                    outputs=[preprocess_model_obj, input_image_render, output_model_obj, output_image_render],
                    fn=do_inference,
                    cache_examples = False,
                    examples_per_page=10
                )
        with gr.Column():
            with gr.Row():
                input_image_render.render()
            with gr.Row():
                with gr.Tab("OBJ"):
                    preprocess_model_obj.render()
            with gr.Row():
                output_image_render.render()
            with gr.Row():
                with gr.Tab("OBJ"):
                    output_model_obj.render()
            with gr.Row():
                gr.Markdown('''Try click random sampling and different <b>Seed Value</b> if the result is unsatisfying''')

    mv_images = gr.State()

    submit.click(
        fn=do_inference,
        inputs=[input_3d, sample_seed, do_sampling, do_marching_cubes, do_smooth_shading],
        outputs=[preprocess_model_obj, input_image_render, output_model_obj, output_image_render],
    )

demo.launch(share=True)