gradio_app.py

import spaces
import os
import tempfile
from typing import Any
import torch
import numpy as np
from PIL import Image
import gradio as gr
import trimesh
from transparent_background import Remover
from pathlib import Path
import uuid
import subprocess

def install_cuda_toolkit():
    # CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run"
    CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/12.2.0/local_installers/cuda_12.2.0_535.54.03_linux.run"
    CUDA_TOOLKIT_FILE = "/tmp/%s" % os.path.basename(CUDA_TOOLKIT_URL)
    subprocess.call(["wget", "-q", CUDA_TOOLKIT_URL, "-O", CUDA_TOOLKIT_FILE])
    subprocess.call(["chmod", "+x", CUDA_TOOLKIT_FILE])
    subprocess.call([CUDA_TOOLKIT_FILE, "--silent", "--toolkit"])

    os.environ["CUDA_HOME"] = "/usr/local/cuda"
    os.environ["PATH"] = "%s/bin:%s" % (os.environ["CUDA_HOME"], os.environ["PATH"])
    os.environ["LD_LIBRARY_PATH"] = "%s/lib:%s" % (
        os.environ["CUDA_HOME"],
        "" if "LD_LIBRARY_PATH" not in os.environ else os.environ["LD_LIBRARY_PATH"],
    )
    # Fix: arch_list[-1] += '+PTX'; IndexError: list index out of range
    os.environ["TORCH_CUDA_ARCH_LIST"] = "8.0;8.6"

install_cuda_toolkit()

# Import and setup SPAR3D 
os.system("USE_CUDA=1 pip install -vv --no-build-isolation ./texture_baker ./uv_unwrapper")
import spar3d.utils as spar3d_utils
from spar3d.system import SPAR3D

# Constants
COND_WIDTH = 512
COND_HEIGHT = 512
COND_DISTANCE = 2.2
COND_FOVY = 0.591627
BACKGROUND_COLOR = [0.5, 0.5, 0.5]
OUTPUT_DIR = "./output"
os.makedirs(OUTPUT_DIR, exist_ok=True)
HF_TOKEN = os.getenv("HF_TOKEN", None)

# Initialize models
device = spar3d_utils.get_device()
bg_remover = Remover()
spar3d_model = SPAR3D.from_pretrained(
    "stabilityai/stable-point-aware-3d",
    config_name="config.yaml",
    weight_name="model.safetensors",
).eval().to(device)

# Initialize camera parameters
c2w_cond = spar3d_utils.default_cond_c2w(COND_DISTANCE)
intrinsic, intrinsic_normed_cond = spar3d_utils.create_intrinsic_from_fov_rad(
    COND_FOVY, COND_HEIGHT, COND_WIDTH
)

def create_rgba_image(rgb_image: Image.Image, mask: np.ndarray = None) -> Image.Image:
    """Create an RGBA image from RGB image and optional mask."""
    rgba_image = rgb_image.convert('RGBA')
    if mask is not None:
        # Ensure mask is 2D before converting to alpha
        if len(mask.shape) > 2:
            mask = mask.squeeze()
        alpha = Image.fromarray((mask * 255).astype(np.uint8))
        rgba_image.putalpha(alpha)
    return rgba_image

def create_batch(input_image: Image.Image) -> dict[str, Any]:
    """Prepare image batch for model input."""
    # Resize and convert input image to numpy array
    resized_image = input_image.resize((COND_WIDTH, COND_HEIGHT))
    img_array = np.array(resized_image).astype(np.float32) / 255.0

    # Extract RGB and alpha channels
    if img_array.shape[-1] == 4:  # RGBA
        rgb = img_array[..., :3]
        mask = img_array[..., 3:4]
    else:  # RGB
        rgb = img_array
        mask = np.ones((*img_array.shape[:2], 1), dtype=np.float32)
    
    # Convert to tensors while keeping channel-last format
    rgb = torch.from_numpy(rgb).float()  # [H, W, 3]
    mask = torch.from_numpy(mask).float()  # [H, W, 1]

    # Create background blend (match channel-last format)
    bg_tensor = torch.tensor(BACKGROUND_COLOR).view(1, 1, 3)  # [1, 1, 3]
 
    # Blend RGB with background using mask (all in channel-last format)
    rgb_cond = torch.lerp(bg_tensor, rgb, mask)  # [H, W, 3]
 
    # Move channels to correct dimension and add batch dimension
    # Important: For SPAR3D image tokenizer, we need [B, H, W, C] format
    rgb_cond = rgb_cond.unsqueeze(0)  # [1, H, W, 3]
    mask = mask.unsqueeze(0)  # [1, H, W, 1]
    
    # Create the batch dictionary
    batch = {
        "rgb_cond": rgb_cond,  # [1, H, W, 3]
        "mask_cond": mask,  # [1, H, W, 1]
        "c2w_cond": c2w_cond.unsqueeze(0),  # [1, 4, 4]
        "intrinsic_cond": intrinsic.unsqueeze(0),  # [1, 3, 3]
        "intrinsic_normed_cond": intrinsic_normed_cond.unsqueeze(0),  # [1, 3, 3]
    }
    
    for k, v in batch.items():
        print(f"[debug] {k} final shape:", v.shape)
 
    return batch

def forward_model(batch, system, guidance_scale=3.0, seed=0, device="cuda"):
    """Process batch through model and generate point cloud."""

    batch_size = batch["rgb_cond"].shape[0]
    assert batch_size == 1, f"Expected batch size 1, got {batch_size}"
    
    # Generate point cloud tokens
    try:
        cond_tokens = system.forward_pdiff_cond(batch)
    except Exception as e:
        print("\n[ERROR] Failed in forward_pdiff_cond:")
        print(e)
        print("\nInput tensor properties:")
        print("rgb_cond dtype:", batch["rgb_cond"].dtype)
        print("rgb_cond device:", batch["rgb_cond"].device)
        print("rgb_cond requires_grad:", batch["rgb_cond"].requires_grad)
        raise
    
    # Sample points
    sample_iter = system.sampler.sample_batch_progressive(
        batch_size,
        cond_tokens,
        guidance_scale=guidance_scale,
        device=device
    )
    
    # Get final samples
    for x in sample_iter:
        samples = x["xstart"]
    
    pc_cond = samples.permute(0, 2, 1).float()

    # Normalize point cloud
    pc_cond = spar3d_utils.normalize_pc_bbox(pc_cond)

    # Subsample to 512 points
    pc_cond = pc_cond[:, torch.randperm(pc_cond.shape[1])[:512]]

    return pc_cond

@spaces.GPU
@torch.inference_mode()
def generate_and_process_3d(image: Image.Image) -> tuple[str | None, str | None]:
    """Generate image from prompt and convert to 3D model."""

    # Generate random seed
    seed = np.random.randint(0, np.iinfo(np.int32).max)
    
    try:
        rgb_image = image.convert('RGB')
        
        # bg_remover returns a PIL Image already, no need to convert
        no_bg_image = bg_remover.process(rgb_image)
        print(f"[debug] no_bg_image type: {type(no_bg_image)}, mode: {no_bg_image.mode}")
        
        # Convert to RGBA if not already
        rgba_image = no_bg_image.convert('RGBA')
        print(f"[debug] rgba_image mode: {rgba_image.mode}")
        
        processed_image = spar3d_utils.foreground_crop(
            rgba_image,
            crop_ratio=1.3,
            newsize=(COND_WIDTH, COND_HEIGHT),
            no_crop=False
        )
        
        # Show the processed image alpha channel for debugging
        alpha = np.array(processed_image)[:, :, 3]
        print(f"[debug] Alpha channel stats - min: {alpha.min()}, max: {alpha.max()}, unique: {np.unique(alpha)}")

        # Prepare batch for processing
        batch = create_batch(processed_image)
        batch = {k: v.to(device) for k, v in batch.items()}

        # Generate point cloud
        pc_cond = forward_model(
            batch,
            spar3d_model,
            guidance_scale=3.0,
            seed=seed,
            device=device
        )
        batch["pc_cond"] = pc_cond

        # Generate mesh
        with torch.no_grad():
            with torch.autocast(device_type='cuda' if torch.cuda.is_available() else 'cpu', dtype=torch.bfloat16):
                trimesh_mesh, _ = spar3d_model.generate_mesh(
                    batch,
                    1024,  # texture_resolution
                    remesh="none",
                    vertex_count=-1,
                    estimate_illumination=True
                )
                trimesh_mesh = trimesh_mesh[0]

        # Export to GLB
        unique_id = str(uuid.uuid4())
        filename = f'model_{unique_id}.glb'
        output_path = os.path.join(OUTPUT_DIR, filename)
        public_url = f"https://john6666-image-to-3d-test.hf.space/gradio_api/file={Path(output_path).resolve()}"
        print(f"public_url: {public_url}")
        
        trimesh_mesh.export(output_path, file_type="glb", include_normals=True)
        
        return output_path, public_url
        
    except Exception as e:
        print(f"Error during generation: {str(e)}")
        import traceback
        traceback.print_exc()
        return None, None

css = """
.image { margin: 0px auto; object-fit: contain; !important; }
.info { align-items: center; text-align: center; }
"""

# Create Gradio app using Blocks
with gr.Blocks(theme=gr.themes.Soft(), fill_width=True, fill_height=True, elem_id="container", css=css) as demo:
    gr.Markdown("This space is based on [Stable Point-Aware 3D](https://huggingface.co/spaces/stabilityai/stable-point-aware-3d) by Stability AI, [Text to 3D](https://huggingface.co/spaces/jbilcke-hf/text-to-3d) by jbilcke-hf.", elem_classes="info")
    
    input_img = gr.Image(type="pil", label="Input Image", sources="upload", image_mode="RGBA", width=128, elem_classes="image")
    model_output = gr.Model3D(label="Generated .GLB model", clear_color=[0.0, 0.0, 0.0, 0.0])
    output_url = gr.Textbox(label="Output URL", value="", lines=1, interactive=False, visible=True, show_copy_button=True)

    # Event handler
    input_img.upload(
        fn=generate_and_process_3d,
        inputs=[input_img],
        outputs=[model_output, output_url],
        api_name="generate"
    )
    
if __name__ == "__main__":
    demo.queue().launch(ssr_mode=False, allowed_paths=[Path(OUTPUT_DIR).resolve()])