gradio_app.py

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 diffusers import DiffusionPipeline

# 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]

# 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 FLUX model
dtype = torch.bfloat16
flux_pipe = DiffusionPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell", 
    torch_dtype=dtype
).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:
        print("[debug] mask shape before alpha:", mask.shape)
        # Ensure mask is 2D before converting to alpha
        if len(mask.shape) > 2:
            mask = mask.squeeze()
        alpha = Image.fromarray((mask * 255).astype(np.uint8))
        print("[debug] alpha size:", alpha.size)
        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
    print("[debug] img_array shape:", img_array.shape)
    
    # 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
    rgb = torch.from_numpy(rgb).float()
    mask = torch.from_numpy(mask).float()
    print("[debug] rgb tensor shape:", rgb.shape)
    print("[debug] mask tensor shape:", mask.shape)
    
    # Create background blend
    bg_tensor = torch.tensor(BACKGROUND_COLOR)[None, None, :]
    print("[debug] bg_tensor shape:", bg_tensor.shape)
    
    # Blend RGB with background using mask
    rgb_cond = torch.lerp(bg_tensor, rgb, mask)
    print("[debug] rgb_cond shape:", rgb_cond.shape)

    # Permute the tensors to match the expected shape [B, C, H, W]
    rgb_cond = torch.movedim(rgb_cond, 2, 0).unsqueeze(0)  # [1, 3, H, W]
    mask = torch.movedim(mask, 2, 0).unsqueeze(0)  # [1, 1, H, W]
    
    print("[debug] rgb_cond after permute shape:", rgb_cond.shape)
    print("[debug] mask after permute shape:", mask.shape)

    batch = {
        "rgb_cond": rgb_cond,
        "mask_cond": mask,
        "c2w_cond": c2w_cond.unsqueeze(0),
        "intrinsic_cond": intrinsic.unsqueeze(0),
        "intrinsic_normed_cond": intrinsic_normed_cond.unsqueeze(0),
    }
    
    # Final shapes check
    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."""
    print("[debug] Starting forward_model")
    batch_size = batch["rgb_cond"].shape[0]
    
    # Generate point cloud tokens
    print("[debug] Generating point cloud tokens")
    cond_tokens = system.forward_pdiff_cond(batch)
    print("[debug] cond_tokens shape:", cond_tokens.shape)
    
    # Sample points
    print("[debug] Sampling 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"]
    
    print("[debug] samples shape before permute:", samples.shape)
    # Convert samples to point cloud format
    pc_cond = samples.permute(0, 2, 1).float()
    print("[debug] pc_cond shape after permute:", pc_cond.shape)
    
    # Normalize point cloud
    pc_cond = spar3d_utils.normalize_pc_bbox(pc_cond)
    print("[debug] pc_cond shape after normalize:", pc_cond.shape)
    
    # Subsample to 512 points
    pc_cond = pc_cond[:, torch.randperm(pc_cond.shape[1])[:512]]
    print("[debug] pc_cond final shape:", pc_cond.shape)
    
    return pc_cond

def generate_and_process_3d(prompt: str, seed: int = 42, width: int = 1024, height: int = 1024) -> tuple[str | None, Image.Image | None]:
    """Generate image from prompt and convert to 3D model."""
    try:
        # Set random seeds
        torch.manual_seed(seed)
        np.random.seed(seed)
        
        # Generate image using FLUX
        generator = torch.Generator(device=device).manual_seed(seed)
        print("[debug] generating the image using Flux")
        generated_image = flux_pipe(
            prompt=prompt,
            width=width,
            height=height,
            num_inference_steps=4,
            generator=generator,
            guidance_scale=0.0
        ).images[0]
        
        print("[debug] converting the image to rgb")
        rgb_image = generated_image.convert('RGB')
        
        print("[debug] removing the background by calling bg_remover.process(rgb_image)")
        no_bg_image = bg_remover.process(rgb_image)
        
        print("[debug] converting to numpy array to extract the mask")
        no_bg_array = np.array(no_bg_image)
        
        # Create mask based on RGB values
        mask = ((no_bg_array > 0).any(axis=2)).astype(np.float32)
        mask = np.expand_dims(mask, axis=2)  # Add channel dimension
        
        print("[debug] creating the RGBA image using create_rgba_image(rgb_image, mask)")
        rgba_image = create_rgba_image(rgb_image, mask)
        
        print("[debug] auto-cropping the rgba_image using spar3d_utils.foreground_crop(...)")
        processed_image = spar3d_utils.foreground_crop(
            rgba_image,
            crop_ratio=1.3,
            newsize=(COND_WIDTH, COND_HEIGHT),
            no_crop=False
        )

        # Prepare batch for processing
        print("[debug] preparing the batch by calling create_batch(processed_image)")
        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():
            print("[debug] calling torch.autocast(....) to generate the mesh")
            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
        print("[debug] creating tmp dir for the .glb output")
        temp_dir = tempfile.mkdtemp()
        output_path = os.path.join(temp_dir, 'output.glb')
        
        print("[debug] calling trimesh_mesh.export(...) to export to .glb")
        trimesh_mesh.export(output_path, file_type="glb", include_normals=True)
        
        return output_path, generated_image
        
    except Exception as e:
        print(f"Error during generation: {str(e)}")
        import traceback
        traceback.print_exc()
        return None, None

# Create Gradio interface
demo = gr.Interface(
    fn=generate_and_process_3d,
    inputs=[
        gr.Text(
            label="Enter your prompt",
            placeholder="Describe what you want to generate..."
        ),
        gr.Slider(
            label="Seed",
            minimum=0,
            maximum=np.iinfo(np.int32).max,
            step=1,
            value=42
        ),
        gr.Slider(
            label="Width",
            minimum=256,
            maximum=2048,
            step=32,
            value=1024
        ),
        gr.Slider(
            label="Height",
            minimum=256,
            maximum=2048,
            step=32,
            value=1024
        )
    ],
    outputs=[
        gr.File(
            label="Download 3D Model",
            file_types=[".glb"]
        ),
        gr.Image(
            label="Generated Image",
            type="pil"
        )
    ],
    title="Text to 3D Model Generator",
    description="Enter a text prompt to generate an image that will be converted into a 3D model",
)

if __name__ == "__main__":
    demo.queue().launch()