app.py

# Copyright (C) 2025, FaceLift Research Group
# https://github.com/weijielyu/FaceLift
#
# This software is free for non-commercial, research and evaluation use 
# under the terms of the LICENSE.md file.
#
# For inquiries contact: [email protected]

"""
FaceLift: Single Image 3D Face Reconstruction
Generates 3D head models from single images using multi-view diffusion and GS-LRM.
"""

import json
from pathlib import Path
from datetime import datetime
import uuid
import time
import shutil

import gradio as gr
import numpy as np
import torch
import yaml
from easydict import EasyDict as edict
from einops import rearrange
from PIL import Image
from huggingface_hub import snapshot_download
import spaces

# Install diff-gaussian-rasterization at runtime (requires GPU)
import subprocess
import sys
import os

# -----------------------------
# Static paths (for viewer files)
# -----------------------------
OUTPUTS_DIR = Path.cwd() / "outputs"
SPLATS_ROOT = OUTPUTS_DIR / "splats"
SPLATS_ROOT.mkdir(parents=True, exist_ok=True)

# Serve ./outputs via Gradio's static router: /gradio_api/file=outputs/...
gr.set_static_paths(paths=[OUTPUTS_DIR])

# -----------------------------
# Per-session helpers
# -----------------------------
def new_session_id() -> str:
    return uuid.uuid4().hex[:10]

def session_dir(session_id: str) -> Path:
    p = SPLATS_ROOT / session_id
    p.mkdir(parents=True, exist_ok=True)
    return p

def cleanup_old_sessions(max_age_hours: int = 6):
    cutoff = time.time() - max_age_hours * 3600
    if not SPLATS_ROOT.exists():
        return
    for child in SPLATS_ROOT.iterdir():
        try:
            if child.is_dir() and child.stat().st_mtime < cutoff:
                shutil.rmtree(child, ignore_errors=True)
        except Exception:
            pass

def copy_to_session_and_get_url(src_path: str, session_id: str) -> str:
    """
    Copy a .splat or .ply into this user's session folder and return a cache-busted URL.
    """
    src = Path(src_path)
    ext = src.suffix.lower() if src.suffix else ".ply"
    fn = f"{int(time.time())}_{uuid.uuid4().hex[:6]}{ext}"
    dst = session_dir(session_id) / fn
    shutil.copy2(src, dst)
    # /gradio_api/file=outputs/...
    return f"/gradio_api/file=outputs/splats/{session_id}/{fn}?v={uuid.uuid4().hex[:6]}"

# -----------------------------
# Ensure diff-gaussian-rasterization builds for current GPU
# -----------------------------
try:
    import diff_gaussian_rasterization  # noqa: F401
except ImportError:
    print("Installing diff-gaussian-rasterization (compiling for detected CUDA arch)...")
    env = os.environ.copy()
    try:
        import torch as _torch
        if _torch.cuda.is_available():
            maj, minr = _torch.cuda.get_device_capability()
            arch = f"{maj}.{minr}"                 # e.g., "9.0" on H100/H200, "8.0" on A100
            env["TORCH_CUDA_ARCH_LIST"] = f"{arch}+PTX"
        else:
            # Build stage may not see a GPU on HF Spaces: compile a cross-arch set
            env["TORCH_CUDA_ARCH_LIST"] = "8.0;8.6;8.9;9.0+PTX"
    except Exception:
        env["TORCH_CUDA_ARCH_LIST"] = "8.0;8.6;8.9;9.0+PTX"

    # (Optional) side-step allocator+NVML quirks in restrictive containers
    env.setdefault("PYTORCH_NO_CUDA_MEMORY_CACHING", "1")

    subprocess.check_call(
        [sys.executable, "-m", "pip", "install",
         "git+https://github.com/graphdeco-inria/diff-gaussian-rasterization"],
        env=env,
    )
    import diff_gaussian_rasterization  # noqa: F401


from gslrm.model.gaussians_renderer import render_turntable, imageseq2video
from mvdiffusion.pipelines.pipeline_mvdiffusion_unclip import StableUnCLIPImg2ImgPipeline
from utils_folder.face_utils import preprocess_image, preprocess_image_without_cropping

# HuggingFace repository configuration
HF_REPO_ID = "wlyu/OpenFaceLift"

def download_weights_from_hf() -> Path:
    """Download model weights from HuggingFace if not already present.
    
    Returns:
        Path to the downloaded repository
    """
    workspace_dir = Path(__file__).parent
    
    # Check if weights already exist locally
    mvdiffusion_path = workspace_dir / "checkpoints/mvdiffusion/pipeckpts"
    gslrm_path = workspace_dir / "checkpoints/gslrm/ckpt_0000000000021125.pt"
    
    if mvdiffusion_path.exists() and gslrm_path.exists():
        print("Using local model weights")
        return workspace_dir
    
    print(f"Downloading model weights from HuggingFace: {HF_REPO_ID}")
    print("This may take a few minutes on first run...")
    
    # Download to local directory
    snapshot_download(
        repo_id=HF_REPO_ID,
        local_dir=str(workspace_dir / "checkpoints"),
        local_dir_use_symlinks=False,
    )
    
    print("Model weights downloaded successfully!")
    return workspace_dir

class FaceLiftPipeline:
    """Pipeline for FaceLift 3D head generation from single images."""
    
    def __init__(self):
        # Download weights from HuggingFace if needed
        workspace_dir = download_weights_from_hf()
        
        # Setup paths
        self.output_dir = workspace_dir / "outputs"
        self.examples_dir = workspace_dir / "examples"
        self.output_dir.mkdir(exist_ok=True)
        
        # Parameters
        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
        self.image_size = 512
        self.camera_indices = [2, 1, 0, 5, 4, 3]
        
        # Load models (keep on CPU for ZeroGPU compatibility)
        print("Loading models...")
        try:
            self.mvdiffusion_pipeline = StableUnCLIPImg2ImgPipeline.from_pretrained(
                str(workspace_dir / "checkpoints/mvdiffusion/pipeckpts"),
                torch_dtype=torch.float16,
            )
            # Don't move to device or enable xformers here - will be done in GPU-decorated function
            self._models_on_gpu = False
            
            with open(workspace_dir / "configs/gslrm.yaml", "r") as f:
                config = edict(yaml.safe_load(f))
            
            module_name, class_name = config.model.class_name.rsplit(".", 1)
            module = __import__(module_name, fromlist=[class_name])
            ModelClass = getattr(module, class_name)
            
            self.gs_lrm_model = ModelClass(config)
            checkpoint = torch.load(
                workspace_dir / "checkpoints/gslrm/ckpt_0000000000021125.pt",
                map_location="cpu"
            )
            # Filter out loss_calculator weights (training-only, not needed for inference)
            state_dict = {k: v for k, v in checkpoint["model"].items() 
                          if not k.startswith("loss_calculator.")}
            self.gs_lrm_model.load_state_dict(state_dict)
            # Keep on CPU initially - will move to GPU in decorated function
            
            self.color_prompt_embedding = torch.load(
                workspace_dir / "mvdiffusion/fixed_prompt_embeds_6view/clr_embeds.pt",
                map_location="cpu"
            )
            
            with open(workspace_dir / "utils_folder/opencv_cameras.json", 'r') as f:
                self.cameras_data = json.load(f)["frames"]
            
            print("Models loaded successfully!")
        except Exception as e:
            print(f"Error loading models: {e}")
            import traceback
            traceback.print_exc()
            raise
    
    def _move_models_to_gpu(self):
        """Move models to GPU and enable optimizations. Called within @spaces.GPU context."""
        if not self._models_on_gpu and torch.cuda.is_available():
            print("Moving models to GPU...")
            self.device = torch.device("cuda:0")
            self.mvdiffusion_pipeline.to(self.device)
            self.mvdiffusion_pipeline.unet.enable_xformers_memory_efficient_attention()
            self.gs_lrm_model.to(self.device)
            self.gs_lrm_model.eval()  # Set to eval mode
            self.color_prompt_embedding = self.color_prompt_embedding.to(self.device)
            self._models_on_gpu = True
            torch.cuda.empty_cache()  # Clear cache after moving models
            print("Models on GPU, xformers enabled!")
    
    @spaces.GPU(duration=120)
    def generate_3d_head(self, image_path, auto_crop=True, guidance_scale=3.0, 
                         random_seed=4, num_steps=50):
        """Generate 3D head from single image."""
        try:
            # Move models to GPU now that we're in the GPU context
            self._move_models_to_gpu()
            # Setup output directory
            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
            output_dir = self.output_dir / timestamp
            output_dir.mkdir(exist_ok=True)
            
            # Preprocess input
            original_img = np.array(Image.open(image_path))
            input_image = preprocess_image(original_img) if auto_crop else \
                         preprocess_image_without_cropping(original_img)
            
            if input_image.size != (self.image_size, self.image_size):
                input_image = input_image.resize((self.image_size, self.image_size))
            
            input_path = output_dir / "input.png"
            input_image.save(input_path)
            
            # Generate multi-view images
            generator = torch.Generator(device=self.mvdiffusion_pipeline.unet.device)
            generator.manual_seed(random_seed)
            
            result = self.mvdiffusion_pipeline(
                input_image, None,
                prompt_embeds=self.color_prompt_embedding,
                height=self.image_size,
                width=self.image_size,
                guidance_scale=guidance_scale,
                num_images_per_prompt=1,
                num_inference_steps=num_steps,
                generator=generator,
                eta=1.0,
            )
            
            selected_views = result.images[:6]
            
            # Save multi-view composite
            multiview_image = Image.new("RGB", (self.image_size * 6, self.image_size))
            for i, view in enumerate(selected_views):
                multiview_image.paste(view, (self.image_size * i, 0))
            
            multiview_path = output_dir / "multiview.png"
            multiview_image.save(multiview_path)
            
            # Move diffusion model to CPU to free GPU memory for GS-LRM
            print("Moving diffusion model to CPU to free memory...")
            self.mvdiffusion_pipeline.to("cpu")
            
            # Delete intermediate variables to free memory
            del result, generator
            torch.cuda.empty_cache()
            torch.cuda.synchronize()
            
            # Prepare 3D reconstruction input
            view_arrays = [np.array(view) for view in selected_views]
            lrm_input = torch.from_numpy(np.stack(view_arrays, axis=0)).float()
            lrm_input = lrm_input[None].to(self.device) / 255.0
            lrm_input = rearrange(lrm_input, "b v h w c -> b v c h w")
            
            # Prepare camera parameters
            selected_cameras = [self.cameras_data[i] for i in self.camera_indices]
            fxfycxcy_list = [[c["fx"], c["fy"], c["cx"], c["cy"]] for c in selected_cameras]
            c2w_list = [np.linalg.inv(np.array(c["w2c"])) for c in selected_cameras]
            
            fxfycxcy = torch.from_numpy(np.stack(fxfycxcy_list, axis=0).astype(np.float32))
            c2w = torch.from_numpy(np.stack(c2w_list, axis=0).astype(np.float32))
            fxfycxcy = fxfycxcy[None].to(self.device)
            c2w = c2w[None].to(self.device)
            
            batch_indices = torch.stack([
                torch.zeros(lrm_input.size(1)).long(),
                torch.arange(lrm_input.size(1)).long(),
            ], dim=-1)[None].to(self.device)
            
            batch = edict({
                "image": lrm_input,
                "c2w": c2w,
                "fxfycxcy": fxfycxcy,
                "index": batch_indices,
            })
            
            # Ensure GS-LRM model is on GPU
            if next(self.gs_lrm_model.parameters()).device.type == "cpu":
                print("Moving GS-LRM model to GPU...")
                self.gs_lrm_model.to(self.device)
                torch.cuda.empty_cache()

            # Final memory cleanup before reconstruction
            torch.cuda.empty_cache()
            
            # Run 3D reconstruction
            with torch.no_grad(), torch.autocast(enabled=True, device_type="cuda", dtype=torch.float16):
                result = self.gs_lrm_model.forward(batch, create_visual=False, split_data=True)
            
            comp_image = result.render[0].unsqueeze(0).detach()
            gaussians = result.gaussians[0]
            
            # Clear CUDA cache after reconstruction
            torch.cuda.empty_cache()
            
            # Save filtered gaussians
            filtered_gaussians = gaussians.apply_all_filters(
                cam_origins=None,
                opacity_thres=0.04,
                scaling_thres=0.2,
                floater_thres=0.75,
                crop_bbx=[-0.91, 0.91, -0.91, 0.91, -1.0, 1.0],
                nearfar_percent=(0.0001, 1.0),
            )
            
            ply_path = output_dir / "gaussians.ply"
            filtered_gaussians.save_ply(str(ply_path))
            
            # Save output image
            comp_image = rearrange(comp_image, "x v c h w -> (x h) (v w) c")
            comp_image = (comp_image.cpu().numpy() * 255.0).clip(0, 255).astype(np.uint8)
            output_path = output_dir / "output.png"
            Image.fromarray(comp_image).save(output_path)
            
            # Generate turntable video
            turntable_resolution = 512
            num_turntable_views = 180
            turntable_frames = render_turntable(gaussians, rendering_resolution=turntable_resolution, 
                                               num_views=num_turntable_views)
            turntable_frames = rearrange(turntable_frames, "h (v w) c -> v h w c", v=num_turntable_views)
            turntable_frames = np.ascontiguousarray(turntable_frames)
            
            turntable_path = output_dir / "turntable.mp4"
            imageseq2video(turntable_frames, str(turntable_path), fps=30)
            
            # Final CUDA cache clear
            torch.cuda.empty_cache()
            
            return str(input_path), str(multiview_path), str(output_path), \
                   str(turntable_path), str(ply_path)
            
        except Exception as e:
            import traceback
            error_details = traceback.format_exc()
            print(f"Error details:\n{error_details}")
            raise gr.Error(f"Generation failed: {str(e)}")

# -----------------------------
# gsplat.js viewer (Option A)
# -----------------------------
GSPLAT_HEAD = """
<script type="module">
  import * as SPLAT from "https://cdn.jsdelivr.net/npm/[email protected]/+esm";
  let renderer, scene, camera, controls;

  function ensureViewer() {
    if (renderer) return;
    const container = document.getElementById("splat-container");
    renderer  = new SPLAT.WebGLRenderer();
    container.appendChild(renderer.canvas);
    scene   = new SPLAT.Scene();
    camera  = new SPLAT.Camera();
    controls = new SPLAT.OrbitControls(camera, renderer.canvas);
    const loop = () => { controls.update(); renderer.render(scene, camera); requestAnimationFrame(loop); };
    requestAnimationFrame(loop);
  }

  async function loadSplat(url) {
    ensureViewer();
    // clear previous
    scene.children.length = 0;
    await SPLAT.Loader.LoadAsync(url, scene, ()=>{});
  }

  // Expose callable function for Gradio
  window.__load_splat__ = loadSplat;
</script>
"""

def main():
    """Run the FaceLift application with an embedded gsplat.js viewer and per-session files."""
    pipeline = FaceLiftPipeline()

    # Prepare examples (same as before)
    examples = []
    if pipeline.examples_dir.exists():
        examples = [[str(f), True, 3.0, 4, 50] for f in sorted(pipeline.examples_dir.iterdir()) 
                   if f.suffix.lower() in {'.png', '.jpg', '.jpeg'}]

    with gr.Blocks(head=GSPLAT_HEAD, title="FaceLift: Single Image 3D Face Reconstruction") as demo:
        session = gr.State()

        # Light GC + session init
        def _init_session():
            cleanup_old_sessions()
            return new_session_id()

        # After generation: copy ply into per-session folder and return viewer URL
        def _prep_viewer_url(ply_path: str, session_id: str) -> str:
            if not ply_path or not os.path.exists(ply_path):
                return ""
            return copy_to_session_and_get_url(ply_path, session_id)

        gr.Markdown("## FaceLift: Single Image 3D Face Reconstruction\nTurn a single portrait image into a 3D head model and preview it interactively.")
        with gr.Row():
            with gr.Column(scale=1):
                in_image  = gr.Image(type="filepath", label="Input Portrait Image")
                auto_crop = gr.Checkbox(value=True, label="Auto Cropping")
                guidance  = gr.Slider(1.0, 10.0, 3.0, step=0.1, label="Guidance Scale")
                seed      = gr.Number(value=4, label="Random Seed")
                steps     = gr.Slider(10, 100, 50, step=5, label="Generation Steps")
                run_btn   = gr.Button("Generate 3D Head", variant="primary")

                # Examples (match input signature)
                if examples:
                    gr.Examples(
                        examples=examples,
                        inputs=[in_image, auto_crop, guidance, seed, steps],
                        examples_per_page=8,
                    )

            with gr.Column(scale=1):
                out_proc = gr.Image(label="Processed Input")
                out_multi = gr.Image(label="Multi-view Generation")
                out_recon = gr.Image(label="3D Reconstruction")
                out_video = gr.PlayableVideo(label="Turntable Animation")
                out_ply   = gr.File(label="3D Model (.ply)")

        gr.Markdown("### Interactive Gaussian Splat Viewer")
        with gr.Row():
            url_box = gr.Textbox(label="Scene URL (auto-filled)", interactive=False)
        viewer = gr.HTML("<div id='splat-container' style='width:100%;height:640px'></div>")
        reload_btn = gr.Button("Reload Viewer")

        # Initialize per-browser session
        demo.load(fn=_init_session, inputs=None, outputs=session)

        # Chain: run → show outputs → prepare viewer URL → load viewer (JS)
        run_btn.click(
            fn=pipeline.generate_3d_head,
            inputs=[in_image, auto_crop, guidance, seed, steps],
            outputs=[out_proc, out_multi, out_recon, out_video, out_ply],
        ).then(
            fn=_prep_viewer_url,
            inputs=[out_ply, session],
            outputs=url_box,
        ).then(
            fn=None, inputs=url_box, outputs=None,
            js="(url)=>window.__load_splat__(url)"
        )

        # Manual reload if needed
        reload_btn.click(fn=None, inputs=url_box, outputs=None, js="(url)=>window.__load_splat__(url)")

        demo.queue(max_size=10)
        demo.launch(share=True, server_name="0.0.0.0", server_port=7860, show_error=True)

if __name__ == "__main__":
    main()