app.py

"""PoseShield: Neural Collision Fields for Human Self-Collision Resolution.

A Gradio demo that takes a colliding SMPL-H pose and resolves self-collisions
using the PoseShield neural collision field with SLSQP optimization.
"""

import os
os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")

import spaces  # MUST come before torch
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import pickle
import tempfile
import time
import yaml
import struct
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from scipy.optimize import minimize
from huggingface_hub import hf_hub_download
import smplx
import gradio as gr


# ---------------------------------------------------------------------------
# Rotation utilities (from poseshield.common.utils)
# ---------------------------------------------------------------------------

def normalize(x, axis=-1, eps=1e-8):
    norm = np.linalg.norm(x, axis=axis, keepdims=True) + eps
    return x / norm


def axis_angle_to_matrix(axis_angle):
    """Convert (N, 3) axis-angle to (N, 3, 3) rotation matrices (Rodrigues)."""
    aa = np.asarray(axis_angle, dtype=np.float64)
    N = aa.shape[0]
    theta = np.linalg.norm(aa, axis=1, keepdims=True)
    eps = 1e-8
    k = aa / (theta + eps)
    kx, ky, kz = k[:, 0], k[:, 1], k[:, 2]
    K = np.zeros((N, 3, 3), dtype=np.float64)
    K[:, 0, 1] = -kz
    K[:, 0, 2] = ky
    K[:, 1, 0] = kz
    K[:, 1, 2] = -kx
    K[:, 2, 0] = -ky
    K[:, 2, 1] = kx
    I = np.eye(3, dtype=np.float64)[None, :, :]
    sin_t = np.sin(theta)[:, None].reshape(N, 1, 1)
    cos_t = np.cos(theta)[:, None].reshape(N, 1, 1)
    K2 = K @ K
    small = (theta.reshape(N) < 1e-4)
    A = np.empty((N, 1, 1), dtype=np.float64)
    B = np.empty((N, 1, 1), dtype=np.float64)
    A[~small] = sin_t[~small]
    B[~small] = (1.0 - cos_t[~small])
    th = theta.reshape(N, 1, 1)
    A[small] = th[small] - (th[small]**3) / 6.0
    B[small] = (th[small]**2) / 2.0 - (th[small]**4) / 24.0
    R = I + A * K + B * K2
    return R


def matrix_to_axis_angle(R):
    """Convert (N, 3, 3) rotation matrices to (N, 3) axis-angle vectors."""
    trace = np.trace(R, axis1=1, axis2=2)
    trace = np.clip(trace, -1.0, 3.0)
    angles = np.arccos((trace - 1.0) / 2.0)
    rx = R[:, 2, 1] - R[:, 1, 2]
    ry = R[:, 0, 2] - R[:, 2, 0]
    rz = R[:, 1, 0] - R[:, 0, 1]
    axes = np.stack([rx, ry, rz], axis=1)
    sin_angles = np.linalg.norm(axes, axis=1, keepdims=True) / 2.0
    axes = axes / (2.0 * (sin_angles + 1e-8))
    axis_angle = axes * angles[:, None]
    return axis_angle


def rotation_6d_to_matrix(d6):
    """Convert (..., 6) 6D rotation to (..., 3, 3) rotation matrix (Gram-Schmidt)."""
    a1 = d6[..., :3]
    a2 = d6[..., 3:]
    b1 = normalize(a1, axis=-1)
    dot = np.sum(b1 * a2, axis=-1, keepdims=True)
    b2 = a2 - dot * b1
    b2 = normalize(b2, axis=-1)
    b3 = np.cross(b1, b2, axis=-1)
    rotation_mats = np.stack((b1, b2, b3), axis=-2)
    return rotation_mats


def matrix_to_rotation_6d(R):
    """Convert (..., 3, 3) rotation matrix to (..., 6) 6D rotation."""
    b1 = R[..., 0, :]
    b2 = R[..., 1, :]
    d6 = np.concatenate([b1, b2], axis=-1)
    return d6


def rotation_6d_to_matrix_torch(d6):
    """Torch differentiable version of rotation_6d_to_matrix."""
    a1, a2 = d6[..., :3], d6[..., 3:]
    b1 = F.normalize(a1, dim=-1)
    b2 = a2 - (b1 * a2).sum(dim=-1, keepdim=True) * b1
    b2 = F.normalize(b2, dim=-1)
    b3 = torch.cross(b1, b2, dim=-1)
    return torch.stack((b1, b2, b3), dim=-2)


def matrix_to_axis_angle_torch(R):
    """Torch differentiable version of matrix_to_axis_angle."""
    trace = R[:, 0, 0] + R[:, 1, 1] + R[:, 2, 2]
    cos_angle = ((trace - 1.0) / 2.0).clamp(-1.0 + 1e-7, 1.0 - 1e-7)
    angle = torch.acos(cos_angle)
    rx = R[:, 2, 1] - R[:, 1, 2]
    ry = R[:, 0, 2] - R[:, 2, 0]
    rz = R[:, 1, 0] - R[:, 0, 1]
    axis_raw = torch.stack([rx, ry, rz], dim=1)
    safe_sin = torch.sin(angle).abs().clamp(min=1e-7).unsqueeze(1)
    unit_axis = axis_raw / (2.0 * safe_sin)
    return unit_axis * angle.unsqueeze(1)


# ---------------------------------------------------------------------------
# PoseShield model (from poseshield.common.network)
# ---------------------------------------------------------------------------

class ResidualMLP(nn.Module):
    """Residual MLP: 21x6 joint rotations -> scalar collision field value."""

    def __init__(self, in_dim=126, hidden_dim=512, num_layers=12, activation="relu"):
        super().__init__()
        self.input_layer = nn.Linear(in_dim, hidden_dim)
        self.hidden_layers = nn.ModuleList([
            nn.Linear(hidden_dim, hidden_dim) for _ in range(num_layers)
        ])
        if activation == "relu":
            self.act = nn.ReLU()
        elif activation == "leaky_relu":
            self.act = nn.LeakyReLU()
        elif activation == "elu":
            self.act = nn.ELU()
        else:
            raise ValueError(f"Unsupported activation: {activation}")
        self.output_layer = nn.Linear(hidden_dim, 1)

    def forward(self, x):
        bs = x.shape[0]
        x_reshaped = x.reshape(-1, 6)
        x_raw = x_reshaped[:, :3]
        x_norm = x_raw / x_raw.norm(dim=1, keepdim=True)
        y_raw = x_reshaped[:, 3:]
        dot = (x_norm * y_raw).sum(dim=1, keepdim=True)
        y_perp = y_raw - dot * x_norm
        y_norm = y_perp / y_perp.norm(dim=1, keepdim=True)
        x_valid = torch.cat([x_norm, y_norm], dim=1).reshape(bs, -1)
        x = self.act(self.input_layer(x_valid))
        for layer in self.hidden_layers:
            x = self.act(layer(x)) + x
        return self.output_layer(x)


# ---------------------------------------------------------------------------
# Cost & constraint functions (from poseshield.pose.utils)
# ---------------------------------------------------------------------------

_SUBTREE_SIZES = [4, 4, 13, 3, 3, 12, 2, 2, 11, 1, 1, 2, 4, 4, 1, 3, 3, 2, 2, 1, 1]
SMPLH_POSE_WEIGHTS = torch.tensor(_SUBTREE_SIZES, dtype=torch.float32)
SMPLH_POSE_WEIGHTS /= SMPLH_POSE_WEIGHTS.sum()


def constraint_function(model, x):
    """Collision field value for a pose. Positive = collision-free."""
    output = model(x.unsqueeze(0))
    return output.squeeze(0).squeeze(0)


def cost_function_weighted(x, x_ref, weights=None):
    """Weighted L2 pose distance preserving kinematic-chain importance."""
    if weights is None:
        weights = SMPLH_POSE_WEIGHTS.to(x.device)
    diff = (x - x_ref).reshape(21, 6)
    per_joint_norm = torch.linalg.norm(diff, dim=-1)
    return (per_joint_norm * weights).sum()


def cost_function(x, x_ref):
    return cost_function_weighted(x, x_ref)


# ---------------------------------------------------------------------------
# SLSQP optimizer (from poseshield.pose.resolve_slsqp)
# ---------------------------------------------------------------------------

def optimize_slsqp(sample, model, device, max_itr=300, threshold=0.1,
                   cost_type="normal", tol=0.03):
    """SLSQP optimization to resolve collisions while preserving pose."""
    x0 = sample.reshape(-1).astype(np.float64)
    x_ref_np = x0.copy()

    def to_torch(x_np, requires_grad=False):
        return torch.tensor(x_np, dtype=torch.float32, device=device,
                            requires_grad=requires_grad)

    x_ref_t = to_torch(x_ref_np, requires_grad=False)

    def cost_fn_np(x_np):
        x_t = to_torch(x_np)
        val = cost_function(x_t, x_ref_t) if cost_type != "weighted" else cost_function_weighted(x_t, x_ref_t)
        return float(val.detach().cpu().item())

    def cost_fn_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        val = cost_function(x_t, x_ref_t) if cost_type != "weighted" else cost_function_weighted(x_t, x_ref_t)
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def cons_ineq_fun(x_np):
        x_t = to_torch(x_np)
        val = constraint_function(model, x_t) - threshold
        return float(val.detach().cpu().item())

    def cons_ineq_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        cons_val = constraint_function(model, x_t)
        grad = torch.autograd.grad(cons_val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def view6(x_t):
        return x_t.view(-1, 6)

    def ineq_r1_upper_fun(x_np):
        x_t = to_torch(x_np)
        r1 = view6(x_t)[:, :3]
        return float((1.0 + tol - r1.norm(dim=1).mean()).detach().cpu().item())

    def ineq_r1_upper_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        r1 = view6(x_t)[:, :3]
        val = 1.0 + tol - r1.norm(dim=1).mean()
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def ineq_r1_lower_fun(x_np):
        x_t = to_torch(x_np)
        r1 = view6(x_t)[:, :3]
        return float((r1.norm(dim=1).mean() - (1.0 - tol)).detach().cpu().item())

    def ineq_r1_lower_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        r1 = view6(x_t)[:, :3]
        val = r1.norm(dim=1).mean() - (1.0 - tol)
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def ineq_r2_upper_fun(x_np):
        x_t = to_torch(x_np)
        r2 = view6(x_t)[:, 3:]
        return float((1.0 + tol - r2.norm(dim=1).mean()).detach().cpu().item())

    def ineq_r2_upper_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        r2 = view6(x_t)[:, 3:]
        val = 1.0 + tol - r2.norm(dim=1).mean()
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def ineq_r2_lower_fun(x_np):
        x_t = to_torch(x_np)
        r2 = view6(x_t)[:, 3:]
        return float((r2.norm(dim=1).mean() - (1.0 - tol)).detach().cpu().item())

    def ineq_r2_lower_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        r2 = view6(x_t)[:, 3:]
        val = r2.norm(dim=1).mean() - (1.0 - tol)
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def ineq_orth_upper_fun(x_np):
        x_t = to_torch(x_np)
        r1, r2 = view6(x_t).split(3, dim=1)
        dot_mean = (r1 * r2).sum(dim=1).mean()
        return float((tol - dot_mean).detach().cpu().item())

    def ineq_orth_upper_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        r1, r2 = view6(x_t).split(3, dim=1)
        val = tol - (r1 * r2).sum(dim=1).mean()
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    def ineq_orth_lower_fun(x_np):
        x_t = to_torch(x_np)
        r1, r2 = view6(x_t).split(3, dim=1)
        dot_mean = (r1 * r2).sum(dim=1).mean()
        return float((tol + dot_mean).detach().cpu().item())

    def ineq_orth_lower_jac(x_np):
        x_t = to_torch(x_np, requires_grad=True)
        r1, r2 = view6(x_t).split(3, dim=1)
        val = tol + (r1 * r2).sum(dim=1).mean()
        grad = torch.autograd.grad(val, x_t)[0]
        return grad.detach().cpu().numpy().astype(np.float64)

    constraints = [
        {"type": "ineq", "fun": cons_ineq_fun, "jac": cons_ineq_jac},
        {"type": "ineq", "fun": ineq_r1_upper_fun, "jac": ineq_r1_upper_jac},
        {"type": "ineq", "fun": ineq_r1_lower_fun, "jac": ineq_r1_lower_jac},
        {"type": "ineq", "fun": ineq_r2_upper_fun, "jac": ineq_r2_upper_jac},
        {"type": "ineq", "fun": ineq_r2_lower_fun, "jac": ineq_r2_lower_jac},
        {"type": "ineq", "fun": ineq_orth_upper_fun, "jac": ineq_orth_upper_jac},
        {"type": "ineq", "fun": ineq_orth_lower_fun, "jac": ineq_orth_lower_jac},
    ]

    loss_history, cons_history = [], []

    def callback(xk):
        try:
            loss_history.append(cost_fn_np(xk))
            cons_history.append(constraint_function(model, to_torch(xk)).detach().cpu().item())
        except Exception:
            pass

    res = minimize(
        cost_fn_np, x0, method="SLSQP", jac=cost_fn_jac,
        constraints=constraints, bounds=None,
        options={"maxiter": max_itr, "ftol": 1e-6, "disp": False},
        callback=callback,
    )
    x_opt = res.x.astype(np.float32)
    if len(loss_history) == 0:
        loss_history.append(cost_fn_np(x_opt))
    if len(cons_history) == 0:
        cons_history.append(constraint_function(model, to_torch(x_opt)).detach().cpu().item())
    return x_opt, loss_history, cons_history, bool(res.success), str(res.message)


# ---------------------------------------------------------------------------
# Mesh visualization
# ---------------------------------------------------------------------------

def pose_to_mesh(smpl_model, r_6d, device):
    """Convert 21x6 6D rotation to SMPL mesh vertices and faces."""
    rot_mats = rotation_6d_to_matrix(r_6d)
    axis_angles = matrix_to_axis_angle(rot_mats)
    body_pose = torch.from_numpy(axis_angles.reshape(1, -1)).float().to(device)
    output = smpl_model(
        global_orient=None,
        body_pose=body_pose,
        betas=None,
        transl=None,
        return_verts=True,
    )
    vertices = output.vertices[0].detach().cpu().numpy()
    faces = smpl_model.faces.astype(np.int32)
    return vertices, faces


def visualize_smpl(vertices, faces, save_path, color="#ff7675"):
    """Render SMPL mesh to PNG using matplotlib."""
    fig = plt.figure(figsize=(8, 6), facecolor="white")
    ax = fig.add_subplot(111, projection="3d")
    ax.set_facecolor("white")
    x_plt = vertices[:, 0]
    y_plt = vertices[:, 2]
    z_plt = vertices[:, 1]
    ax.plot_trisurf(x_plt, y_plt, z_plt, triangles=faces,
                    shade=True, color=color, edgecolor="none", alpha=0.9)
    all_coords = np.stack([x_plt, y_plt, z_plt], axis=-1)
    min_vals = np.min(all_coords, axis=0)
    max_vals = np.max(all_coords, axis=0)
    ranges = max_vals - min_vals
    max_range = max(ranges)
    mid = (max_vals + min_vals) / 2
    ax.set_xlim(mid[0] - max_range / 2, mid[0] + max_range / 2)
    ax.set_ylim(mid[1] - max_range / 2, mid[1] + max_range / 2)
    ax.set_zlim(mid[2] - max_range / 2, mid[2] + max_range / 2)
    ax.view_init(elev=15, azim=90)
    ax.axis("off")
    ax.grid(False)
    plt.tight_layout()
    plt.savefig(save_path, dpi=150, bbox_inches="tight", facecolor="white")
    plt.close(fig)


# ---------------------------------------------------------------------------
# Model loading at module scope
# ---------------------------------------------------------------------------

# Download PoseShield model from HF
_model_dir = hf_hub_download("ZYYY99/PoseShield", "model.pth", repo_type="model")
_config_path = hf_hub_download("ZYYY99/PoseShield", "config.yaml", repo_type="model")

with open(_config_path, "r") as f:
    _config = yaml.safe_load(f)

MODEL_HIDDEN_DIM = _config["MODEL"]["HIDDEN_DIM"]
MODEL_NUM_LAYERS = _config["MODEL"]["NUM_LAYERS"]
MODEL_ACTIVATION = _config["MODEL"].get("ACTIVATION", "relu")

# Load collision field model
collision_model = ResidualMLP(
    in_dim=126,
    hidden_dim=MODEL_HIDDEN_DIM,
    num_layers=MODEL_NUM_LAYERS,
    activation=MODEL_ACTIVATION,
).to("cuda")

_ckpt = torch.load(_model_dir, map_location="cuda", weights_only=True)
collision_model.load_state_dict(_ckpt)
collision_model.eval()
print(f"PoseShield collision field loaded: hidden_dim={MODEL_HIDDEN_DIM}, layers={MODEL_NUM_LAYERS}")

# Download SMPL-H neutral body model and set up directory structure for smplx
# The community SMPLH model lacks hand PCA components, so we add dummy ones.
_smplh_file = hf_hub_download("Tevior/smplh", "neutral/model.npz", repo_type="model")
_smplh_root = tempfile.mkdtemp()
_smplh_target_dir = os.path.join(_smplh_root, "smplh")
os.makedirs(_smplh_target_dir, exist_ok=True)

# Load the npz, add missing hand component keys, and re-save for smplx
_orig_smplh = dict(np.load(_smplh_file, allow_pickle=True))
_orig_smplh["hands_componentsl"] = np.zeros((1, 45), dtype=np.float32)
_orig_smplh["hands_componentsr"] = np.zeros((1, 45), dtype=np.float32)
_orig_smplh["hands_meanl"] = np.zeros((45,), dtype=np.float32)
_orig_smplh["hands_meanr"] = np.zeros((45,), dtype=np.float32)
np.savez(os.path.join(_smplh_target_dir, "SMPLH_NEUTRAL.npz"), **_orig_smplh)

_smpl_model = smplx.create(
    _smplh_root,
    model_type="smplh",
    gender="neutral",
    ext="npz",
    use_pca=False,
).to("cuda")
print("SMPL-H neutral body model loaded")


# ---------------------------------------------------------------------------
# Pre-bundled example poses (from the PoseShield demo_asset directory)
# ---------------------------------------------------------------------------

EXAMPLE_POSES = {
    "Colliding Pose #210": "x_ori_210.pkl",
    "Colliding Pose #408": "x_ori_408.pkl",
    "Colliding Pose #436": "x_ori_436.pkl",
}

# The example .pkl files are bundled in the Space repo root (same dir as app.py)
_APP_DIR = os.path.dirname(os.path.abspath(__file__))
_example_files = {}
for name, fname in EXAMPLE_POSES.items():
    local_path = os.path.join(_APP_DIR, fname)
    if os.path.exists(local_path):
        _example_files[name] = local_path
        print(f"Loaded example pose: {name} -> {local_path}")
    else:
        # Fallback: download from the Space repo
        _path = hf_hub_download("hugging-apps/poseshield-collision-fix", fname, repo_type="space")
        _example_files[name] = _path
        print(f"Downloaded example pose: {name} -> {_path}")


def load_pose_pickle(path):
    """Load a pose from a pickle file (format: dict with 'pose' key, shape (63,))."""
    with open(path, "rb") as f:
        data = pickle.load(f, encoding="latin1")
    pose_aa = data["pose"].reshape(21, 3)
    rot_mat = axis_angle_to_matrix(pose_aa)
    rot_6d = matrix_to_rotation_6d(rot_mat)
    sample_flat = rot_6d.reshape(-1)
    return sample_flat, data


# ---------------------------------------------------------------------------
# Inference
# ---------------------------------------------------------------------------

@spaces.GPU(duration=60)
def resolve_pose(pose_file, threshold=0.1, max_itr=150, progress=gr.Progress()):
    """Resolve self-collisions in a SMPL-H pose using the PoseShield collision field.

    Args:
        pose_file: A .pkl file containing a SMPL-H pose with a 'pose' key (21 joints x 3 axis-angle).
        threshold: Collision field threshold (lower = stricter constraint).
        max_itr: Maximum SLSQP optimization iterations.
    """
    if pose_file is None:
        return None, None, "Please upload a pose file or select an example.", ""

    progress(0.1, desc="Loading pose...")
    sample_flat, raw_data = load_pose_pickle(pose_file)

    device = torch.device("cuda")

    progress(0.2, desc="Computing initial collision score...")
    with torch.no_grad():
        init_val = constraint_function(
            collision_model,
            torch.from_numpy(sample_flat).float().to(device),
        ).item()

    progress(0.3, desc="Running SLSQP optimization...")
    start_time = time.time()
    optimized_x, loss_hist, cons_hist, success, message = optimize_slsqp(
        sample_flat, collision_model, device,
        max_itr=max_itr, threshold=threshold,
    )
    elapsed = time.time() - start_time

    progress(0.7, desc="Computing final collision score...")
    final_val = constraint_function(
        collision_model,
        torch.from_numpy(optimized_x).float().to(device),
    ).item()
    final_error = cost_function(
        torch.from_numpy(optimized_x).to(device),
        torch.from_numpy(sample_flat).float().to(device),
    ).item()
    constraint_satisfied = final_val >= threshold

    progress(0.85, desc="Rendering meshes...")

    # Render before and after meshes
    with tempfile.NamedTemporaryFile(suffix="_before.png", delete=False) as f_before:
        before_path = f_before.name
    with tempfile.NamedTemporaryFile(suffix="_after.png", delete=False) as f_after:
        after_path = f_after.name

    r_6d_orig = sample_flat.reshape(21, 6)
    r_6d_opt = optimized_x.reshape(21, 6)

    verts_orig, faces = pose_to_mesh(_smpl_model, r_6d_orig, device)
    visualize_smpl(verts_orig, faces, before_path, color="#ff7675")

    verts_opt, _ = pose_to_mesh(_smpl_model, r_6d_opt, device)
    visualize_smpl(verts_opt, faces, after_path, color="#2ed573")

    progress(1.0, desc="Done!")

    status_text = (
        f"**Optimization Results**\n"
        f"- Time: {elapsed:.2f}s\n"
        f"- Solver success: {success}\n"
        f"- Solver message: {message}\n"
        f"- Initial collision score: {init_val:.6f}\n"
        f"- Final collision score: {final_val:.6f} (threshold: {threshold})\n"
        f"- Constraint satisfied: {constraint_satisfied}\n"
        f"- Mean Vertex Deviation: {final_error:.6f}\n"
        f"- Iterations: {len(loss_hist)}"
    )

    return before_path, after_path, status_text, f"{elapsed:.2f}s"


# ---------------------------------------------------------------------------
# Gradio UI
# ---------------------------------------------------------------------------

CSS = """
#col-container { max-width: 1100px; margin: 0 auto; }
.dark .gradio-container { color: var(--body-text-color); }
"""

with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
    gr.Markdown("""
    # PoseShield: Neural Collision Fields for Human Self-Collision Resolution

    Upload a colliding SMPL-H pose (`.pkl` with a `pose` key of 21×3 axis-angle rotations) or try one of the example poses below.
    PoseShield resolves self-collisions using a learned neural collision field as a differentiable constraint.

    [Paper](https://arxiv.org/abs/2606.29686) | [GitHub](https://github.com/lzhyu/PoseShield) | [Model](https://huggingface.co/ZYYY99/PoseShield)
    """)

    with gr.Row():
        with gr.Column():
            pose_input = gr.File(label="Upload SMPL-H Pose (.pkl)", file_types=[".pkl"])
            with gr.Accordion("Advanced Settings", open=False):
                threshold = gr.Slider(0.01, 1.0, value=0.1, step=0.01,
                                     label="Constraint Threshold (lower = stricter)")
                max_itr = gr.Slider(50, 300, value=150, step=10,
                                    label="Max Optimization Iterations")
            run_btn = gr.Button("Resolve Collisions", variant="primary")
        with gr.Column():
            before_img = gr.Image(label="Before (Colliding)", type="filepath")
            after_img = gr.Image(label="After (Resolved)", type="filepath")
            status_output = gr.Markdown(label="Status")
            time_output = gr.Textbox(label="Inference Time", visible=True)

    gr.Examples(
        examples=[
            [_example_files["Colliding Pose #210"]],
            [_example_files["Colliding Pose #408"]],
            [_example_files["Colliding Pose #436"]],
        ],
        inputs=[pose_input],
        outputs=[before_img, after_img, status_output, time_output],
        fn=resolve_pose,
        cache_examples=True,
        cache_mode="lazy",
    )

    run_btn.click(
        fn=resolve_pose,
        inputs=[pose_input, threshold, max_itr],
        outputs=[before_img, after_img, status_output, time_output],
        api_name="resolve",
    )

demo.launch(mcp_server=True)