scripts/convert_mine_blender.py

"""Convert ``mine_blender`` voxel exports into LiveWorld_comp input cases.

Produces a depth-derived **static** background point cloud — what a depth sensor
would observe of the static scene. Mirroring SPMEM, we raycast the static voxel
cubes from ``--depth-views`` viewpoints sampled uniformly across the model's
generation range and union the first-hit surface points, then voxel-downsample
to dedupe the multi-view overlap. Dynamic (foreground) voxels are deliberately
kept out of ``pointcloud.npz`` (auto fg/bg split): the foreground is conditioned
separately via ``fg_projection.mp4`` / ``fg_mask_first.png``, and the background
point cloud is re-coloured per chunk at inference time.

Per case output (matches ``LiveWorld_comp.core.inputs.load_user_inputs``):

    <name>/
      first_frame.png        copied from ``input_image`` (resized to target_hw)
      prompt.txt             from ``prompt``
      geometry.npz           keys: poses_c2w (N, 4, 4), K (3, 3), intrinsics_size
      pointcloud.npz         keys: points (M, 3)  — STATIC background only
      fg_mask_first.png      from frame_0001 dyn voxels projected into pose 0
"""
from __future__ import annotations

import argparse
import json
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, List, Optional, Tuple

import cv2
import numpy as np
from PIL import Image


TARGET_W_DEFAULT = 832
TARGET_H_DEFAULT = 480


# ---------------------------------------------------------------------------
# Camera helpers (mirrors versecrafter_compare/render_controls.py)
# ---------------------------------------------------------------------------

def compute_intrinsics_px(intr: dict,
                          target_w: int,
                          target_h: int) -> Tuple[float, float, float, float]:
    """Convert Blender sensor intrinsics to pixel intrinsics at target_hw."""
    focal = float(intr["focal_length"])
    sw = float(intr["sensor_width"])
    sh = float(intr["sensor_height"])
    res_x = int(intr["resolution_x"])
    res_y = int(intr["resolution_y"])
    fit = intr.get("sensor_fit", "AUTO")

    if fit == "VERTICAL":
        fpx = focal / sh * res_y
    elif fit == "HORIZONTAL":
        fpx = focal / sw * res_x
    else:  # AUTO
        fpx = focal / sw * res_x if res_x >= res_y else focal / sh * res_y

    fx = fpx * target_w / res_x
    fy = fpx * target_h / res_y
    cx = target_w / 2.0
    cy = target_h / 2.0
    return fx, fy, cx, cy


def blender_c2w_to_opencv_c2w(c2w_blender: np.ndarray) -> np.ndarray:
    """Flip cam-local Y and Z columns to convert Blender → OpenCV camera frame."""
    c2w = c2w_blender.astype(np.float32).copy()
    c2w[..., :3, 1:3] *= -1
    return c2w


def K_matrix(fx: float, fy: float, cx: float, cy: float) -> np.ndarray:
    return np.array([[fx, 0.0, cx],
                     [0.0, fy, cy],
                     [0.0, 0.0, 1.0]], dtype=np.float32)


# ---------------------------------------------------------------------------
# Cube-surface point sampling (with per-point face normals)
# ---------------------------------------------------------------------------

_FACE_NORMALS = np.array([
    [-1, 0, 0], [+1, 0, 0],   # ±X
    [0, -1, 0], [0, +1, 0],   # ±Y
    [0, 0, -1], [0, 0, +1],   # ±Z
], dtype=np.float32)


def sample_cube_surface_points(centers: np.ndarray,
                               sizes: np.ndarray,
                               points_per_voxel: int,
                               rng: np.random.Generator,
                               return_normals: bool = False
                               ) -> Tuple[np.ndarray, np.ndarray]:
    """Stratified per-face sampling on each voxel's surface.

    Returns
    -------
    points  : (N * K, 3) float32  — surface samples
    normals : (N * K, 3) float32  — face normal per point (only if requested)
    """
    N = len(centers)
    if N == 0:
        empty = np.zeros((0, 3), dtype=np.float32)
        return empty, empty
    K = int(points_per_voxel)
    if K <= 0:
        # Fall back to one sample at each voxel center; cannot derive normals.
        pts = centers.astype(np.float32).copy()
        nm = np.zeros_like(pts)
        return pts, nm

    per_face = max(1, (K + 5) // 6)
    total_per_voxel = per_face * 6

    face_id = np.repeat(np.arange(6, dtype=np.int32), per_face)        # (6P,)
    face_axis = face_id // 2                                            # 0/1/2
    face_sign = ((face_id % 2) * 2 - 1).astype(np.float32)              # -1/+1
    uv = rng.uniform(-0.5, 0.5, size=(total_per_voxel, 2)).astype(np.float32)

    unit_pts = np.zeros((total_per_voxel, 3), dtype=np.float32)
    for f in range(6):
        s, e = f * per_face, (f + 1) * per_face
        axis = f // 2
        sign = float(face_sign[s])
        other_axes = [a for a in range(3) if a != axis]
        unit_pts[s:e, axis] = sign * 0.5
        unit_pts[s:e, other_axes[0]] = uv[s:e, 0]
        unit_pts[s:e, other_axes[1]] = uv[s:e, 1]

    unit_normals = _FACE_NORMALS[face_id]                               # (6P, 3)

    keep_slice = slice(None)
    if total_per_voxel > K:
        keep_idx = rng.choice(total_per_voxel, size=K, replace=False)
        unit_pts = unit_pts[keep_idx]
        unit_normals = unit_normals[keep_idx]
        keep_slice = keep_idx  # noqa: F841 — kept for clarity

    sizes_b = sizes.astype(np.float32).reshape(N, 1, 1)
    pts = centers.astype(np.float32).reshape(N, 1, 3) + unit_pts.reshape(1, -1, 3) * sizes_b
    pts = pts.reshape(-1, 3)

    if return_normals:
        nm = np.tile(unit_normals.reshape(1, -1, 3), (N, 1, 1)).reshape(-1, 3)
        return pts, nm
    return pts, np.zeros_like(pts)


# ---------------------------------------------------------------------------
# Visibility filters
# ---------------------------------------------------------------------------

def cull_back_faces(points: np.ndarray,
                    normals: np.ndarray,
                    cam_positions: np.ndarray) -> np.ndarray:
    """True if the point's face is front-facing toward AT LEAST one camera."""
    visible = np.zeros(len(points), dtype=bool)
    for cam in cam_positions:
        diff = cam[None] - points
        dot = (normals * diff).sum(axis=1)
        visible |= dot > 0.0
    return visible


def voxel_downsample_points(points: np.ndarray, voxel_size: float) -> np.ndarray:
    if voxel_size <= 0 or len(points) == 0:
        return points
    keys = np.floor(points / voxel_size).astype(np.int64)
    _, unique_idx = np.unique(keys, axis=0, return_index=True)
    unique_idx.sort()
    return points[unique_idx]


def voxel_downsample_with_colors(points: np.ndarray,
                                 colors: np.ndarray,
                                 voxel_size: float
                                 ) -> Tuple[np.ndarray, np.ndarray]:
    """Voxel-downsample ``points`` while keeping the matching color of the
    representative sample. Returns ``(pts, cols)``."""
    if voxel_size <= 0 or len(points) == 0:
        return points, colors
    keys = np.floor(points / voxel_size).astype(np.int64)
    _, unique_idx = np.unique(keys, axis=0, return_index=True)
    unique_idx.sort()
    return points[unique_idx], colors[unique_idx]


def build_unit_cube_surface_samples(K: int, seed: int = 0) -> np.ndarray:
    """Return ``(K, 3) float32`` points uniformly distributed on the surface
    of a unit cube centred at origin (so in ``[-0.5, 0.5]^3``).

    Used as a SHARED per-voxel template applied via::

        world_pt = unit_sample * voxel_size + voxel_center

    By sharing one template across all voxels we get deterministic, time-
    consistent per-voxel splats: each voxel's K samples follow the voxel
    rigidly across frames, instead of re-randomising per-frame.
    """
    rng = np.random.default_rng(seed)
    per_face = max(1, (int(K) + 5) // 6)
    total = per_face * 6
    face_id = np.repeat(np.arange(6, dtype=np.int32), per_face)
    axis = face_id // 2
    side = ((face_id % 2) * 2 - 1).astype(np.float32) * 0.5
    uv = rng.uniform(-0.5, 0.5, size=(total, 2)).astype(np.float32)
    samples = np.zeros((total, 3), dtype=np.float32)
    for f in range(6):
        s, e = f * per_face, (f + 1) * per_face
        a = f // 2
        others = [aa for aa in range(3) if aa != a]
        samples[s:e, a] = float(side[s])
        samples[s:e, others[0]] = uv[s:e, 0]
        samples[s:e, others[1]] = uv[s:e, 1]
    if total > K:
        idx = rng.choice(total, size=int(K), replace=False)
        samples = samples[idx]
    return samples


def transform_unit_samples_per_voxel(unit_samples: np.ndarray,
                                     centers: np.ndarray,
                                     sizes: np.ndarray) -> np.ndarray:
    """Apply ``(size_i, center_i)`` to a SHARED unit template per voxel.
    Returns ``(N * K, 3)`` where rows ``[i*K:(i+1)*K)`` belong to voxel ``i``.
    """
    N = len(centers)
    if N == 0:
        return np.zeros((0, 3), dtype=np.float32)
    s = sizes.reshape(N, 1, 1).astype(np.float32)
    c = centers.reshape(N, 1, 3).astype(np.float32)
    u = unit_samples.reshape(1, -1, 3).astype(np.float32)
    return (u * s + c).reshape(-1, 3)


# ---------------------------------------------------------------------------
# Depth-image based PC via Open3D raycasting (the correct way)
# ---------------------------------------------------------------------------

_CUBE_UNIT_VERTS = np.array([
    [-0.5, -0.5, -0.5], [+0.5, -0.5, -0.5], [+0.5, +0.5, -0.5], [-0.5, +0.5, -0.5],
    [-0.5, -0.5, +0.5], [+0.5, -0.5, +0.5], [+0.5, +0.5, +0.5], [-0.5, +0.5, +0.5],
], dtype=np.float32)

_CUBE_UNIT_FACES = np.array([
    [0, 2, 1], [0, 3, 2],  # -Z
    [4, 5, 6], [4, 6, 7],  # +Z
    [0, 1, 5], [0, 5, 4],  # -Y
    [3, 6, 2], [3, 7, 6],  # +Y
    [0, 7, 3], [0, 4, 7],  # -X
    [1, 2, 6], [1, 6, 5],  # +X
], dtype=np.int32)


def build_cube_triangle_mesh(centers: np.ndarray,
                             sizes: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
    """Build axis-aligned cube mesh: returns (verts (N*8, 3), faces (N*12, 3))."""
    N = len(centers)
    if N == 0:
        return (np.zeros((0, 3), np.float32),
                np.zeros((0, 3), np.int32))
    sizes_b = sizes.astype(np.float32).reshape(N, 1, 1)
    verts = (_CUBE_UNIT_VERTS[None] * sizes_b +
             centers[:, None, :].astype(np.float32)).reshape(-1, 3)
    offsets = (np.arange(N, dtype=np.int32) * 8).reshape(N, 1, 1)
    faces = (_CUBE_UNIT_FACES[None] + offsets).reshape(-1, 3).astype(np.int32)
    return verts.astype(np.float32), faces.astype(np.int32)


def _build_scene_for_frame(static_v: np.ndarray, static_f: np.ndarray,
                           dyn_c: np.ndarray, dyn_s: np.ndarray
                           ) -> Tuple[Optional[object], int, int]:
    """Build an Open3D ``RaycastingScene`` for one frame with **separate**
    geometry ids for static and dyn cubes.

    Returns ``(scene, static_id, dyn_id)`` where ``static_id`` / ``dyn_id`` are
    Open3D geometry handles (or ``-1`` when that layer is empty). ``scene`` is
    ``None`` if both layers are empty.
    """
    import open3d as o3d
    import open3d.core as o3c

    dyn_v, dyn_f = build_cube_triangle_mesh(dyn_c, dyn_s)
    if len(static_v) == 0 and len(dyn_v) == 0:
        return None, -1, -1

    scene = o3d.t.geometry.RaycastingScene()
    static_id = -1
    dyn_id = -1
    if len(static_v) > 0:
        static_id = int(scene.add_triangles(
            o3c.Tensor(static_v, dtype=o3c.float32),
            o3c.Tensor(static_f.astype(np.uint32), dtype=o3c.uint32),
        ))
    if len(dyn_v) > 0:
        dyn_id = int(scene.add_triangles(
            o3c.Tensor(dyn_v, dtype=o3c.float32),
            o3c.Tensor(dyn_f.astype(np.uint32), dtype=o3c.uint32),
        ))
    return scene, static_id, dyn_id


def raycast_static_at_camera0(static_centers: np.ndarray,
                              static_sizes: np.ndarray,
                              dyn_centers_at_0: np.ndarray,
                              dyn_sizes_at_0: np.ndarray,
                              pose_c2w_0: np.ndarray,
                              K: np.ndarray,
                              H: int, W: int,
                              max_depth: float = 200.0
                              ) -> np.ndarray:
    """Depth-derived static PC visible from **camera 0 only**.

    Why frame 0 only: with the current coloring strategy (only points whose
    projection to ``first_frame.png`` is in-view AND outside ``fg_mask``
    get a color), any static surface visible only from later frames CANNOT
    be coloured and will be dropped downstream. Casting rays for frames
    1..N-1 produces hits that are guaranteed to be discarded — pure waste.

    Includes ``dyn[0]`` as occluder so dyn-occluded static doesn't leak in.
    Hits are un-deduped; caller voxel-downsamples.
    """
    import open3d.core as o3c

    static_v, static_f = build_cube_triangle_mesh(static_centers, static_sizes)
    scene, static_id, _ = _build_scene_for_frame(
        static_v, static_f, dyn_centers_at_0, dyn_sizes_at_0
    )
    if scene is None or static_id < 0:
        return np.zeros((0, 3), dtype=np.float32)

    intrinsic = o3c.Tensor(K.astype(np.float64), dtype=o3c.float64)
    w2c = np.linalg.inv(pose_c2w_0).astype(np.float64)
    rays = scene.create_rays_pinhole(
        intrinsic_matrix=intrinsic,
        extrinsic_matrix=o3c.Tensor(w2c, dtype=o3c.float64),
        width_px=W, height_px=H,
    )
    ans = scene.cast_rays(rays)
    t_hit = ans["t_hit"].numpy()
    geom_ids = ans["geometry_ids"].numpy()
    valid = np.isfinite(t_hit) & (t_hit > 1e-3) & (t_hit < max_depth)
    sm = valid & (geom_ids == static_id)
    if not sm.any():
        return np.zeros((0, 3), dtype=np.float32)

    rays_np = rays.numpy()
    hits = rays_np[..., :3] + rays_np[..., 3:] * t_hit[..., None]
    return hits[sm].reshape(-1, 3).astype(np.float32)


def raycast_static_multiview(static_centers: np.ndarray,
                             static_sizes: np.ndarray,
                             poses_c2w_views: np.ndarray,
                             K: np.ndarray,
                             H: int, W: int,
                             max_depth: float = 200.0
                             ) -> np.ndarray:
    """Depth-derived static PC visible from a set of camera views (union).

    Mirrors SPMEM's ``voxels_to_depth_pointcloud``: raycast the **static** voxel
    cubes from several viewpoints sampled uniformly across the trajectory and
    union the first-hit surface points. This yields a depth-sensor-like cloud
    that covers everything the camera eventually sees — not just frame 0.

    No dynamic occluders are added: the static background hidden behind a moving
    object from one view is still captured from other views, and gets coloured
    later by the per-chunk progressive coloring. Hits are un-deduped; the caller
    voxel-downsamples.
    """
    import open3d.core as o3c

    static_v, static_f = build_cube_triangle_mesh(static_centers, static_sizes)
    scene, static_id, _ = _build_scene_for_frame(
        static_v, static_f,
        np.zeros((0, 3), dtype=np.float32),
        np.zeros((0,), dtype=np.float32),
    )
    if scene is None or static_id < 0:
        return np.zeros((0, 3), dtype=np.float32)

    intrinsic = o3c.Tensor(K.astype(np.float64), dtype=o3c.float64)
    views = np.asarray(poses_c2w_views, dtype=np.float64)
    if views.ndim == 2:
        views = views[None, ...]

    all_hits: List[np.ndarray] = []
    for c2w in views:
        w2c = np.linalg.inv(c2w)
        rays = scene.create_rays_pinhole(
            intrinsic_matrix=intrinsic,
            extrinsic_matrix=o3c.Tensor(w2c, dtype=o3c.float64),
            width_px=W, height_px=H,
        )
        ans = scene.cast_rays(rays)
        t_hit = ans["t_hit"].numpy()
        geom_ids = ans["geometry_ids"].numpy()
        valid = (np.isfinite(t_hit) & (t_hit > 1e-3)
                 & (t_hit < max_depth) & (geom_ids == static_id))
        if not valid.any():
            continue
        rays_np = rays.numpy()
        hits = rays_np[..., :3] + rays_np[..., 3:] * t_hit[..., None]
        all_hits.append(hits[valid].reshape(-1, 3).astype(np.float32))

    if not all_hits:
        return np.zeros((0, 3), dtype=np.float32)
    return np.concatenate(all_hits, axis=0)


def _project_points_uvz(points_world: np.ndarray,
                        c2w: np.ndarray,
                        K: np.ndarray,
                        H: int, W: int,
                        max_depth: float = 200.0
                        ) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
    """Project (N, 3) points into camera ``c2w``. Returns ``(u_i, v_i, z, valid)``
    arrays of shape ``(N,)``. ``valid`` is the in-view & in-front mask."""
    if len(points_world) == 0:
        empty_i = np.zeros((0,), dtype=np.int64)
        empty_f = np.zeros((0,), dtype=np.float32)
        empty_b = np.zeros((0,), dtype=bool)
        return empty_i, empty_i, empty_f, empty_b

    w2c = np.linalg.inv(c2w.astype(np.float64)).astype(np.float32)
    R, t = w2c[:3, :3], w2c[:3, 3]
    pts_cam = points_world.astype(np.float32) @ R.T + t
    z = pts_cam[:, 2]
    safe_z = np.where(z > 1e-6, z, 1.0)
    fx, fy = float(K[0, 0]), float(K[1, 1])
    cx, cy = float(K[0, 2]), float(K[1, 2])
    u_f = pts_cam[:, 0] / safe_z * fx + cx
    v_f = pts_cam[:, 1] / safe_z * fy + cy
    u_i = np.rint(u_f).astype(np.int64)
    v_i = np.rint(v_f).astype(np.int64)
    valid = (
        (z > 1e-3) & (z < max_depth)
        & np.isfinite(u_f) & np.isfinite(v_f)
        & (u_i >= 0) & (u_i < W) & (v_i >= 0) & (v_i < H)
    )
    return u_i, v_i, z.astype(np.float32), valid


def sample_pc_colors_from_image(points_world: np.ndarray,
                                c2w: np.ndarray,
                                K: np.ndarray,
                                image_rgb: np.ndarray,
                                fg_mask: Optional[np.ndarray] = None,
                                default_color: Tuple[int, int, int] = (128, 128, 128),
                                max_depth: float = 200.0,
                                return_keep_mask: bool = False):
    """Sample per-point RGB from ``image_rgb`` at where each point projects.

    Points that fall outside the view, behind the camera, or onto pixels
    where ``fg_mask`` is ``True`` get ``default_color`` instead. Returns
    ``(N, 3) uint8`` colors; if ``return_keep_mask`` is True, also returns
    a ``(N,) bool`` mask that is True exactly for points that successfully
    sampled a real color (in-view AND not masked out).
    """
    H, W = image_rgb.shape[:2]
    N = len(points_world)
    out = np.tile(np.array(default_color, dtype=np.uint8), (N, 1))
    keep = np.zeros(N, dtype=bool)
    if N == 0:
        return (out, keep) if return_keep_mask else out
    u_i, v_i, _, valid = _project_points_uvz(points_world, c2w, K, H, W, max_depth)
    if not valid.any():
        return (out, keep) if return_keep_mask else out
    sampled = image_rgb[v_i[valid], u_i[valid]]                # (M, 3) uint8
    if fg_mask is not None:
        on_fg = fg_mask[v_i[valid], u_i[valid]].astype(bool)
        ok = ~on_fg
        idx_valid = np.nonzero(valid)[0]
        out[idx_valid[ok]] = sampled[ok]
        keep[idx_valid[ok]] = True
    else:
        out[valid] = sampled
        keep[valid] = True
    return (out, keep) if return_keep_mask else out


def color_points_from_first_frame(points_frame0: np.ndarray,
                                  c2w_0: np.ndarray,
                                  K: np.ndarray,
                                  first_rgb: np.ndarray,
                                  fg_mask: Optional[np.ndarray] = None,
                                  ) -> Tuple[np.ndarray, np.ndarray]:
    """Sample per-point RGB from ``first_rgb`` at each point's camera-0 projection.

    Points that (a) project outside the camera, or (b) land outside
    ``fg_mask`` (when provided) are flagged as un-colored.

    Returns ``(point_color (N, 3) uint8, point_has_color (N,) bool)``.
    """
    N = len(points_frame0)
    point_color = np.zeros((N, 3), dtype=np.uint8)
    point_has_color = np.zeros(N, dtype=bool)
    if N == 0:
        return point_color, point_has_color
    H, W = first_rgb.shape[:2]
    u_i, v_i, _, valid = _project_points_uvz(points_frame0, c2w_0, K, H, W)
    if not valid.any():
        return point_color, point_has_color
    sampled = first_rgb[v_i[valid], u_i[valid]]
    if fg_mask is not None:
        in_fg = fg_mask[v_i[valid], u_i[valid]].astype(bool)
        ok = in_fg
    else:
        ok = np.ones(int(valid.sum()), dtype=bool)
    idx_valid = np.nonzero(valid)[0]
    point_color[idx_valid[ok]] = sampled[ok]
    point_has_color[idx_valid[ok]] = True
    return point_color, point_has_color


def project_pc_zbuffer_depth(points_world: np.ndarray,
                             c2w: np.ndarray,
                             K: np.ndarray,
                             H: int, W: int,
                             max_depth: float = 200.0) -> np.ndarray:
    """LiveWorld-style point splat: each 3-D point → 1 pixel, z-buffer keeps
    the closest one. Returns an ``(H, W)`` depth map (``np.inf`` where no
    point projected).

    This mirrors ``LiveWorld/scripts/create_train_data/_projection.py::
    _compute_zbuffer`` — the very same projection the diffusion model sees
    after VAE encoding.
    """
    out = np.full((H, W), np.inf, dtype=np.float32)
    if len(points_world) == 0:
        return out

    u_i, v_i, z, valid = _project_points_uvz(points_world, c2w, K, H, W, max_depth)
    if not valid.any():
        return out

    flat = (v_i[valid] * W + u_i[valid]).astype(np.int64)
    z_val = z[valid].astype(np.float32)
    order = np.argsort(z_val)
    flat_sorted = flat[order]
    z_sorted = z_val[order]
    unique_flat, first_idx = np.unique(flat_sorted, return_index=True)
    flat_buf = np.full(H * W, np.inf, dtype=np.float32)
    flat_buf[unique_flat] = z_sorted[first_idx]
    return flat_buf.reshape(H, W)


def project_pc_zbuffer_rgb(points_world: np.ndarray,
                           colors_uint8: np.ndarray,
                           c2w: np.ndarray,
                           K: np.ndarray,
                           H: int, W: int,
                           max_depth: float = 200.0,
                           bg_rgb: Tuple[int, int, int] = (0, 0, 0),
                           splat_size: int = 1,
                           ) -> np.ndarray:
    """LiveWorld-style point splat with **per-point RGB colors**: each point
    paints a ``splat_size × splat_size`` block of pixels with its color;
    z-buffer keeps the closest point per pixel. Pixels with no point retain
    ``bg_rgb``. Returns ``(H, W, 3) uint8``.

    ``splat_size=1`` (default) → exact LiveWorld behaviour (1 pixel per
    point). ``splat_size=2`` fills typical 1cm-grid moiré gaps from oblique
    angles at ~4× the splat cost.
    """
    out = np.tile(np.array(bg_rgb, dtype=np.uint8), (H, W, 1))
    if len(points_world) == 0:
        return out

    u_i, v_i, z, valid = _project_points_uvz(points_world, c2w, K, H, W, max_depth)
    if not valid.any():
        return out

    u_v = u_i[valid]
    v_v = v_i[valid]
    z_v = z[valid].astype(np.float32)
    c_v = colors_uint8[valid].astype(np.uint8)

    if splat_size > 1:
        # Replicate each projected point with offsets in a splat_size × splat_size
        # block centred near the projected pixel. Z and color stay the same.
        offsets = np.array(
            [(du, dv) for du in range(splat_size) for dv in range(splat_size)],
            dtype=np.int64,
        )
        N_off = len(offsets)
        u_v = np.repeat(u_v, N_off) + np.tile(offsets[:, 0], len(z_v))
        v_v = np.repeat(v_v, N_off) + np.tile(offsets[:, 1], len(z_v))
        z_v = np.repeat(z_v, N_off)
        c_v = np.repeat(c_v, N_off, axis=0)
        inb = (u_v >= 0) & (u_v < W) & (v_v >= 0) & (v_v < H)
        u_v = u_v[inb]
        v_v = v_v[inb]
        z_v = z_v[inb]
        c_v = c_v[inb]
        if len(z_v) == 0:
            return out

    flat = (v_v * W + u_v).astype(np.int64)
    order = np.argsort(z_v)
    flat_sorted = flat[order]
    cols_sorted = c_v[order]
    unique_flat, first_idx = np.unique(flat_sorted, return_index=True)
    rgb_flat = out.reshape(H * W, 3).copy()
    rgb_flat[unique_flat] = cols_sorted[first_idx]
    return rgb_flat.reshape(H, W, 3)


def render_pc_video_depth(pc_or_dict,
                          poses_c2w: np.ndarray,
                          K: np.ndarray,
                          H: int, W: int,
                          *,
                          is_static: bool,
                          max_depth: float = 200.0) -> np.ndarray:
    """Project a PC through every camera in ``poses_c2w`` via z-buffer splat.

    - If ``is_static`` is ``True``, ``pc_or_dict`` is a ``(M, 3)`` array and
      the same PC is projected every frame.
    - Else, ``pc_or_dict`` is a ``Dict[int, (Mt, 3)]`` of per-frame PCs.

    Returns ``(T, H, W)`` float depth (``np.inf`` at uncovered pixels).
    """
    T = len(poses_c2w)
    out = np.full((T, H, W), np.inf, dtype=np.float32)
    for t in range(T):
        if is_static:
            pts = pc_or_dict
        else:
            pts = pc_or_dict.get(t, np.zeros((0, 3), dtype=np.float32))
        if len(pts) == 0:
            continue
        out[t] = project_pc_zbuffer_depth(pts, poses_c2w[t], K, H, W, max_depth)
    return out


def render_pc_video_rgb(pc_or_dict,
                        colors_or_dict,
                        poses_c2w: np.ndarray,
                        K: np.ndarray,
                        H: int, W: int,
                        *,
                        is_static: bool,
                        max_depth: float = 200.0,
                        splat_size: int = 1) -> np.ndarray:
    """Like :func:`render_pc_video_depth` but writes RGB per-point colors.

    - ``is_static=True``: ``pc_or_dict`` is ``(M, 3)`` and ``colors_or_dict``
      is ``(M, 3)``; the same colored PC is projected every frame.
    - ``is_static=False``: both are ``Dict[int, ndarray]`` keyed by frame.

    ``splat_size`` is forwarded to :func:`project_pc_zbuffer_rgb`.
    Returns ``(T, H, W, 3) uint8``.
    """
    T = len(poses_c2w)
    out = np.zeros((T, H, W, 3), dtype=np.uint8)
    for t in range(T):
        if is_static:
            pts = pc_or_dict
            cols = colors_or_dict
        else:
            pts = pc_or_dict.get(t, np.zeros((0, 3), dtype=np.float32))
            cols = colors_or_dict.get(t, np.zeros((0, 3), dtype=np.uint8))
        if len(pts) == 0:
            continue
        out[t] = project_pc_zbuffer_rgb(
            pts, cols, poses_c2w[t], K, H, W, max_depth,
            splat_size=splat_size,
        )
    return out


# ---------------------------------------------------------------------------
# Depth-frame visualization helpers
# ---------------------------------------------------------------------------

def depth_to_grayscale_frames(depths: np.ndarray,
                              percentile: Tuple[float, float] = (2.0, 98.0),
                              colormap: Optional[int] = None) -> np.ndarray:
    """Convert (T, H, W) float depth to (T, H, W, 3) uint8 RGB frames.

    Near pixels are bright, far are dim, no-hit (inf) is black. Depth range is
    derived from percentile clipping across all valid pixels in the stack so
    every frame uses the same brightness mapping.
    """
    valid_mask = np.isfinite(depths)
    if not valid_mask.any():
        return np.zeros((*depths.shape, 3), dtype=np.uint8)
    valid_d = depths[valid_mask]
    near = float(np.percentile(valid_d, percentile[0]))
    far = float(np.percentile(valid_d, percentile[1]))
    if far <= near:
        far = near + 1.0

    norm = np.clip((depths - near) / (far - near), 0.0, 1.0)
    gray = ((1.0 - norm) * 255).astype(np.uint8)
    gray[~valid_mask] = 0

    if colormap is None:
        return np.stack([gray, gray, gray], axis=-1)

    out = np.empty((*depths.shape, 3), dtype=np.uint8)
    for t in range(depths.shape[0]):
        rgb = cv2.applyColorMap(gray[t], colormap)
        rgb = cv2.cvtColor(rgb, cv2.COLOR_BGR2RGB)
        rgb[~valid_mask[t]] = 0
        out[t] = rgb
    return out


def _resolve_ffmpeg_binary() -> Optional[str]:
    """Locate an ffmpeg binary, preferring system ffmpeg, falling back to
    the one bundled with ``imageio_ffmpeg`` (always installed in xfuser env).
    """
    import shutil as _sh
    bin_path = _sh.which("ffmpeg")
    if bin_path:
        return bin_path
    try:
        import imageio_ffmpeg
        return imageio_ffmpeg.get_ffmpeg_exe()
    except Exception:
        return None


def save_video_mp4(path: Path,
                   frames: np.ndarray,
                   fps: float = 16.0) -> None:
    """Write ``(T, H, W, 3) uint8`` RGB frames as MP4 (H.264, yuv420p,
    +faststart) via ffmpeg subprocess — same recipe as
    ``liveworld.utils.save_video_h264``. Falls back to ``cv2.VideoWriter``
    only if no ffmpeg binary is reachable.

    Why not cv2's mp4v writer? It writes raw RGB into an MPEG-4 Part 2
    container without proper YUV color-space conversion. Many players
    (Cursor's preview, Chrome, QuickTime) interpret the resulting frames as
    YUV420 and render them as a uniform green tint. Using libx264 +
    yuv420p + faststart is the universal MP4 recipe.
    """
    import subprocess

    p = Path(path)
    p.parent.mkdir(parents=True, exist_ok=True)
    if frames.size == 0:
        return
    if frames.dtype != np.uint8:
        frames = frames.astype(np.uint8)
    T, H, W, C = frames.shape
    if C != 3:
        raise ValueError(f"save_video_mp4 expects (T, H, W, 3) RGB; got {frames.shape}")

    ffmpeg_bin = _resolve_ffmpeg_binary()
    if ffmpeg_bin is not None:
        cmd = [
            ffmpeg_bin, "-y",
            "-f", "rawvideo",
            "-pix_fmt", "rgb24",
            "-s", f"{W}x{H}",
            "-r", str(float(fps)),
            "-i", "-",
            "-an",
            "-c:v", "libx264",
            "-preset", "medium",
            "-crf", "18",
            "-pix_fmt", "yuv420p",
            "-movflags", "+faststart",
            str(p),
        ]
        proc = subprocess.Popen(
            cmd, stdin=subprocess.PIPE,
            stdout=subprocess.DEVNULL, stderr=subprocess.PIPE,
        )
        try:
            assert proc.stdin is not None
            for i in range(T):
                proc.stdin.write(frames[i].tobytes())
            proc.stdin.close()
            assert proc.stderr is not None
            stderr = proc.stderr.read()
            ret = proc.wait()
            if ret != 0:
                raise RuntimeError(
                    f"ffmpeg failed ({ret}) writing {p}:\n"
                    f"{stderr.decode('utf-8', 'replace')}"
                )
        except Exception:
            proc.kill()
            raise
        return

    # Last-resort fallback: cv2 mp4v (may render green in some players).
    print(f"[warn] ffmpeg not found; falling back to cv2 mp4v writer for {p}")
    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
    writer = cv2.VideoWriter(str(p), fourcc, float(fps), (W, H))
    if not writer.isOpened():
        raise RuntimeError(f"cv2.VideoWriter failed to open: {p}")
    for f in frames:
        writer.write(cv2.cvtColor(f, cv2.COLOR_RGB2BGR))
    writer.release()


# ---------------------------------------------------------------------------
# Voxel collection
# ---------------------------------------------------------------------------

@dataclass
class VoxelChunk:
    centers: np.ndarray
    sizes: np.ndarray


def _collect_static_voxels(static_npz_path: Path,
                           objects_info: Dict[str, dict],
                           default_voxel_size: float,
                           voxel_size_scale: float) -> VoxelChunk:
    if not static_npz_path.exists():
        raise FileNotFoundError(f"static.npz not found: {static_npz_path}")
    npz = np.load(static_npz_path)
    try:
        centers_chunks, size_chunks = [], []
        for name in npz.files:
            info = objects_info.get(name, {})
            if info.get("type") != "static":
                continue
            pts = npz[name].astype(np.float32)
            if pts.size == 0:
                continue
            vs = float(info.get("voxel_size", default_voxel_size)) * voxel_size_scale
            centers_chunks.append(pts)
            size_chunks.append(np.full(pts.shape[0], vs, dtype=np.float32))
    finally:
        npz.close()
    if not centers_chunks:
        return VoxelChunk(np.zeros((0, 3), np.float32), np.zeros((0,), np.float32))
    return VoxelChunk(
        np.concatenate(centers_chunks, axis=0),
        np.concatenate(size_chunks, axis=0),
    )


def _collect_dyn_voxels_for_frame(ved_dir: Path,
                                  frame_meta: dict,
                                  objects_info: Dict[str, dict],
                                  default_voxel_size: float,
                                  voxel_size_scale: float) -> VoxelChunk:
    npz_path = ved_dir / frame_meta["data_file"]
    if not npz_path.exists():
        return VoxelChunk(np.zeros((0, 3), np.float32), np.zeros((0,), np.float32))
    npz = np.load(npz_path)
    centers_chunks, size_chunks = [], []
    try:
        for name in npz.files:
            info = objects_info.get(name, {})
            if info.get("type") != "dynamic":
                continue
            pts = npz[name].astype(np.float32)
            if pts.size == 0:
                continue
            vs = float(info.get("voxel_size", default_voxel_size)) * voxel_size_scale
            centers_chunks.append(pts)
            size_chunks.append(np.full(pts.shape[0], vs, dtype=np.float32))
    finally:
        npz.close()
    if not centers_chunks:
        return VoxelChunk(np.zeros((0, 3), np.float32), np.zeros((0,), np.float32))
    return VoxelChunk(
        np.concatenate(centers_chunks, axis=0),
        np.concatenate(size_chunks, axis=0),
    )


def _collect_dyn_per_object_persistent(ved_dir: Path,
                                       frames_meta: List[dict],
                                       objects_info: Dict[str, dict],
                                       default_voxel_size: float,
                                       voxel_size_scale: float,
                                       n_use: int
                                       ) -> Tuple[Dict[int, np.ndarray],
                                                  Dict[int, np.ndarray],
                                                  np.ndarray,
                                                  List[str]]:
    """Collect dyn voxels per frame with **persistent per-voxel identity**.

    Unlike :func:`_collect_dyn_voxels_for_frame` (which iterates ``npz.files``
    in arbitrary order and then dedups), this function:

    - Iterates dynamic object names in a deterministic (alphabetical) order
      so the row layout of the returned ``(K_total, 3)`` array is identical
      across frames.
    - Skips dedup so row ``i`` in frame ``t`` is the same physical voxel as
      row ``i`` in frame ``0`` (per-object index → global combined index).
    - Verifies that every frame has the same ``K_total`` and same per-object
      counts; raises if the upstream data breaks identity.

    Returns:
        ``(centers_per_frame, sizes_per_frame, voxel_obj_id, dyn_names)``
        where ``voxel_obj_id`` is ``(K_total,) int32`` indexing into
        ``dyn_names`` (same array reused for every frame).
    """
    dyn_names = sorted(
        n for n, info in objects_info.items() if info.get("type") == "dynamic"
    )
    centers_per_frame: Dict[int, np.ndarray] = {}
    sizes_per_frame: Dict[int, np.ndarray] = {}
    K_total_ref: Optional[int] = None
    voxel_obj_id_ref: Optional[np.ndarray] = None
    per_obj_count_ref: Optional[List[int]] = None

    for t in range(n_use):
        fm = frames_meta[t]
        npz_path = ved_dir / fm["data_file"]
        if not npz_path.exists():
            raise FileNotFoundError(f"missing dyn voxel npz: {npz_path}")
        per_obj_centers: List[np.ndarray] = []
        per_obj_sizes: List[np.ndarray] = []
        per_obj_id: List[np.ndarray] = []
        per_obj_count: List[int] = []
        npz = np.load(npz_path)
        try:
            for obj_idx, name in enumerate(dyn_names):
                if name in npz.files and npz[name].size:
                    arr = npz[name].astype(np.float32)
                else:
                    arr = np.zeros((0, 3), dtype=np.float32)
                vs = float(
                    objects_info[name].get("voxel_size", default_voxel_size)
                ) * voxel_size_scale
                per_obj_centers.append(arr)
                per_obj_sizes.append(np.full(len(arr), vs, dtype=np.float32))
                per_obj_id.append(np.full(len(arr), obj_idx, dtype=np.int32))
                per_obj_count.append(len(arr))
        finally:
            npz.close()
        centers = (np.concatenate(per_obj_centers, axis=0)
                   if per_obj_centers else np.zeros((0, 3), dtype=np.float32))
        sizes = (np.concatenate(per_obj_sizes, axis=0)
                 if per_obj_sizes else np.zeros((0,), dtype=np.float32))
        obj_id = (np.concatenate(per_obj_id, axis=0)
                  if per_obj_id else np.zeros((0,), dtype=np.int32))

        if K_total_ref is None:
            K_total_ref = len(centers)
            voxel_obj_id_ref = obj_id
            per_obj_count_ref = per_obj_count
        else:
            if len(centers) != K_total_ref or per_obj_count != per_obj_count_ref:
                raise RuntimeError(
                    f"dyn voxel count mismatch at frame {t}: "
                    f"got per-object {per_obj_count} (total {len(centers)}) vs "
                    f"reference {per_obj_count_ref} (total {K_total_ref}). "
                    "Per-voxel identity is broken — fg color propagation cannot work."
                )

        centers_per_frame[t] = centers
        sizes_per_frame[t] = sizes

    assert voxel_obj_id_ref is not None
    return centers_per_frame, sizes_per_frame, voxel_obj_id_ref, dyn_names


def _voxel_dedup(chunk: VoxelChunk) -> VoxelChunk:
    if len(chunk.centers) == 0:
        return chunk
    keys = np.concatenate(
        [np.round(chunk.centers, 3), chunk.sizes.reshape(-1, 1).round(4)], axis=1
    )
    _, unique_idx = np.unique(keys, axis=0, return_index=True)
    unique_idx.sort()
    return VoxelChunk(chunk.centers[unique_idx], chunk.sizes[unique_idx])


# ---------------------------------------------------------------------------
# Foreground mask
# ---------------------------------------------------------------------------

def render_fg_mask_from_voxels(centers: np.ndarray,
                               sizes: np.ndarray,
                               c2w_opencv: np.ndarray,
                               K: np.ndarray,
                               H: int, W: int,
                               dilate_px: int = 5) -> np.ndarray:
    """Per-voxel 8-corner projected bbox fill (then dilate)."""
    mask = np.zeros((H, W), dtype=bool)
    if len(centers) == 0:
        return mask

    offs = np.array([[a, b, c] for a in (-0.5, 0.5) for b in (-0.5, 0.5)
                     for c in (-0.5, 0.5)], dtype=np.float32)
    corners_w = (centers[:, None, :]
                 + offs[None, :, :] * sizes[:, None, None])
    flat = corners_w.reshape(-1, 3)

    w2c = np.linalg.inv(c2w_opencv).astype(np.float32)
    pts_cam = flat @ w2c[:3, :3].T + w2c[:3, 3]
    z = pts_cam[:, 2]
    safe_z = np.where(z > 1e-3, z, 1.0)
    uv = pts_cam @ K.T
    u = uv[:, 0] / safe_z
    v = uv[:, 1] / safe_z

    z = z.reshape(-1, 8)
    u = u.reshape(-1, 8)
    v = v.reshape(-1, 8)
    in_front = (z > 1e-3).all(axis=1)

    for i in np.nonzero(in_front)[0]:
        umin = max(0, int(np.floor(u[i].min())))
        vmin = max(0, int(np.floor(v[i].min())))
        umax = min(W - 1, int(np.ceil(u[i].max())))
        vmax = min(H - 1, int(np.ceil(v[i].max())))
        if umin > umax or vmin > vmax:
            continue
        mask[vmin:vmax + 1, umin:umax + 1] = True

    if dilate_px > 0:
        k = max(1, int(dilate_px))
        kernel = np.ones((k, k), np.uint8)
        mask = cv2.dilate(mask.astype(np.uint8), kernel, iterations=1).astype(bool)
    return mask


# ---------------------------------------------------------------------------
# Per-case conversion
# ---------------------------------------------------------------------------

def _resolve_metadata_path(voxels_path: Path) -> Path:
    return voxels_path.parent.parent / "metadata.json"


def _resolve_ved_dir(voxels_path: Path) -> Path:
    return voxels_path.parent.parent.parent


def convert_case(entry: dict,
                 output_root: Path,
                 *,
                 target_w: int,
                 target_h: int,
                 num_frames: Optional[int],
                 viz_num_frames: Optional[int],
                 points_per_voxel: int,
                 dyn_samples_per_voxel: int,
                 voxel_size_scale: float,
                 visibility: str,
                 depth_views: int,
                 z_tol_rel: float,
                 z_tol_abs: float,
                 final_voxel_downsample: float,
                 fg_dilate: int,
                 seed: int,
                 overwrite: bool,
                 save_projection_videos: bool = True,
                 save_bg_projection_video: bool = True,
                 viz_fps: float = 16.0,
                 viz_colormap: str = "turbo") -> Optional[Path]:
    name = entry["name"]
    out_dir = output_root / name
    if out_dir.exists() and not overwrite:
        if all((out_dir / f).exists() for f in
               ("first_frame.png", "prompt.txt", "geometry.npz", "pointcloud.npz")):
            print(f"[skip] {name}: already exists (use --overwrite to redo)")
            return out_dir
    out_dir.mkdir(parents=True, exist_ok=True)

    print(f"\n=== converting {name} ===")

    voxels_path = Path(entry["voxels"])
    metadata_path = _resolve_metadata_path(voxels_path)
    ved_dir = _resolve_ved_dir(voxels_path)
    input_image = Path(entry["input_image"])
    if not metadata_path.exists():
        raise FileNotFoundError(f"metadata.json missing: {metadata_path}")
    if not input_image.exists():
        raise FileNotFoundError(f"input_image missing: {input_image}")

    metadata = json.loads(metadata_path.read_text(encoding="utf-8"))
    objects_info: Dict[str, dict] = metadata["objects_info"]
    frames_meta: List[dict] = metadata["frames"]
    default_voxel_size = float(metadata.get("voxel_size", 0.25))
    static_data_file = metadata.get("static_data_file", "static.npz")
    static_npz_path = ved_dir / static_data_file

    N_total = len(frames_meta)
    pc_n_use = N_total if num_frames is None else min(int(num_frames), N_total)
    # Default viz to the PC range so previews match the model's actual
    # generation range. Users can pass --viz-num-frames > --num-frames
    # explicitly to preview the full trajectory beyond what the model will
    # generate (e.g. for debugging / scouting).
    viz_n_use = (pc_n_use if viz_num_frames is None
                 else min(int(viz_num_frames), N_total))
    # One combined raycast over the longer range — PC gets a strict subset.
    n_use = max(pc_n_use, viz_n_use)
    frames_use = frames_meta[:n_use]
    print(f"  pc range: {pc_n_use}/{N_total} frames "
          f"(saved to geometry.npz; model will iterate to fill this)")
    print(f"  viz range: {viz_n_use}/{N_total} frames "
          f"(bg/fg projection videos)"
          + ("  [extended beyond PC range]" if viz_n_use > pc_n_use else ""))
    print(f"  raycast spans {n_use} frames")

    # ---- Cameras ----------------------------------------------------------
    c2w_blender = np.stack(
        [np.array(fm["camera_extrinsics"], dtype=np.float32) for fm in frames_use],
        axis=0,
    )
    poses_c2w = blender_c2w_to_opencv_c2w(c2w_blender)
    poses_c2w_pc = poses_c2w[:pc_n_use]
    poses_c2w_viz = poses_c2w[:viz_n_use]
    fx, fy, cx, cy = compute_intrinsics_px(frames_use[0]["camera_intrinsics"],
                                           target_w, target_h)
    K = K_matrix(fx, fy, cx, cy)
    intrinsics_size = np.array([target_h, target_w], dtype=np.int32)
    print(f"  K: fx={fx:.2f}, fy={fy:.2f}, cx={cx:.2f}, cy={cy:.2f}")

    rng = np.random.default_rng(seed)

    # ---- Static voxels ----------------------------------------------------
    static_chunk = _collect_static_voxels(
        static_npz_path, objects_info, default_voxel_size, voxel_size_scale,
    )
    static_chunk = _voxel_dedup(static_chunk)
    print(f"  static voxels: {len(static_chunk.centers)}")

    # ---- Dynamic voxels per frame (persistent per-voxel identity) --------
    # We DO NOT dedup or shuffle so that row `i` in frame `t` is the same
    # physical voxel as row `i` in frame `0` (required for the fg viz to
    # propagate frame-0 colors forward through time by voxel index).
    (dyn_centers_per_frame, dyn_sizes_per_frame,
     dyn_voxel_obj_id, dyn_obj_names) = _collect_dyn_per_object_persistent(
        ved_dir, frames_meta, objects_info,
        default_voxel_size, voxel_size_scale, n_use,
    )
    K_dyn_total = dyn_centers_per_frame[0].shape[0] if n_use else 0
    total_dyn_voxels = K_dyn_total * n_use
    print(f"  dyn voxels: {K_dyn_total} voxels x {n_use} frames "
          f"(objects={dyn_obj_names}; persistent index)")

    # ---- Visibility ------------------------------------------------------
    static_pc_for_viz: np.ndarray = np.zeros((0, 3), dtype=np.float32)
    dyn_pc_per_frame_rgb: Dict[int, np.ndarray] = {}
    dyn_colors_per_frame_rgb: Dict[int, np.ndarray] = {}

    if visibility == "depth":
        # ============================================================
        # NEW FAST PATH (no per-frame raycast):
        # ------------------------------------------------------------
        # 1) Static: raycast ONLY at camera 0 (with dyn[0] as occluder).
        #    Anything visible only from later frames cannot be coloured
        #    by first_frame, so it would be dropped — pointless to cast.
        # 2) Dyn: skip raycast entirely. For every dyn surface sample at
        #    frame 0, project to camera 0 and sample first_frame RGB at
        #    that pixel (must fall inside fg_mask). Transform the SHARED
        #    unit-cube surface template by (size_t, center_t) per frame;
        #    colours stay fixed per sample index. Z-buffer at splat time
        #    handles silhouette / inter-voxel occlusion.
        # ============================================================
        # SPMEM-style multi-view static back-projection: sample views uniformly
        # across the model's generation range and union the first-hit surface
        # points, so the background covers everything the camera eventually
        # sees (progressive coloring fills in colours per chunk).
        total_traj = max(1, pc_n_use)
        n_views = min(max(1, int(depth_views)), total_traj)
        if n_views <= 1:
            view_indices = [0]
        else:
            view_indices = np.unique(
                np.linspace(0, total_traj - 1, n_views).round().astype(int)
            ).tolist()
        print(f"  [depth] static: raycasting {len(view_indices)} views uniformly "
              f"across {total_traj} frames (target {target_h}x{target_w}, "
              f"static-only)...")
        static_pts_raw = raycast_static_multiview(
            static_chunk.centers, static_chunk.sizes,
            poses_c2w[view_indices], K, target_h, target_w,
        )
        print(f"  [depth] static raw hits: {len(static_pts_raw)}")

        # Load first_frame in memory now so we can build the per-voxel
        # color table without round-tripping via disk. (We'll still save
        # both first_frame.png and fg_mask_first.png below.)
        _img_pil = Image.open(input_image).convert("RGB")
        if _img_pil.size != (target_w, target_h):
            _img_pil = _img_pil.resize((target_w, target_h), Image.LANCZOS)
        first_rgb_mem = np.asarray(_img_pil, dtype=np.uint8)
        if K_dyn_total > 0:
            fg_mask_mem = render_fg_mask_from_voxels(
                centers=dyn_centers_per_frame[0],
                sizes=dyn_sizes_per_frame[0],
                c2w_opencv=poses_c2w[0], K=K,
                H=target_h, W=target_w, dilate_px=fg_dilate,
            )
        else:
            fg_mask_mem = None

        K_per_voxel = int(dyn_samples_per_voxel)
        unit_samples = build_unit_cube_surface_samples(K_per_voxel, seed=seed)
        samples_0 = transform_unit_samples_per_voxel(
            unit_samples,
            dyn_centers_per_frame[0],
            dyn_sizes_per_frame[0],
        )
        print(f"  [color] per-point frame-0 colour sampling "
              f"({K_dyn_total} voxels × {K_per_voxel} samples = "
              f"{len(samples_0)} pts)...")
        point_colors, point_has_color = color_points_from_first_frame(
            samples_0, poses_c2w[0], K, first_rgb_mem,
            fg_mask=fg_mask_mem,
        )
        n_colored = int(point_has_color.sum())
        print(f"  [color] -> {n_colored}/{len(samples_0)} surface pts coloured "
              f"({100.0 * point_has_color.mean():.1f}%)")

        print(f"  [dyn] skipping raycast; using {n_colored} coloured surface pts "
              f"× {n_use} frames")

        if n_colored > 0:
            # Colours are fixed from each point's camera-0 projection onto
            # first_frame; only positions are rigidly propagated per frame.
            shared_colors = point_colors[point_has_color].astype(np.uint8)
            for t in range(n_use):
                samples_t = transform_unit_samples_per_voxel(
                    unit_samples,
                    dyn_centers_per_frame[t],
                    dyn_sizes_per_frame[t],
                )
                dyn_pc_per_frame_rgb[t] = samples_t[point_has_color].astype(np.float32)
                dyn_colors_per_frame_rgb[t] = shared_colors
        total_dyn_samples = sum(len(p) for p in dyn_pc_per_frame_rgb.values())

        if final_voxel_downsample <= 0:
            # 1cm matches LiveWorld's default scene_voxel_size.
            final_voxel_downsample = 0.01
            print(f"  [depth] auto-setting --final-voxel-downsample to "
                  f"{final_voxel_downsample} m")

        static_pc_for_viz = voxel_downsample_points(
            static_pts_raw, final_voxel_downsample
        )

        # Background point cloud = STATIC voxels only (auto fg/bg split, matching
        # SPMEM's static-only cache). Dynamic foreground is NOT baked into
        # pointcloud.npz — it is conditioned separately via fg_projection.mp4 /
        # fg_mask_first.png and re-rendered per chunk at inference time.
        points = static_pc_for_viz
        print(f"  [depth] downsample @ {final_voxel_downsample} m: "
              f"static-only PC {len(points)} pts "
              f"(views={len(view_indices)}; dyn samples kept only for fg viz: "
              f"{total_dyn_samples})")
    else:
        # Surface-sampling modes (legacy / debug only).
        static_pts, static_normals = sample_cube_surface_points(
            static_chunk.centers, static_chunk.sizes, points_per_voxel, rng,
            return_normals=True,
        )
        dyn_pts_per_frame: Dict[int, np.ndarray] = {}
        dyn_normals_per_frame: Dict[int, np.ndarray] = {}
        for t in range(n_use):
            c = dyn_centers_per_frame[t]
            s = dyn_sizes_per_frame[t]
            if len(c) == 0:
                dyn_pts_per_frame[t] = np.zeros((0, 3), np.float32)
                dyn_normals_per_frame[t] = np.zeros((0, 3), np.float32)
            else:
                p, n = sample_cube_surface_points(c, s, points_per_voxel, rng,
                                                  return_normals=True)
                dyn_pts_per_frame[t] = p
                dyn_normals_per_frame[t] = n

        if visibility == "none":
            static_keep = np.ones(len(static_pts), dtype=bool)
            dyn_keeps = {t: np.ones(len(p), dtype=bool)
                         for t, p in dyn_pts_per_frame.items()}
        elif visibility == "backface":
            cam_positions = poses_c2w[:, :3, 3]
            static_keep = cull_back_faces(static_pts, static_normals, cam_positions)
            dyn_keeps = {}
            for t, p in dyn_pts_per_frame.items():
                if len(p) == 0:
                    dyn_keeps[t] = np.zeros(0, dtype=bool)
                else:
                    dyn_keeps[t] = cull_back_faces(
                        p, dyn_normals_per_frame[t], poses_c2w[t:t + 1, :3, 3],
                    )
        else:
            raise ValueError(f"--visibility must be 'none'|'backface'|'depth', "
                             f"got {visibility}")

        final_static = static_pts[static_keep]
        final_dyn_chunks = [dyn_pts_per_frame[t][dyn_keeps[t]]
                            for t in sorted(dyn_pts_per_frame)]
        final_dyn_chunks = [c for c in final_dyn_chunks if len(c) > 0]
        final_dyn = (np.concatenate(final_dyn_chunks, axis=0)
                     if final_dyn_chunks else np.zeros((0, 3), dtype=np.float32))
        points = np.concatenate([final_static, final_dyn], axis=0)
        total_dyn_pts = sum(len(p) for p in dyn_pts_per_frame.values())
        print(f"  [{visibility}] kept static: {len(final_static)}/{len(static_pts)},  "
              f"dyn: {len(final_dyn)}/{total_dyn_pts}")
        if final_voxel_downsample > 0:
            before = len(points)
            points = voxel_downsample_points(points, final_voxel_downsample)
            print(f"  downsample @ {final_voxel_downsample} m: {before} → {len(points)}")

    print(f"  final PC: {points.shape[0]} pts")

    # ---- First-frame image -----------------------------------------------
    img = Image.open(input_image).convert("RGB")
    if img.size != (target_w, target_h):
        img = img.resize((target_w, target_h), Image.LANCZOS)
    img.save(out_dir / "first_frame.png")

    # ---- Foreground mask at first frame (frame_0001 dyn voxels) ----------
    dyn_first = _collect_dyn_voxels_for_frame(
        ved_dir, frames_meta[0], objects_info, default_voxel_size, voxel_size_scale,
    )
    if len(dyn_first.centers) > 0:
        fg_mask = render_fg_mask_from_voxels(
            centers=dyn_first.centers, sizes=dyn_first.sizes,
            c2w_opencv=poses_c2w[0], K=K,
            H=target_h, W=target_w, dilate_px=fg_dilate,
        )
        cv2.imwrite(str(out_dir / "fg_mask_first.png"),
                    (fg_mask.astype(np.uint8) * 255))
        print(f"  fg_mask coverage: {fg_mask.mean():.3%}")
    else:
        print("  fg_mask: no dynamic voxels in frame 1 → skipped")

    # ---- Prompt -----------------------------------------------------------
    (out_dir / "prompt.txt").write_text(entry["prompt"].strip() + "\n", encoding="utf-8")

    # ---- Save -------------------------------------------------------------
    np.savez_compressed(
        out_dir / "geometry.npz",
        poses_c2w=poses_c2w_pc.astype(np.float32),
        K=K.astype(np.float32),
        intrinsics_size=intrinsics_size,
    )
    np.savez_compressed(out_dir / "pointcloud.npz", points=points.astype(np.float32))

    # ---- Projection videos (bg / fg) -------------------------------------
    #
    # spmem_comp + LiveWorld both do **PC z-buffer splat** (each 3-D point →
    # 1 pixel, keep the closest one) with RGB colors sampled from the first
    # frame — NOT raw depth. So we mirror that here:
    #   1. For each point, project to camera 0 and sample first_frame's RGB.
    #      Out-of-view points get gray (128,128,128).
    #   2. For bg PC, mask out pixels under fg_mask_first.png so character
    #      colors don't bleed onto static points hidden behind the character.
    #   3. Z-buffer splat the colored PC through every camera.
    if save_projection_videos and visibility == "depth":
        # ---- bg: drop uncolored points ----------------------------------
        # Points that (a) project outside cam 0, or (b) project onto a pixel
        # inside fg_mask, get dropped entirely instead of rendered gray.
        # The result: a clean, honest PC where every rendered pixel comes
        # from a real first_frame sample.
        #
        # NOTE: bg_projection.mp4 is ONLY used when running inference with
        # condition_source="mp4". The default workflow renders the scene from
        # the point cloud at run time (splat2d), so this slow 401-camera splat
        # over the full static cloud is pure waste there — skip it with
        # --no-bg-projection-video.
        if save_bg_projection_video:
            first_rgb = np.asarray(
                Image.open(out_dir / "first_frame.png").convert("RGB"),
                dtype=np.uint8,
            )
            fg_mask_path = out_dir / "fg_mask_first.png"
            fg_mask_bool: Optional[np.ndarray] = None
            if fg_mask_path.exists():
                m = cv2.imread(str(fg_mask_path), cv2.IMREAD_GRAYSCALE)
                if m is not None:
                    fg_mask_bool = m > 0

            print(f"  [viz] sampling bg colors from first_frame "
                  f"({len(static_pc_for_viz)} pts; fg region masked, "
                  f"uncolored points dropped)...")
            bg_colors_full, bg_keep = sample_pc_colors_from_image(
                static_pc_for_viz, poses_c2w[0], K,
                first_rgb, fg_mask=fg_mask_bool, return_keep_mask=True,
            )
            bg_pts_colored = static_pc_for_viz[bg_keep]
            bg_cols_colored = bg_colors_full[bg_keep]
            print(f"  [viz] kept {len(bg_pts_colored)}/{len(static_pc_for_viz)} "
                  f"static pts that successfully sampled a color "
                  f"({100.0 * bg_keep.mean():.1f}%)")
            print(f"  [viz] z-buffer splat bg PC through {viz_n_use} cameras "
                  f"(splat 2x2)...")
            bg_frames = render_pc_video_rgb(
                bg_pts_colored, bg_cols_colored,
                poses_c2w_viz, K, target_h, target_w, is_static=True,
                splat_size=2,
            )
            bg_path = out_dir / "bg_projection.mp4"
            save_video_mp4(bg_path, bg_frames, fps=viz_fps)
            nz_bg = (bg_frames.reshape(bg_frames.shape[0], -1, 3).sum(-1) > 0).mean()
            print(f"  [viz] -> {bg_path}  "
                  f"({bg_frames.shape[0]} frames @ {viz_fps} fps, "
                  f"non-empty pixels avg {100.0 * nz_bg:.1f}%)")
        else:
            print("  [viz] bg_projection.mp4 skipped "
                  "(scene uses runtime point-cloud splat2d)")

        # ---- fg: time-propagated per-point colors -----------------------
        # `dyn_pc_per_frame_rgb` and `dyn_colors_per_frame_rgb` were
        # already built in the depth branch above: each surface sample
        # gets its RGB from where it projects on first_frame at camera 0;
        # per-frame positions come from the shared unit-cube template
        # rigidly transformed by (size_t, center_t).
        n_dyn_samples_total = sum(len(p) for p in dyn_pc_per_frame_rgb.values())
        # Restrict to the viz range (raycast may have spanned more if
        # pc_n_use > viz_n_use, though usually viz_n_use is the larger one).
        dyn_viz_pts = {t: dyn_pc_per_frame_rgb.get(
                          t, np.zeros((0, 3), dtype=np.float32))
                       for t in range(viz_n_use)}
        dyn_viz_cols = {t: dyn_colors_per_frame_rgb.get(
                           t, np.zeros((0, 3), dtype=np.uint8))
                        for t in range(viz_n_use)}
        print(f"  [viz] fg: {n_dyn_samples_total} sampled pts across "
              f"{len(dyn_pc_per_frame_rgb)} frames (all already coloured)")
        print(f"  [viz] z-buffer splat fg PC through {viz_n_use} cameras "
              f"(splat 2x2)...")
        fg_frames = render_pc_video_rgb(
            dyn_viz_pts, dyn_viz_cols,
            poses_c2w_viz, K, target_h, target_w, is_static=False,
            splat_size=2,
        )
        fg_path = out_dir / "fg_projection.mp4"
        save_video_mp4(fg_path, fg_frames, fps=viz_fps)
        nz_fg = (fg_frames.reshape(fg_frames.shape[0], -1, 3).sum(-1) > 0).mean()
        print(f"  [viz] -> {fg_path}  "
              f"({fg_frames.shape[0]} frames @ {viz_fps} fps, "
              f"non-empty pixels avg {100.0 * nz_fg:.1f}%)")
    elif save_projection_videos:
        print("  [viz] skipped (projection videos only supported for "
              "--visibility depth)")

    print(f"  -> {out_dir}")
    return out_dir


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------

def _parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(description="Convert mine_blender voxel exports to LiveWorld_comp inputs (depth-derived PC)")
    p.add_argument("--meta", required=True, help="Path to meta_long_updated.json")
    p.add_argument("--output-root", required=True)
    p.add_argument("--names", nargs="*", default=None,
                   help="Only convert these entry names (default: all)")
    p.add_argument("--target-w", type=int, default=TARGET_W_DEFAULT)
    p.add_argument("--target-h", type=int, default=TARGET_H_DEFAULT)
    p.add_argument("--num-frames", type=int, default=None,
                   help="Number of frames the model will generate "
                        "(default: use all frames in metadata). Saved as "
                        "geometry.npz's poses_c2w and used to scope the dyn "
                        "voxels included in pointcloud.npz.")
    p.add_argument("--viz-num-frames", type=int, default=None,
                   help="Number of frames in the bg/fg projection videos. "
                        "Defaults to --num-frames so previews stay aligned "
                        "with the model's actual generation range. Set to a "
                        "larger value (e.g. metadata full length) only when "
                        "you want to inspect the full trajectory beyond what "
                        "the model will generate — this preview will then NOT "
                        "correspond 1:1 to model output. Raycast scope = "
                        "max(--num-frames, --viz-num-frames).")
    p.add_argument("--points-per-voxel", type=int, default=48,
                   help="(legacy/debug modes only) Surface samples per "
                        "voxel for --visibility none|backface.")
    p.add_argument("--dyn-samples-per-voxel", type=int, default=96,
                   help="(depth mode) Number of surface samples per coloured "
                        "dyn voxel used to build the per-frame dyn PC without "
                        "raycasting. Default 96 ≈ 16/face — denser = smoother "
                        "fg viz / model input but slower splat. Drop to 48 if "
                        "you want closer parity with the legacy raycast output.")
    p.add_argument("--voxel-size-scale", type=float, default=1.0)
    p.add_argument("--depth-views", type=int, default=10,
                   help="(depth mode) Number of camera views, sampled uniformly "
                        "across the model's generation range, used to raycast "
                        "the STATIC voxels into a depth-derived background point "
                        "cloud (matches SPMEM's --depth_views). 1 = frame-0 only. "
                        "Default 20.")
    p.add_argument("--visibility", choices=["none", "backface", "depth"], default="depth",
                   help="none: keep all faces (legacy); "
                        "backface: cull faces facing away from every cam (fast); "
                        "depth: per-frame z-buffer culling (depth-sensor-like, default)")
    p.add_argument("--z-tol-rel", type=float, default=0.005,
                   help="Relative z-buffer tolerance (frac of ref_z)")
    p.add_argument("--z-tol-abs", type=float, default=0.01,
                   help="Absolute z-buffer tolerance in metres")
    p.add_argument("--final-voxel-downsample", type=float, default=0.0,
                   help="Optional final voxel downsample in metres (0 = off)")
    p.add_argument("--fg-dilate", type=int, default=5)
    p.add_argument("--seed", type=int, default=0)
    p.add_argument("--overwrite", action="store_true")
    p.add_argument("--limit", type=int, default=None)
    p.add_argument("--indices", type=int, nargs="*", default=None,
                   help="Pick specific meta entries by 0-based index (e.g. "
                        "--indices 0 3 5). Applied after --names filtering, "
                        "before --limit.")
    p.add_argument("--no-projection-videos", action="store_true",
                   help="Skip writing bg_projection.mp4 / fg_projection.mp4")
    p.add_argument("--no-bg-projection-video", action="store_true",
                   help="Skip only bg_projection.mp4 (the slow full-cloud "
                        "splat). Safe for the default workflow, where the scene "
                        "is rendered from the point cloud at inference time; "
                        "bg_projection.mp4 is only needed for "
                        "condition_source='mp4'. fg_projection.mp4 is still written.")
    p.add_argument("--viz-fps", type=float, default=16.0)
    p.add_argument("--viz-colormap", default="turbo",
                   choices=["turbo", "jet", "magma", "viridis", "gray", "grey"])
    return p.parse_args()


def main() -> None:
    args = _parse_args()
    meta_path = Path(args.meta)
    if not meta_path.exists():
        raise SystemExit(f"meta not found: {meta_path}")

    meta = json.loads(meta_path.read_text(encoding="utf-8"))
    if args.names:
        wanted = set(args.names)
        entries = [e for e in meta if e.get("name") in wanted]
        missing = wanted - {e.get("name") for e in entries}
        if missing:
            print(f"[warn] entries not found: {sorted(missing)}", file=sys.stderr)
    else:
        entries = list(meta)

    if args.indices is not None:
        bad = [i for i in args.indices if i < 0 or i >= len(entries)]
        if bad:
            raise SystemExit(
                f"--indices out of range {bad} (have {len(entries)} entries)"
            )
        entries = [entries[i] for i in args.indices]

    if args.limit is not None:
        entries = entries[: args.limit]

    out_root = Path(args.output_root)
    out_root.mkdir(parents=True, exist_ok=True)

    print(f"[boot] {len(entries)} cases → {out_root}  (visibility={args.visibility})")

    n_ok, n_fail = 0, 0
    for entry in entries:
        try:
            convert_case(
                entry, out_root,
                target_w=args.target_w, target_h=args.target_h,
                num_frames=args.num_frames,
                viz_num_frames=args.viz_num_frames,
                points_per_voxel=args.points_per_voxel,
                dyn_samples_per_voxel=args.dyn_samples_per_voxel,
                voxel_size_scale=args.voxel_size_scale,
                visibility=args.visibility,
                depth_views=args.depth_views,
                z_tol_rel=args.z_tol_rel,
                z_tol_abs=args.z_tol_abs,
                final_voxel_downsample=args.final_voxel_downsample,
                fg_dilate=args.fg_dilate,
                seed=args.seed,
                overwrite=args.overwrite,
                save_projection_videos=not args.no_projection_videos,
                save_bg_projection_video=not args.no_bg_projection_video,
                viz_fps=args.viz_fps,
                viz_colormap=args.viz_colormap,
            )
            n_ok += 1
        except Exception as e:
            n_fail += 1
            print(f"[FAIL] {entry.get('name')}: {e}", file=sys.stderr)

    print(f"\n[done] ok={n_ok}, fail={n_fail}, total={len(entries)}")


if __name__ == "__main__":
    main()