Spaces:
Running on Zero
Running on Zero
File size: 15,461 Bytes
6febba8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 | """Level-of-Token layouts.
A layout is a partition of the uniform token lattice into axis-aligned
rectangles. Extents are measured in uniform tokens, matching Kiyohiro Nakayama, Brian Chao, Jan Ackermann, Hansheng Chen, Federico Tombari, Leonidas Guibas, Lior Yariv and Gordon Wetzstein.
2026, Sec. 3.1. Video layouts in the paper keep a temporal extent of 1 and
vary only the spatial rectangle (Appendix A.10).
"""
from __future__ import annotations
import math
from dataclasses import dataclass
import torch
@dataclass(frozen=True)
class TokenRect:
"""One LoT token covering ``[t, t+et) x [u, u+eh) x [v, v+ew)``."""
t: int
u: int
v: int
et: int
eh: int
ew: int
@property
def extent(self) -> tuple[int, int, int]:
return (self.et, self.eh, self.ew)
def center(self) -> tuple[float, float, float]:
"""Finest-grid center. Unit extents recover the integer lattice (eq. 13)."""
return (
self.t + (self.et - 1) / 2,
self.u + (self.eh - 1) / 2,
self.v + (self.ew - 1) / 2,
)
@dataclass(frozen=True)
class LotLayout:
time: int
height: int
width: int
rects: tuple[TokenRect, ...]
def __post_init__(self) -> None:
validate_layout(self)
@property
def count(self) -> int:
return len(self.rects)
@property
def dense_count(self) -> int:
return self.time * self.height * self.width
@property
def compression(self) -> float:
"""Uniform-token count divided by the LoT sequence length."""
return self.dense_count / self.count
def centers(self) -> torch.Tensor:
"""``(L, 3)`` centers in ``(t, h, w)`` order on the uniform lattice."""
return torch.tensor([rect.center() for rect in self.rects], dtype=torch.float32)
def groups(self) -> list[tuple[tuple[int, int, int], list[int]]]:
"""Rect indices grouped by extent, in order of first appearance."""
order: list[tuple[int, int, int]] = []
buckets: dict[tuple[int, int, int], list[int]] = {}
for index, rect in enumerate(self.rects):
bucket = buckets.get(rect.extent)
if bucket is None:
order.append(rect.extent)
bucket = []
buckets[rect.extent] = bucket
bucket.append(index)
return [(extent, buckets[extent]) for extent in order]
def validate_layout(layout: LotLayout) -> None:
if layout.time < 1 or layout.height < 1 or layout.width < 1:
raise ValueError("layout axes must be positive")
if not layout.rects:
raise ValueError("layout has no tokens")
cover = torch.zeros(layout.time, layout.height, layout.width, dtype=torch.int32)
for rect in layout.rects:
if min(rect.et, rect.eh, rect.ew) < 1:
raise ValueError(f"non-positive extent on {rect}")
t1 = rect.t + rect.et
u1 = rect.u + rect.eh
v1 = rect.v + rect.ew
if rect.t < 0 or rect.u < 0 or rect.v < 0 or t1 > layout.time or u1 > layout.height or v1 > layout.width:
raise ValueError(f"rect {rect} leaves the {layout.time}x{layout.height}x{layout.width} lattice")
cover[rect.t:t1, rect.u:u1, rect.v:v1] += 1
if int(cover.min()) != 1 or int(cover.max()) != 1:
raise ValueError("rects are not a partition of the uniform token lattice")
def dense_layout(time: int, height: int, width: int) -> LotLayout:
"""One 1x1x1 token per uniform site, in t-then-row-major order."""
rects = [
TokenRect(t, u, v, 1, 1, 1)
for t in range(time)
for u in range(height)
for v in range(width)
]
return LotLayout(time, height, width, tuple(rects))
def layout_from_rects(time: int, height: int, width: int, rects: list[TokenRect]) -> LotLayout:
return LotLayout(time, height, width, tuple(rects))
def _walk_squares(time: int, height: int, width: int, root: int, keep, edge_rects: bool = False) -> LotLayout:
"""Quadtree over any ``height x width`` grid; ``keep(t, u, v, eh, ew)`` decides a merge.
Blocks are laid on the ``root`` lattice from the top-left. A block that
crosses the far edge is clipped: it splits until its pieces fit (squares
only), or, with ``edge_rects``, is kept as the clipped rectangle when both
sides are powers of two (e.g. 4x2 on the 2-token strip of H3's 24x42).
"""
if root < 1 or (root & (root - 1)) != 0:
raise ValueError(f"root extent must be a power of two, got {root}")
rects: list[TokenRect] = []
def visit(t: int, u: int, v: int, size: int) -> None:
if u >= height or v >= width:
return
eh, ew = min(size, height - u), min(size, width - v)
if size == 1:
rects.append(TokenRect(t, u, v, 1, 1, 1))
return
if eh == size and ew == size:
if keep(t, u, v, size, size):
rects.append(TokenRect(t, u, v, 1, size, size))
return
elif edge_rects and max(eh, ew) > 1 and _power_of_two(eh) and _power_of_two(ew) and keep(t, u, v, eh, ew):
rects.append(TokenRect(t, u, v, 1, eh, ew))
return
half = size // 2
for du in (0, half):
for dv in (0, half):
visit(t, u + du, v + dv, half)
for t in range(time):
for u in range(0, height, root):
for v in range(0, width, root):
visit(t, u, v, root)
return LotLayout(time, height, width, tuple(rects))
def _power_of_two(n: int) -> bool:
return n >= 1 and (n & (n - 1)) == 0
def _as_time_field(field: torch.Tensor) -> torch.Tensor:
if field.ndim == 2:
return field.unsqueeze(0)
if field.ndim == 3:
return field
raise ValueError(f"expected (H, W) or (T, H, W), got {tuple(field.shape)}")
def layout_from_detail(
detail: torch.Tensor,
thresholds: dict[int, float],
root: int = 8,
edge_rects: bool = False,
) -> LotLayout:
"""Quadtree of eq. 23.
A block of extent ``b`` splits when its maximum detail is at least
``thresholds[b]``. Thresholds must be nondecreasing as blocks get finer:
``t_root <= ... <= t_2``. ``detail`` is already on the uniform token lattice.
"""
field = _as_time_field(detail).detach().float()
sizes = [size for size in (8, 4, 2) if size <= root]
missing = [size for size in sizes if size not in thresholds]
if missing:
raise ValueError(f"thresholds missing extents {missing}")
ordered = [float(thresholds[size]) for size in sizes]
if ordered != sorted(ordered):
raise ValueError("thresholds must satisfy t8 <= t4 <= t2")
def keep(t: int, u: int, v: int, eh: int, ew: int) -> bool:
block = field[t, u:u + eh, v:v + ew]
return float(block.max()) < float(thresholds[max(eh, ew)])
return _walk_squares(field.shape[0], field.shape[1], field.shape[2], root, keep, edge_rects)
def layout_from_desired(desired: torch.Tensor, root: int = 8, edge_rects: bool = False) -> LotLayout:
"""Keep a square only when every cell inside allows that extent.
``desired`` stores the coarsest square extent each finest cell may belong
to. This is the merge rule used for mask levels and for blur radii: a
block is coarsened only when its most demanding cell still allows it.
"""
field = _as_time_field(desired).detach().float()
def keep(t: int, u: int, v: int, eh: int, ew: int) -> bool:
block = field[t, u:u + eh, v:v + ew]
return float(block.min()) >= float(max(eh, ew))
return _walk_squares(field.shape[0], field.shape[1], field.shape[2], root, keep, edge_rects)
def levels_to_desired(level: torch.Tensor, max_level: int = 3) -> torch.Tensor:
"""Map paper levels ``{0,1,2,3}`` to extents ``{8,4,2,1}`` (eq. 19)."""
if int(level.min()) < 0 or int(level.max()) > max_level:
raise ValueError(f"levels must lie in 0..{max_level}")
return 2 ** (max_level - level.to(torch.int64))
def layout_from_regions(
height: int,
width: int,
background_level: int,
regions: list[tuple[torch.Tensor, int]],
root: int = 8,
time: int = 1,
edge_rects: bool = False,
) -> LotLayout:
"""Eq. 19. Overlapping regions keep the finer (larger) level.
Each mask is ``(H, W)`` or ``(T, H, W)`` with true on the selected cells.
The same masks are applied on every frame when ``time > 1`` and the mask
has no time axis. Per-frame masks must already have length ``time``.
"""
level = torch.full((time, height, width), int(background_level), dtype=torch.int64)
for mask, region_level in regions:
if region_level <= background_level:
raise ValueError("selected regions must be finer than the background")
mask_field = mask.to(dtype=torch.bool)
if mask_field.ndim == 2:
mask_field = mask_field.unsqueeze(0).expand(time, -1, -1)
if tuple(mask_field.shape) != (time, height, width):
raise ValueError(f"mask shape {tuple(mask_field.shape)} != {(time, height, width)}")
picked = torch.full_like(level, int(region_level))
level = torch.where(mask_field, torch.maximum(level, picked), level)
return layout_from_desired(levels_to_desired(level), root=root, edge_rects=edge_rects)
def shape_features(extent: tuple[int, int, int], include_time: bool = False) -> torch.Tensor:
"""Eq. 11, optionally followed by ``log2(et)`` for temporal extents."""
et, eh, ew = extent
if min(et, eh, ew) < 1:
raise ValueError(f"bad extent {extent}")
values = [
math.log2(eh),
math.log2(ew),
math.log2(eh * ew),
math.log2(eh / ew),
]
if include_time:
values.append(math.log2(et))
return torch.tensor(values, dtype=torch.float32)
def smooth_detail(detail: torch.Tensor, sigma: float) -> torch.Tensor:
"""Separable Gaussian on the last two axes. ``sigma <= 0`` is a no-op."""
if sigma <= 0:
return detail
radius = max(1, int(math.ceil(3 * sigma)))
coords = torch.arange(-radius, radius + 1, dtype=torch.float32, device=detail.device)
kernel = torch.exp(-0.5 * (coords / float(sigma)) ** 2)
kernel = kernel / kernel.sum()
return _conv_last(detail.float(), kernel, radius).to(dtype=detail.dtype)
def _conv_last(field: torch.Tensor, kernel: torch.Tensor, radius: int) -> torch.Tensor:
"""Gaussian along height, then along width."""
flat = field.reshape(-1, 1, field.shape[-2], field.shape[-1])
weight_w = kernel.view(1, 1, 1, -1)
weight_h = kernel.view(1, 1, -1, 1)
flat = torch.nn.functional.pad(flat, (radius, radius, 0, 0), mode="reflect")
flat = torch.nn.functional.conv2d(flat, weight_w)
flat = torch.nn.functional.pad(flat, (0, 0, radius, radius), mode="reflect")
flat = torch.nn.functional.conv2d(flat, weight_h)
return flat.reshape(field.shape).to(dtype=field.dtype)
def prepare_detail(scores: torch.Tensor, gain: float = 1.0, sigma: float = 0.0) -> torch.Tensor:
"""Gain, optional smooth, clip to ``[0, 1]``. Scores are on the token lattice."""
return smooth_detail(scores.float() * float(gain), sigma).clamp(0, 1)
def area_average(pixel_scores: torch.Tensor, token_height: int, token_width: int) -> torch.Tensor:
"""Area-average a pixel detail map onto the uniform token lattice."""
if pixel_scores.ndim == 2:
field = pixel_scores[None, None]
squeeze = 2
elif pixel_scores.ndim == 3:
field = pixel_scores[:, None]
squeeze = 1
else:
raise ValueError("pixel scores must be (H, W) or (T, H, W)")
pooled = torch.nn.functional.adaptive_avg_pool2d(field.float(), (token_height, token_width))
if squeeze == 2:
return pooled[0, 0]
return pooled[:, 0]
def blur_radius(depth: torch.Tensor, focal_depth: float, r_target: float) -> torch.Tensor:
"""Circle-of-confusion radius, eqs. 21–22.
``K`` is set so the 95th percentile of absolute inverse-depth deviation
equals ``r_target``. Depth must be positive.
"""
if focal_depth <= 0 or r_target < 0:
raise ValueError("focal depth and r_target must be positive")
deviation = (depth.float().clamp(min=1e-8).reciprocal() - (1.0 / float(focal_depth))).abs()
flat = deviation.reshape(-1)
percentile = torch.quantile(flat, 0.95).clamp(min=1e-8)
return deviation * (float(r_target) / percentile)
def layout_from_blur(radius: torch.Tensor, support: dict[int, float], root: int = 8,
edge_rects: bool = False) -> LotLayout:
"""Merge a block only when its smallest blur radius supports that extent.
``support[b]`` is the minimum radius that still permits an extent-``b``
token. ``support`` values must increase with ``b``. Cells with a small
radius (in focus) force the block to split.
"""
sizes = [size for size in (8, 4, 2) if size <= root]
missing = [size for size in sizes if size not in support]
if missing:
raise ValueError(f"blur support missing extents {missing}")
ordered = [float(support[size]) for size in sizes]
if ordered != sorted(ordered, reverse=True):
raise ValueError("blur support must be larger for coarser extents")
field = _as_time_field(radius).detach().float()
def keep(t: int, u: int, v: int, eh: int, ew: int) -> bool:
block = field[t, u:u + eh, v:v + ew]
return float(block.min()) >= float(support[max(eh, ew)])
return _walk_squares(field.shape[0], field.shape[1], field.shape[2], root, keep, edge_rects)
def directional_variation(luminance: torch.Tensor) -> torch.Tensor:
"""Per-cell max of absolute forward differences, padded at the far edge.
This is the error field fed to the VRS tolerance in Appendix A.3. The
paper cites Yang et al. 2019 for the shading rule and does not publish
the finite-difference stencil; the forward difference is the stencil
used here.
"""
field = _as_time_field(luminance).float()
dx = (field[..., :, 1:] - field[..., :, :-1]).abs()
dy = (field[..., 1:, :] - field[..., :-1, :]).abs()
dx = torch.nn.functional.pad(dx, (0, 1))
dy = torch.nn.functional.pad(dy, (0, 0, 0, 1))
return torch.maximum(dx, dy).reshape(luminance.shape).to(dtype=luminance.dtype)
def layout_from_vrs(
luminance: torch.Tensor,
sensitivity: float,
ambient: float,
root: int = 8,
edge_rects: bool = False,
) -> LotLayout:
"""Coarsen a block when its variation stays under ``s * (mean(I) + a)``.
Training in the paper draws ``s`` from ``[0.01, 0.1]`` and ``a`` from
``[0.01, 1]``. Larger values permit coarser tokens. The decision is
square, on the same quadtree as the detail maps. Directional rectangles
can still be passed through ``layout_from_rects``.
"""
if sensitivity < 0 or ambient < 0:
raise ValueError("VRS sensitivity and ambient must be non-negative")
luma = _as_time_field(luminance).float()
error = _as_time_field(directional_variation(luminance)).float()
def keep(t: int, u: int, v: int, eh: int, ew: int) -> bool:
block_luma = luma[t, u:u + eh, v:v + ew]
block_error = error[t, u:u + eh, v:v + ew]
tau = float(sensitivity) * (float(block_luma.mean()) + float(ambient))
return float(block_error.max()) < tau
return _walk_squares(luma.shape[0], luma.shape[1], luma.shape[2], root, keep, edge_rects)
|