File size: 16,597 Bytes
9463e5c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 | """
GLADIUS β Gaussian Head Data
Procedural Gaussian scene generation for training the specialist head.
No external datasets needed β generates ground truth Gaussians from
geometric primitives (spheres, cubes, multi-object scenes).
Also includes a differentiable 2D renderer for computing training loss.
The renderer projects 3D Gaussians to 2D and alpha-composites them.
"""
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset
import math
import random
def fibonacci_sphere(n: int, device: str = 'cpu') -> torch.Tensor:
"""
Generate n points uniformly distributed on a unit sphere.
Fibonacci/golden spiral method.
Returns: (n, 3) β unit vectors on sphere surface
"""
golden_ratio = (1 + math.sqrt(5)) / 2
indices = torch.arange(n, dtype=torch.float32, device=device)
theta = 2 * math.pi * indices / golden_ratio
phi = torch.acos(1 - 2 * (indices + 0.5) / n)
x = torch.sin(phi) * torch.cos(theta)
y = torch.sin(phi) * torch.sin(theta)
z = torch.cos(phi)
return torch.stack([x, y, z], dim=-1)
def random_quaternion(n: int, device: str = 'cpu') -> torch.Tensor:
"""Generate n random unit quaternions (uniform on SO(3))."""
u = torch.rand(n, 3, device=device)
q = torch.stack([
torch.sqrt(1 - u[:, 0]) * torch.sin(2 * math.pi * u[:, 1]),
torch.sqrt(1 - u[:, 0]) * torch.cos(2 * math.pi * u[:, 1]),
torch.sqrt(u[:, 0]) * torch.sin(2 * math.pi * u[:, 2]),
torch.sqrt(u[:, 0]) * torch.cos(2 * math.pi * u[:, 2]),
], dim=-1)
return q
def quaternion_from_direction(normals: torch.Tensor) -> torch.Tensor:
"""
Compute quaternion that aligns Z-axis with given normal vectors.
normals: (N, 3) β unit vectors
Returns: (N, 4) β unit quaternions
"""
N = normals.shape[0]
z = torch.tensor([0.0, 0.0, 1.0], device=normals.device).expand(N, 3)
# Cross product z Γ normal
cross = torch.cross(z, normals, dim=-1)
dot = (z * normals).sum(dim=-1, keepdim=True)
# Quaternion: q = [cross, 1 + dot] then normalize
q = torch.cat([cross, 1.0 + dot], dim=-1)
# Handle anti-parallel case (dot β -1)
anti = dot.squeeze(-1) < -0.999
if anti.any():
# Use perpendicular vector
perp = torch.zeros_like(normals[anti])
perp[:, 0] = 1.0 # arbitrary perpendicular
q[anti] = torch.cat([perp, torch.zeros(anti.sum(), 1, device=normals.device)], dim=-1)
return F.normalize(q, dim=-1)
def generate_sphere_gaussians(center: torch.Tensor, radius: float,
color: torch.Tensor, n_gaussians: int,
device: str = 'cpu') -> torch.Tensor:
"""
Generate Gaussians for a sphere surface.
Returns: (n_gaussians, 14) β full Gaussian params
[pos(3), scale(3), rot(4), opacity(1), color(3)]
"""
# Points on sphere surface
normals = fibonacci_sphere(n_gaussians, device)
positions = center.unsqueeze(0) + radius * normals
# Scale: flat discs covering the surface
splat_radius = radius * math.sqrt(4 * math.pi / n_gaussians) * 0.5
log_scale = math.log(max(splat_radius, 1e-6))
scales = torch.full((n_gaussians, 3), log_scale, device=device)
scales[:, 2] = log_scale - 1.5 # Thin in normal direction
# Rotation: align with surface normal
rotations = quaternion_from_direction(normals)
# Opacity (pre-sigmoid): high opacity
opacity = torch.full((n_gaussians, 1), 2.0, device=device) # sigmoid(2) β 0.88
# Color: uniform with slight variation
colors = color.unsqueeze(0).expand(n_gaussians, 3).clone()
colors += torch.randn_like(colors) * 0.02 # Slight variation
colors = colors.clamp(0, 1)
return torch.cat([positions, scales, rotations, opacity, colors], dim=-1)
def generate_cube_gaussians(center: torch.Tensor, size: float,
color: torch.Tensor, n_gaussians: int,
device: str = 'cpu') -> torch.Tensor:
"""Generate Gaussians for a cube surface."""
per_face = n_gaussians // 6
remainder = n_gaussians - per_face * 6
half = size / 2
all_gaussians = []
# 6 faces: +X, -X, +Y, -Y, +Z, -Z
face_normals = torch.tensor([
[1, 0, 0], [-1, 0, 0], [0, 1, 0], [0, -1, 0], [0, 0, 1], [0, 0, -1]
], dtype=torch.float32, device=device)
for i, normal in enumerate(face_normals):
n = per_face + (1 if i < remainder else 0)
if n == 0:
continue
# Random points on face
uv = torch.rand(n, 2, device=device) * size - half
# Face center offset
face_center = center + normal * half
# Map UV to 3D based on face normal axis
abs_normal = normal.abs()
if abs_normal[0] > 0.5: # X face
positions = torch.stack([
torch.full((n,), face_center[0].item(), device=device),
uv[:, 0] + center[1],
uv[:, 1] + center[2],
], dim=-1)
elif abs_normal[1] > 0.5: # Y face
positions = torch.stack([
uv[:, 0] + center[0],
torch.full((n,), face_center[1].item(), device=device),
uv[:, 1] + center[2],
], dim=-1)
else: # Z face
positions = torch.stack([
uv[:, 0] + center[0],
uv[:, 1] + center[1],
torch.full((n,), face_center[2].item(), device=device),
], dim=-1)
splat_radius = size / math.sqrt(per_face) * 0.6
log_scale = math.log(max(splat_radius, 1e-6))
scales = torch.full((n, 3), log_scale, device=device)
scales[:, 2] = log_scale - 1.5
rotations = quaternion_from_direction(normal.unsqueeze(0).expand(n, 3))
opacity = torch.full((n, 1), 2.0, device=device)
colors = color.unsqueeze(0).expand(n, 3).clone()
colors += torch.randn_like(colors) * 0.02
colors = colors.clamp(0, 1)
all_gaussians.append(torch.cat([positions, scales, rotations, opacity, colors], dim=-1))
return torch.cat(all_gaussians, dim=0)
def generate_scene(n_objects: int = None, n_gaussians_per_object: int = 128,
device: str = 'cpu', scene_scale: float = 2.0) -> dict:
"""
Generate a random procedural scene with Gaussian splats.
Returns:
dict with:
gaussians: (N, 14) β all Gaussians in the scene
description: str β text description of the scene
objects: list of dicts β object metadata
"""
if n_objects is None:
n_objects = random.randint(1, 5)
all_gaussians = []
objects = []
descriptions = []
color_names = {
'red': torch.tensor([0.9, 0.15, 0.1]),
'green': torch.tensor([0.1, 0.85, 0.2]),
'blue': torch.tensor([0.1, 0.2, 0.9]),
'yellow': torch.tensor([0.95, 0.9, 0.1]),
'purple': torch.tensor([0.7, 0.1, 0.8]),
'orange': torch.tensor([0.95, 0.5, 0.05]),
'white': torch.tensor([0.9, 0.9, 0.9]),
'cyan': torch.tensor([0.1, 0.9, 0.9]),
}
color_list = list(color_names.items())
shape_types = ['sphere', 'cube']
for i in range(n_objects):
shape = random.choice(shape_types)
color_name, color_val = random.choice(color_list)
color_val = color_val.to(device)
# Random position within scene bounds
center = (torch.rand(3, device=device) - 0.5) * scene_scale * 1.2
size = random.uniform(0.3, 0.8)
if shape == 'sphere':
gs = generate_sphere_gaussians(center, size, color_val,
n_gaussians_per_object, device)
desc = f"a {color_name} sphere"
else:
gs = generate_cube_gaussians(center, size, color_val,
n_gaussians_per_object, device)
desc = f"a {color_name} cube"
all_gaussians.append(gs)
objects.append({'shape': shape, 'color': color_name, 'center': center.cpu(),
'size': size})
descriptions.append(desc)
gaussians = torch.cat(all_gaussians, dim=0)
description = " and ".join(descriptions)
return {
'gaussians': gaussians,
'description': description,
'objects': objects,
'n_gaussians': gaussians.shape[0],
}
class ProceduralGaussianDataset(Dataset):
"""
On-the-fly procedural Gaussian scene dataset.
Generates random scenes with spheres and cubes.
Each sample: (gaussians, description_text)
For VQ-VAE Phase 1: use .get_gaussian_params() to get pooled params.
For Head Phase 2: use __getitem__ for (text, gaussians) pairs.
"""
def __init__(self, size: int = 10000, gaussians_per_object: int = 128,
max_objects: int = 5, scene_scale: float = 2.0,
device: str = 'cpu', fixed_seed: bool = False):
self.size = size
self.gaussians_per_object = gaussians_per_object
self.max_objects = max_objects
self.scene_scale = scene_scale
self.device = device
self.fixed_seed = fixed_seed
def __len__(self) -> int:
return self.size
def __getitem__(self, idx: int) -> dict:
if self.fixed_seed:
# Deterministic for eval
torch.manual_seed(idx)
random.seed(idx)
scene = generate_scene(
n_objects=random.randint(1, self.max_objects),
n_gaussians_per_object=self.gaussians_per_object,
device=self.device,
scene_scale=self.scene_scale,
)
return scene
def get_gaussian_params(self, n_samples: int = 100000) -> torch.Tensor:
"""
Get a flat pool of Gaussian non-position params for VQ-VAE training.
Returns: (N, 11) β scale(3), rot(4), opacity(1), color(3)
"""
all_params = []
collected = 0
while collected < n_samples:
scene = generate_scene(
n_objects=random.randint(1, self.max_objects),
n_gaussians_per_object=self.gaussians_per_object,
device='cpu',
scene_scale=self.scene_scale,
)
# Extract non-position params: [3:14] = scale(3) + rot(4) + opacity(1) + color(3)
params = scene['gaussians'][:, 3:] # (N, 11)
all_params.append(params)
collected += params.shape[0]
return torch.cat(all_params, dim=0)[:n_samples]
# βββ Differentiable 2D Renderer βββ
def render_gaussians_2d(gaussians: torch.Tensor, camera_pos: torch.Tensor,
camera_target: torch.Tensor, image_size: int = 64,
fov: float = 60.0, bg_color: torch.Tensor = None) -> torch.Tensor:
"""
Simple differentiable Gaussian splatting renderer.
Projects 3D Gaussians to 2D, evaluates 2D Gaussian at each pixel,
alpha-composites front-to-back (depth-sorted).
This is NOT a full 3DGS renderer β it's a training-grade approximation
that provides useful gradients. Good enough for learning, not for display.
Args:
gaussians: (N, 14) β [pos(3), scale(3), rot(4), opacity(1), color(3)]
camera_pos: (3,) β camera position in world space
camera_target: (3,) β camera look-at point
image_size: H = W of output image
fov: vertical field of view in degrees
bg_color: (3,) β background color, default black
Returns:
image: (3, H, W) β rendered RGB image [0, 1]
"""
device = gaussians.device
N = gaussians.shape[0]
H = W = image_size
if bg_color is None:
bg_color = torch.zeros(3, device=device)
# Parse Gaussian params
pos = gaussians[:, :3] # (N, 3)
scale = gaussians[:, 3:6].exp() # (N, 3) β from log-scale
opacity = torch.sigmoid(gaussians[:, 10:11]) # (N, 1)
color = gaussians[:, 11:14] # (N, 3) β already [0,1] from sigmoid in head
# ββ Camera setup ββ
forward = F.normalize(camera_target - camera_pos, dim=0)
world_up = torch.tensor([0.0, 1.0, 0.0], device=device)
right = F.normalize(torch.cross(forward, world_up), dim=0)
up = torch.cross(right, forward)
# View matrix (world β camera)
R = torch.stack([right, up, -forward], dim=0) # (3, 3)
t = -R @ camera_pos # (3,)
# Project Gaussians to camera space
cam_pos = (R @ pos.t()).t() + t.unsqueeze(0) # (N, 3)
# Perspective projection
fov_rad = fov * math.pi / 180
focal = 0.5 * H / math.tan(fov_rad / 2)
# Filter: only Gaussians in front of camera
valid = cam_pos[:, 2] < -0.1 # Z points backward in our convention
if valid.sum() == 0:
return bg_color.view(3, 1, 1).expand(3, H, W)
cam_pos_valid = cam_pos[valid]
color_valid = color[valid]
opacity_valid = opacity[valid]
scale_valid = scale[valid]
# Project to pixel coords
px = -focal * cam_pos_valid[:, 0] / cam_pos_valid[:, 2] + W / 2 # (M,)
py = -focal * cam_pos_valid[:, 1] / cam_pos_valid[:, 2] + H / 2
# Sort by depth (front to back)
depths = -cam_pos_valid[:, 2] # Positive = further
sort_idx = depths.argsort()
px = px[sort_idx]
py = py[sort_idx]
color_sorted = color_valid[sort_idx]
opacity_sorted = opacity_valid[sort_idx]
scale_sorted = scale_valid[sort_idx]
depth_sorted = depths[sort_idx]
# Projected 2D scale (approximate)
# Use average of X,Y scale projected by focal/depth
s2d = (scale_sorted[:, :2].mean(dim=-1) * focal / depth_sorted).clamp(min=0.5) # (M,)
# ββ Rasterize via 2D Gaussian evaluation ββ
# Create pixel grid
yy, xx = torch.meshgrid(
torch.arange(H, dtype=torch.float32, device=device),
torch.arange(W, dtype=torch.float32, device=device),
indexing='ij'
) # (H, W) each
# Initialize image
image = bg_color.view(3, 1, 1).expand(3, H, W).clone()
transmittance = torch.ones(H, W, device=device)
# Splat each Gaussian (front to back alpha compositing)
M = px.shape[0]
# Batch for efficiency: process in chunks
chunk_size = min(M, 512)
for start in range(0, M, chunk_size):
end = min(start + chunk_size, M)
chunk_px = px[start:end] # (C,)
chunk_py = py[start:end]
chunk_s = s2d[start:end] # (C,)
chunk_color = color_sorted[start:end] # (C, 3)
chunk_alpha = opacity_sorted[start:end] # (C, 1)
# Distance from each pixel to each Gaussian center
dx = xx.unsqueeze(0) - chunk_px.view(-1, 1, 1) # (C, H, W)
dy = yy.unsqueeze(0) - chunk_py.view(-1, 1, 1)
# 2D Gaussian evaluation
inv_s2 = 1.0 / (chunk_s.view(-1, 1, 1) ** 2 + 1e-6)
power = -0.5 * (dx ** 2 + dy ** 2) * inv_s2 # (C, H, W)
# Gaussian contribution
gauss = torch.exp(power.clamp(min=-8)) # Clamp for numerical stability
alpha = chunk_alpha.view(-1, 1, 1) * gauss # (C, H, W)
# Alpha composite per Gaussian in chunk (sequential β maintains depth order)
for i in range(end - start):
a = alpha[i] * transmittance # (H, W)
c = chunk_color[i] # (3,)
image += c.view(3, 1, 1) * a.unsqueeze(0) # (3, H, W)
transmittance = transmittance * (1 - alpha[i])
return image.clamp(0, 1)
def multi_view_render(gaussians: torch.Tensor, n_views: int = 4,
image_size: int = 64, orbit_radius: float = 4.0,
fov: float = 60.0) -> torch.Tensor:
"""
Render Gaussians from multiple viewpoints on an orbit.
Args:
gaussians: (N, 14)
n_views: number of views
image_size: H = W
orbit_radius: camera distance from origin
fov: field of view
Returns:
images: (n_views, 3, H, W)
"""
device = gaussians.device
images = []
target = torch.zeros(3, device=device)
for i in range(n_views):
angle = 2 * math.pi * i / n_views
# Camera on orbit (XZ plane, Y slightly elevated)
cam_pos = torch.tensor([
orbit_radius * math.cos(angle),
orbit_radius * 0.4, # Slight elevation
orbit_radius * math.sin(angle),
], device=device)
img = render_gaussians_2d(gaussians, cam_pos, target,
image_size=image_size, fov=fov)
images.append(img)
return torch.stack(images, dim=0)
|