File size: 21,219 Bytes
bc50ea0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 | """Public general masked-target nodes for MiniMax H3 video editing."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from .masking_ops import (
GRID_MODES,
H3_PIXEL_CELL,
H3_VIDEO_FPS,
MASK_CONVERSION_MODES,
floor_h3_frame_count,
normalize_comfy_mask,
quantize_h3_pixel_mask,
reduce_h3_mask_to_video_latent,
resize_mask_to_video_latent,
temporal_audio_mask,
trim_audio_to_frames,
)
from .masking_support import require_h3_mask_support
MASK_MEANINGS = ("white = generate", "white = preserve")
AUDIO_MASK_MODES = (
"preserve source audio",
"generate all audio",
"follow video mask",
"custom audio mask",
)
def _nested_parts(value, label):
if isinstance(value, (tuple, list)):
return list(value)
if bool(getattr(value, "is_nested", False)) and hasattr(value, "unbind"):
return list(value.unbind())
raise ValueError(
"%s must be a nested H3 video/audio value, not a single tensor." %
label
)
def _target_streams(latent):
if not isinstance(latent, dict) or latent.get("samples") is None:
raise ValueError("H3 Masked Target needs a LATENT containing samples.")
streams = _nested_parts(latent["samples"], "H3 target latent")
if len(streams) < 2:
raise ValueError(
"H3 Masked Target needs both video and audio streams. Encode the "
"source media and combine it as a joint AV latent first."
)
video, audio = streams[:2]
if not isinstance(video, torch.Tensor) or video.ndim != 5:
raise ValueError(
"H3 target video latent must be [B,C,T,H,W], got %s." %
(list(getattr(video, "shape", ())),)
)
if not isinstance(audio, torch.Tensor) or audio.ndim != 4:
raise ValueError(
"H3 target audio latent must be [B,C,channels,T], got %s." %
(list(getattr(audio, "shape", ())),)
)
if int(video.shape[0]) != 1 or int(audio.shape[0]) != 1:
raise ValueError("H3 masked targets currently support batch size 1.")
return video, audio
def _existing_video_mask(value, video):
if value.ndim == 5:
work = value[:, :1].to(device=video.device, dtype=torch.float32)
if int(work.shape[0]) != 1:
raise ValueError("Existing H3 video mask has an unsupported batch.")
if tuple(work.shape[2:]) != tuple(video.shape[2:]):
work = F.interpolate(
work,
size=tuple(int(v) for v in video.shape[2:]),
mode="trilinear",
align_corners=False,
)
return work.clamp(0.0, 1.0).contiguous()
if value.ndim == 4 and int(value.shape[0]) == 1:
return _existing_video_mask(value.unsqueeze(1), video)
return resize_mask_to_video_latent(value, video)
def _existing_audio_mask(value, audio):
work = value.to(device=audio.device, dtype=torch.float32)
if work.ndim == 1:
work = work[None, None, None]
elif work.ndim == 2:
work = work[None, None]
elif work.ndim == 3:
work = work.unsqueeze(0)
if work.ndim != 4 or int(work.shape[0]) != 1:
raise ValueError(
"Existing H3 audio mask must broadcast to [1,C,channels,T]; got "
"%s." % (list(work.shape),)
)
work = work.amax(dim=1, keepdim=True)
if int(work.shape[2]) != int(audio.shape[2]):
work = work.amax(dim=2, keepdim=True).expand(
1, 1, int(audio.shape[2]), int(work.shape[3]))
if int(work.shape[-1]) != int(audio.shape[-1]):
lead = int(work.shape[2])
work = F.interpolate(
work.reshape(lead, 1, int(work.shape[-1])),
size=int(audio.shape[-1]),
mode="linear",
align_corners=False,
).reshape(1, 1, lead, int(audio.shape[-1]))
return work.clamp(0.0, 1.0).contiguous()
def _compose_existing_masks(latent, video_mask, audio_mask, video, audio):
existing = latent.get("noise_mask")
if existing is None:
return video_mask, audio_mask, False
parts = _nested_parts(existing, "Existing H3 noise mask")
if len(parts) < 2:
raise ValueError(
"Existing H3 noise mask has no audio component and cannot be "
"composed safely."
)
old_video = _existing_video_mask(parts[0], video)
old_audio = _existing_audio_mask(parts[1], audio)
# Preserve wins. This lets a spatial edit compose with masked_av's exact
# protected prefix regardless of which node executes first.
return (
torch.minimum(video_mask, old_video),
torch.minimum(audio_mask, old_audio),
True,
)
class MiniMaxH3ContexTrimSourceAV:
@classmethod
def INPUT_TYPES(cls):
return {
"required": {
"frames": ("IMAGE", {
"tooltip": "Source video frames. Trailing frames are "
"dropped to the largest valid H3 17k+5 length; "
"frames are never padded or resized.",
}),
},
"optional": {
"audio": ("AUDIO", {
"tooltip": "Optional synchronized source audio, trimmed "
"to the returned frame duration at 24 fps.",
}),
},
}
RETURN_TYPES = ("IMAGE", "AUDIO", "INT", "STRING")
RETURN_NAMES = ("trimmed_frames", "trimmed_audio", "h3_length",
"trim_info")
OUTPUT_TOOLTIPS = (
"Source frames trimmed to the largest valid H3 17k+5 length.",
"Synchronized audio trimmed to the returned frame duration, or an "
"empty output when no audio was connected.",
"The valid H3 frame count represented by both returned streams.",
"Human-readable summary of the frame and audio trimming performed.",
)
FUNCTION = "trim"
CATEGORY = "conditioning/minimax/context_loop/masking"
DESCRIPTION = ("Trim source video and optional audio to the largest H3 "
"17k+5 run without inventing frames or silence.")
def trim(self, frames, audio=None):
if not isinstance(frames, torch.Tensor) or frames.ndim != 4:
raise ValueError(
"H3 source AV trim expects IMAGE [T,H,W,C], got %s." %
(list(getattr(frames, "shape", ())),)
)
source_count = int(frames.shape[0])
h3_length = floor_h3_frame_count(source_count)
trimmed_frames = frames[:h3_length]
trimmed_audio, target_samples = trim_audio_to_frames(audio, h3_length)
dropped = source_count - h3_length
duration = h3_length / float(H3_VIDEO_FPS)
audio_text = "no audio connected"
if trimmed_audio is not None:
actual = int(trimmed_audio["waveform"].shape[-1])
sample_rate = int(trimmed_audio["sample_rate"])
audio_text = "audio %d/%d samples at %d Hz" % (
actual, target_samples, sample_rate)
if actual < target_samples:
audio_text += " (source shorter; no silence added)"
info = (
"H3 source trim: %d -> %d frames; dropped %d; %.3fs at 24 fps; %s."
% (source_count, h3_length, dropped, duration, audio_text)
)
return trimmed_frames, trimmed_audio, h3_length, info
class MiniMaxH3ContexMaskedTarget:
@classmethod
def INPUT_TYPES(cls):
return {
"required": {
"target_latent": ("LATENT", {
"tooltip": "Joint source video/audio latent used as the "
"actual H3 target. Known regions are preserved "
"from this latent; do not add the same source "
"again merely as a video reference.",
}),
"mask": ("MASK", {
"tooltip": "Static or per-frame video-space mask. It is "
"mapped to H3's causal video-latent frames and "
"2x2 latent-token cells. Exact mode accepts "
"one static mask or one mask per source frame.",
}),
"mask_meaning": (list(MASK_MEANINGS), {
"default": MASK_MEANINGS[0],
"tooltip": "Choose whether white pixels regenerate or "
"remain protected.",
}),
"audio_mode": (list(AUDIO_MASK_MODES), {
"default": AUDIO_MASK_MODES[0],
"tooltip": "Preserve or generate all source audio, map "
"the video mask onto time, or use the optional "
"custom audio mask.",
}),
},
"optional": {
"audio_mask": ("MASK", {
"tooltip": "Used only by custom audio mask. Spatial "
"dimensions are reduced; white time regions "
"generate and black regions remain protected.",
}),
"mask_conversion": (list(MASK_CONVERSION_MODES), {
"default": MASK_CONVERSION_MODES[0],
"tooltip": "H3 exact maps pixel masks with the video "
"VAE's causal 1/4/4/4/4 frame groups and "
"2x2 latent-token max coverage. Legacy "
"trilinear retains the older interpolated "
"conversion for workflow comparison. Changing "
"this changes generation; update the Plan's "
"generation fingerprint before resuming.",
}),
},
}
RETURN_TYPES = ("LATENT", "STRING")
RETURN_NAMES = ("masked_target", "mask_info")
OUTPUT_TOOLTIPS = (
"Copy of target_latent carrying the composed H3 video and audio "
"denoise mask.",
"Summary of the latent dimensions and generated percentages for "
"each stream.",
)
FUNCTION = "apply"
CATEGORY = "conditioning/minimax/context_loop/masking"
DESCRIPTION = ("Attach causal, token-aligned H3 video/audio denoise masks "
"to a real source AV target for inpainting and temporal "
"edits. Ordinary AV extension needs no user mask.")
def apply(
self,
target_latent,
mask,
mask_meaning,
audio_mode,
audio_mask=None,
mask_conversion=MASK_CONVERSION_MODES[0],
):
require_h3_mask_support("general masked-target generation")
video, audio = _target_streams(target_latent)
generation_mask = normalize_comfy_mask(mask)
if mask_meaning == "white = preserve":
generation_mask = 1.0 - generation_mask
elif mask_meaning != "white = generate":
raise ValueError("Unknown H3 mask meaning %r." % mask_meaning)
if mask_conversion == MASK_CONVERSION_MODES[0]:
video_mask = reduce_h3_mask_to_video_latent(
generation_mask, video)
conversion_summary = "H3 exact causal/token max"
elif mask_conversion == MASK_CONVERSION_MODES[1]:
video_mask = resize_mask_to_video_latent(generation_mask, video)
conversion_summary = "legacy trilinear"
else:
raise ValueError(
"Unknown H3 mask conversion mode %r." % mask_conversion)
if audio_mode == "preserve source audio":
out_audio_mask = torch.zeros(
(1, 1, int(audio.shape[2]), int(audio.shape[3])),
device=audio.device,
dtype=torch.float32,
)
elif audio_mode == "generate all audio":
out_audio_mask = torch.ones(
(1, 1, int(audio.shape[2]), int(audio.shape[3])),
device=audio.device,
dtype=torch.float32,
)
elif audio_mode == "follow video mask":
out_audio_mask = temporal_audio_mask(generation_mask, audio)
elif audio_mode == "custom audio mask":
if audio_mask is None:
raise ValueError(
"custom audio mask mode requires the optional audio_mask "
"input. White generates; black preserves."
)
out_audio_mask = temporal_audio_mask(audio_mask, audio)
else:
raise ValueError("Unknown H3 audio mask mode %r." % audio_mode)
video_mask, out_audio_mask, composed = _compose_existing_masks(
target_latent,
video_mask,
out_audio_mask,
video,
audio,
)
import comfy.nested_tensor
output = target_latent.copy()
output["noise_mask"] = comfy.nested_tensor.NestedTensor(
(video_mask, out_audio_mask))
video_percent = 100.0 * float(video_mask.mean().item())
audio_percent = 100.0 * float(out_audio_mask.mean().item())
info = (
"H3 masked target: latent video %dx%dx%d, generate %.2f%%; "
"audio %d steps, generate %.2f%%; %s; mask %s; existing mask %s."
% (
int(video.shape[4]),
int(video.shape[3]),
int(video.shape[2]),
video_percent,
int(audio.shape[-1]),
audio_percent,
audio_mode,
conversion_summary,
"intersected" if composed else "none",
)
)
return output, info
class MiniMaxH3ContexMaskGridPreview:
@classmethod
def INPUT_TYPES(cls):
return {
"required": {
"image": ("IMAGE", {
"tooltip": "Exact resized H3 canvas used by the video "
"VAE. Width and height must be divisible by 32.",
}),
"mask": ("MASK", {
"tooltip": "Static or tracked ComfyUI mask interpreted "
"using mask_meaning.",
}),
"cell_selection": (list(GRID_MODES), {
"default": GRID_MODES[0],
"tooltip": "Runtime exact matches H3 latent reduction. "
"Coverage modes provide deliberate conservative "
"or strict alternatives.",
}),
"cell_adjust": ("INT", {
"default": 0, "min": -8, "max": 8, "step": 1,
"tooltip": "Grow or shrink by complete 32x32 H3 cells.",
}),
"preview_frame": ("INT", {
"default": 0, "min": 0, "max": 10000, "step": 1,
"tooltip": "Frame displayed by grid_preview. The snapped "
"MASK output still contains the complete batch.",
}),
"overlay_opacity": ("FLOAT", {
"default": 0.38, "min": 0.0, "max": 1.0, "step": 0.01,
"tooltip": "Opacity of the orange snapped-generation "
"overlay in grid_preview.",
}),
"show_grid": ("BOOLEAN", {
"default": True,
"tooltip": "Draw the effective 32x32 H3 source-pixel "
"cell grid over the preview.",
}),
"show_source_outline": ("BOOLEAN", {
"default": True,
"tooltip": "Outline the original unsnapped mask so its "
"difference from the effective mask is visible.",
}),
},
"optional": {
"mask_meaning": (list(MASK_MEANINGS), {
"default": MASK_MEANINGS[0],
"tooltip": "Use the same white-pixel convention here "
"and on Apply Target Mask. The overlay always "
"shows the effective generation region.",
}),
},
}
RETURN_TYPES = ("MASK", "IMAGE", "STRING")
RETURN_NAMES = ("snapped_mask", "grid_preview", "grid_info")
OUTPUT_TOOLTIPS = (
"Complete mask batch snapped to H3's effective causal 32x32 "
"source-pixel cells, using the selected white-pixel convention.",
"Selected source frame with the snapped mask and optional grid or "
"source outline overlaid.",
"Summary of the selected frame, grid dimensions, and snapped-mask "
"coverage.",
)
FUNCTION = "preview"
CATEGORY = "conditioning/minimax/context_loop/masking"
DESCRIPTION = ("Preview and snap a video mask to H3's effective 32x32 "
"source-pixel generation cells.")
def preview(
self,
image,
mask,
cell_selection,
cell_adjust,
preview_frame,
overlay_opacity,
show_grid,
show_source_outline,
mask_meaning=MASK_MEANINGS[0],
):
if not isinstance(image, torch.Tensor) or image.ndim != 4:
raise ValueError(
"H3 mask preview expects IMAGE [frames,H,W,C], got %s." %
(list(getattr(image, "shape", ())),)
)
frames, height, width = (int(v) for v in image.shape[:3])
generation_mask = normalize_comfy_mask(mask)
if mask_meaning == "white = preserve":
generation_mask = 1.0 - generation_mask
elif mask_meaning != "white = generate":
raise ValueError("Unknown H3 mask meaning %r." % mask_meaning)
raw, snapped_generation, cells = quantize_h3_pixel_mask(
generation_mask,
frames,
height,
width,
mode=cell_selection,
cell_adjust=int(cell_adjust),
)
selected_frame = min(max(0, int(preview_frame)), frames - 1)
preview = image[selected_frame:selected_frame + 1, ..., :3].to(
torch.float32).clamp(0.0, 1.0).clone()
selected = snapped_generation[selected_frame].to(
device=preview.device, dtype=preview.dtype)[None, ..., None]
orange = torch.tensor(
[1.0, 0.18, 0.0], device=preview.device, dtype=preview.dtype)
opacity = float(overlay_opacity)
preview = preview * (1.0 - selected * opacity) + (
orange * selected * opacity)
if bool(show_grid):
grid = torch.zeros(
(height, width), device=preview.device, dtype=torch.bool)
grid[::H3_PIXEL_CELL, :] = True
grid[:, ::H3_PIXEL_CELL] = True
grid[-1, :] = True
grid[:, -1] = True
cyan = torch.tensor(
[0.0, 0.82, 1.0], device=preview.device,
dtype=preview.dtype)
preview[0, grid] = preview[0, grid] * 0.28 + cyan * 0.72
if bool(show_source_outline):
source = (raw[selected_frame:selected_frame + 1].to(
device=preview.device) >= 0.5).to(preview.dtype)
dilated = F.max_pool2d(
source, kernel_size=3, stride=1, padding=1)
eroded = 1.0 - F.max_pool2d(
1.0 - source, kernel_size=3, stride=1, padding=1)
outline = ((dilated - eroded) > 0).movedim(1, -1)
preview = torch.where(outline, torch.ones_like(preview), preview)
counts = (cells > 0).sum(
dim=(1, 2, 3)).to(torch.int64).cpu().tolist()
cells_per_frame = int(cells.shape[-2]) * int(cells.shape[-1])
if len(counts) <= 8:
count_text = ", ".join(str(value) for value in counts)
else:
count_text = "min %d, max %d, total %d" % (
min(counts), max(counts), sum(counts))
info = (
"H3 grid: %dx%d cells at 32x32 pixels; selected/frame %s / %d; "
"mode %s; %s; adjust %+d; preview frame %d/%d."
% (
int(cells.shape[-1]),
int(cells.shape[-2]),
count_text,
cells_per_frame,
cell_selection,
mask_meaning,
int(cell_adjust),
selected_frame,
frames - 1,
)
)
snapped = (
1.0 - snapped_generation
if mask_meaning == "white = preserve"
else snapped_generation
)
return snapped.contiguous(), preview, info
NODE_CLASS_MAPPINGS = {
"MiniMaxH3ContexTrimSourceAV": MiniMaxH3ContexTrimSourceAV,
"MiniMaxH3ContexMaskedTarget": MiniMaxH3ContexMaskedTarget,
"MiniMaxH3ContexMaskGridPreview": MiniMaxH3ContexMaskGridPreview,
}
NODE_DISPLAY_NAME_MAPPINGS = {
"MiniMaxH3ContexTrimSourceAV": "MiniMax H3 Masking 路 Trim Source AV",
"MiniMaxH3ContexMaskedTarget": "MiniMax H3 Masking 路 Apply Target Mask",
"MiniMaxH3ContexMaskGridPreview": "MiniMax H3 Masking 路 Grid Preview",
}
|