File size: 12,849 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 | #!/usr/bin/env python3
"""CPU regression for exact master-audio latent masking."""
import os
import sys
import types
import torch
TESTS = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, TESTS)
import _masked_prefix_unit_test as harness # noqa: E402
def main():
harness._install_comfy_stubs()
package = types.ModuleType(harness.PACKAGE)
package.__path__ = [harness.ROOT]
sys.modules[harness.PACKAGE] = package
harness._load("patch_layout")
harness._load("patch_payload")
nodes = harness._load("nodes")
harness._load("masked_context")
audio_context = harness._load("master_audio_context")
audio_context.require_h3_mask_support = lambda _operation: True
target_frames = 141
video_steps = 42
audio_steps = 235
assert nodes._pixel_frames(video_steps) == target_frames
target_video = torch.zeros((1, 16, video_steps, 2, 4))
target_audio = torch.zeros((1, 32, 2, audio_steps))
latent = {
"samples": harness.NestedTensor((target_video, target_audio)),
}
previous_frames = torch.rand((120, 32, 64, 3))
master_audio = {
"waveform": torch.rand((1, 2, 32_000 * 30)),
"sample_rate": 32_000,
}
class VideoVAE:
def encode(self, frames):
count = int(frames.shape[0])
steps = max(1, (count - 5) // 17 * 5 + 2)
return torch.full((1, 16, steps, 2, 4), 0.25)
class AudioVAE:
audio_sample_rate = 32_000
def encode(self, waveform):
steps = round(int(waveform.shape[1]) / 32_000 * 40)
return torch.full((1, 32, 2, steps), 0.5)
output, prefix, clip_audio = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
latent,
AudioVAE(),
master_audio,
clip_start_seconds=3.25,
context_length=39,
source_fps=24.0,
crop="disabled",
vae=VideoVAE(),
source_frames=previous_frames,
))
video, audio = output["samples"].unbind()
video_mask, audio_mask = output["noise_mask"].unbind()
assert prefix == 39
assert torch.allclose(
video[:, :, :12], torch.full_like(video[:, :, :12], 0.25))
assert not torch.count_nonzero(video[:, :, 12:])
assert torch.allclose(audio, torch.full_like(audio, 0.5))
assert not torch.count_nonzero(video_mask[:, :, :12])
assert torch.all(video_mask[:, :, 12:] == 1.0)
assert not torch.count_nonzero(audio_mask)
assert clip_audio["sample_rate"] == 32_000
assert int(clip_audio["waveform"].shape[-1]) == round(
target_frames / 24 * 32_000)
# Preferred live chaining copies the sampled H3 video tail directly. The
# source latent's audio stream must not replace the authoritative master.
source_video = torch.arange(
video_steps, dtype=torch.float32).reshape(
1, 1, video_steps, 1, 1).expand_as(target_video).clone()
source_audio = torch.full_like(target_audio, 0.9)
source_latent = {
"samples": harness.NestedTensor((source_video, source_audio)),
}
live, live_prefix, _ = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
latent,
AudioVAE(),
master_audio,
clip_start_seconds=3.25,
context_length=39,
source_latent=source_latent,
))
live_video, live_audio = live["samples"].unbind()
live_video_mask, live_audio_mask = live["noise_mask"].unbind()
assert live_prefix == 39
assert torch.equal(live_video[:, :, :12], source_video[:, :, -12:])
assert not torch.count_nonzero(live_video[:, :, 12:])
assert torch.allclose(live_audio, torch.full_like(live_audio, 0.5))
assert not torch.count_nonzero(live_video_mask[:, :, :12])
assert torch.all(live_video_mask[:, :, 12:] == 1.0)
assert not torch.count_nonzero(live_audio_mask)
try:
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
latent,
AudioVAE(),
master_audio,
context_length=39,
vae=VideoVAE(),
source_frames=previous_frames,
source_latent=source_latent,
)
except ValueError as exc:
assert "either source_latent or source_frames" in str(exc)
else:
raise AssertionError("both master-audio video sources were accepted")
assert "source_latent" in (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV.INPUT_TYPES()[
"optional"])
assert "voice" in (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV.INPUT_TYPES()[
"optional"])
assert audio_context.MiniMaxH3ContexMasterAudioMaskedAV.INPUT_TYPES()[
"required"]["preroll_seconds"][1]["default"] == 0.0
# Chain Context may already own the protected visual prefix. Replacing
# the audio target must retain that video mask while making the complete
# current master-audio stream protected.
existing_video_mask = torch.ones((1, 1, video_steps, 2, 4))
existing_video_mask[:, :, :7] = 0.0
existing_audio_mask = torch.ones((1, 1, 2, audio_steps))
chained_latent = {
"samples": harness.NestedTensor((target_video, target_audio)),
"noise_mask": harness.NestedTensor((
existing_video_mask, existing_audio_mask)),
}
chained, chained_prefix, _ = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
chained_latent,
AudioVAE(),
master_audio,
clip_start_seconds=3.25,
context_length=0,
))
chained_video_mask, chained_audio_mask = chained["noise_mask"].unbind()
assert chained_prefix == 0
assert not torch.count_nonzero(chained_video_mask[:, :, :7])
assert torch.all(chained_video_mask[:, :, 7:] == 1.0)
assert not torch.count_nonzero(chained_audio_mask)
first, first_prefix, _ = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
latent,
AudioVAE(),
master_audio,
clip_start_seconds=0.0,
context_length=0,
))
first_video_mask, first_audio_mask = first["noise_mask"].unbind()
assert first_prefix == 0
assert torch.all(first_video_mask == 1.0)
assert not torch.count_nonzero(first_audio_mask)
class FloorAudioVAE:
"""Simulate an audio encoder that floors its temporal output."""
audio_sample_rate = 32_000
def __init__(self):
self.encoded_samples = []
def encode(self, waveform):
samples = int(waveform.shape[1])
self.encoded_samples.append(samples)
steps = int(samples / self.audio_sample_rate * 40)
return torch.full((1, 32, 2, steps), 0.75)
rounding_frames = 124
rounding_video_steps = 37
rounding_audio_steps = 207
assert nodes._pixel_frames(rounding_video_steps) == rounding_frames
rounding_latent = {
"samples": harness.NestedTensor((
torch.zeros((1, 16, rounding_video_steps, 2, 4)),
torch.zeros((1, 32, 2, rounding_audio_steps)),
)),
}
floor_vae = FloorAudioVAE()
rounded, rounded_prefix, rounded_clip_audio = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
rounding_latent,
floor_vae,
master_audio,
clip_start_seconds=1.0,
context_length=0,
))
_, rounded_audio = rounded["samples"].unbind()
_, rounded_audio_mask = rounded["noise_mask"].unbind()
picture_samples = round(rounding_frames / 24 * 32_000)
grid_samples = int(rounding_audio_steps / 40 * 32_000)
assert floor_vae.encoded_samples == [grid_samples]
assert grid_samples > picture_samples
assert int(rounded_audio.shape[-1]) == rounding_audio_steps
assert torch.allclose(
rounded_audio, torch.full_like(rounded_audio, 0.75))
assert not torch.count_nonzero(rounded_audio_mask)
assert rounded_prefix == 0
assert int(rounded_clip_audio["waveform"].shape[-1]) == picture_samples
# Absolute endpoints avoid the occasional one-sample error caused by
# rounding a relative duration and adding it to an already-rounded start.
awkward_start = 0.3 / 32_000.0
absolute_start = round(awkward_start * 32_000)
absolute_end = round(
(awkward_start + rounding_frames / 24.0) * 32_000)
_, _, absolute_clip_audio = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
rounding_latent,
FloorAudioVAE(),
master_audio,
clip_start_seconds=awkward_start,
context_length=0,
))
assert int(absolute_clip_audio["waveform"].shape[-1]) == (
absolute_end - absolute_start)
class ShortFirstAudioVAE(FloorAudioVAE):
def encode(self, waveform):
encoded = super().encode(waveform)
if len(self.encoded_samples) == 1:
return encoded[..., :-1]
return encoded
retry_vae = ShortFirstAudioVAE()
retried, _, _ = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
rounding_latent,
retry_vae,
master_audio,
clip_start_seconds=1.0,
context_length=0,
))
_, retried_audio = retried["samples"].unbind()
assert len(retry_vae.encoded_samples) == 2
assert retry_vae.encoded_samples[0] == grid_samples
assert retry_vae.encoded_samples[1] > grid_samples
assert int(retried_audio.shape[-1]) == rounding_audio_steps
assert torch.allclose(
retried_audio, torch.full_like(retried_audio, 0.75))
class IndexedAudioVAE:
audio_sample_rate = 32_000
def __init__(self):
self.encoded_samples = []
def encode(self, waveform):
samples = int(waveform.shape[1])
self.encoded_samples.append(samples)
steps = round(samples / self.audio_sample_rate * 40)
ticks = torch.arange(steps, dtype=torch.float32).reshape(
1, 1, 1, steps)
return ticks.expand(1, 32, 2, steps).clone()
indexed_vae = IndexedAudioVAE()
contextual, _, contextual_clip = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
rounding_latent,
indexed_vae,
master_audio,
clip_start_seconds=2.0,
context_length=0,
preroll_seconds=1.0,
lookahead_seconds=0.2,
))
_, contextual_audio = contextual["samples"].unbind()
_, contextual_mask = contextual["noise_mask"].unbind()
expected_context_steps = 40 + rounding_audio_steps + 8
assert indexed_vae.encoded_samples == [
int(expected_context_steps / 40 * 32_000)]
assert torch.equal(
contextual_audio[0, 0, 0],
torch.arange(40, 40 + rounding_audio_steps, dtype=torch.float32))
assert not torch.count_nonzero(contextual_mask)
assert int(contextual_clip["waveform"].shape[-1]) == picture_samples
voice = {
"waveform": torch.zeros((1, 1, 32_000 * 30)),
"sample_rate": 32_000,
}
voice["waveform"][..., 32_000:64_000] = 1.0
gated, _, _ = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
latent,
AudioVAE(),
master_audio,
clip_start_seconds=0.0,
context_length=0,
audio_denoise=0.0,
gap_denoise=0.3,
gate_hold_seconds=0.0,
voice=voice,
))
_, gated_mask = gated["noise_mask"].unbind()
assert torch.all(gated_mask[..., 10] == 0.3)
assert not torch.count_nonzero(gated_mask[..., 50])
assert torch.all(gated_mask[..., 100] == 0.3)
softened, _, _ = (
audio_context.MiniMaxH3ContexMasterAudioMaskedAV().prepare(
latent,
AudioVAE(),
master_audio,
clip_start_seconds=0.0,
context_length=0,
audio_denoise=0.25,
))
_, softened_mask = softened["noise_mask"].unbind()
assert torch.all(softened_mask == 0.25)
assert (
"MiniMaxH3ContexMasterAudioMaskedAV"
in audio_context.NODE_CLASS_MAPPINGS)
assert (
"MiniMaxH3ContexSongMaskedAVContext"
not in audio_context.NODE_CLASS_MAPPINGS)
assert (
"MiniMaxH3SongMaskedAVContext"
not in audio_context.NODE_CLASS_MAPPINGS)
print(
"master-audio masking: absolute-endpoint timeline slice, complete protected audio, "
"40 Hz target-grid lookahead, contextual tick crop, vocal gate, "
"fractional song denoise, decoded and direct-latent 39-frame "
"protected video continuation, "
"and clip-1 video generation pass")
if __name__ == "__main__":
main()
|