Skip to content

stages

Classes

fastvideo.pipelines.basic.magi_human.stages.MagiHumanAudioDecodingStage 

MagiHumanAudioDecodingStage(audio_vae, time_stretching: float = _UPSTREAM_AUDIO_TIME_STRETCH)

Bases: PipelineStage

Decode batch.audio_latents to a waveform using Stable Audio's VAE.

The VAE is loaded lazily by SAAudioVAEModel.sa_audio_vae_model — the first call triggers a snapshot_download (requires HF token + accepted terms on stabilityai/stable-audio-open-1.0).

Source code in fastvideo/pipelines/basic/magi_human/stages/audio_decoding.py 
68
69
70
71
72
73
74
75
def __init__(
    self,
    audio_vae,
    time_stretching: float = _UPSTREAM_AUDIO_TIME_STRETCH,
) -> None:
    super().__init__()
    self.audio_vae = audio_vae
    self.time_stretching = time_stretching

fastvideo.pipelines.basic.magi_human.stages.MagiHumanDenoisingStage 

MagiHumanDenoisingStage(transformer, scheduler, patch_size: tuple[int, int, int] = (1, 2, 2), video_in_channels: int = 192, audio_in_channels: int = 64, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, coords_style: str = 'v2', video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0)

Bases: PipelineStage

UniPC-flow joint denoising with CFG=2 over (video, audio) latents.

Source code in fastvideo/pipelines/basic/magi_human/stages/denoising.py 
 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
def __init__(
    self,
    transformer,
    scheduler,
    patch_size: tuple[int, int, int] = (1, 2, 2),
    video_in_channels: int = 192,
    audio_in_channels: int = 64,
    video_txt_guidance_scale: float = 5.0,
    audio_txt_guidance_scale: float = 5.0,
    cfg_number: int = 2,
    coords_style: str = "v2",
    video_guidance_high_t_threshold: int = 500,
    video_guidance_low_t_value: float = 2.0,
) -> None:
    super().__init__()
    self.transformer = transformer
    self.scheduler = scheduler
    self.patch_size = patch_size
    self.video_in_channels = video_in_channels
    self.audio_in_channels = audio_in_channels
    self.video_txt_guidance_scale = video_txt_guidance_scale
    self.audio_txt_guidance_scale = audio_txt_guidance_scale
    self.cfg_number = cfg_number
    self.coords_style = coords_style
    self.video_guidance_high_t_threshold = video_guidance_high_t_threshold
    self.video_guidance_low_t_value = video_guidance_low_t_value

fastvideo.pipelines.basic.magi_human.stages.MagiHumanLatentPreparationStage 

MagiHumanLatentPreparationStage(vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, patch_size: tuple[int, int, int] = (1, 2, 2), fps: int = 25, t5_gemma_target_length: int = 640, coords_style: Literal['v1', 'v2'] = 'v2', text_offset: int = 0, audio_in_channels: int = 64)

Bases: PipelineStage

Prepare latents, coords, modality maps, and padded text embed.

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
def __init__(
    self,
    vae_stride: tuple[int, int, int] = (4, 16, 16),
    z_dim: int = 48,
    patch_size: tuple[int, int, int] = (1, 2, 2),
    fps: int = 25,
    t5_gemma_target_length: int = 640,
    coords_style: Literal["v1", "v2"] = "v2",
    text_offset: int = 0,
    audio_in_channels: int = 64,
) -> None:
    super().__init__()
    self.vae_stride = vae_stride
    self.z_dim = z_dim
    self.patch_size = patch_size
    self.fps = fps
    self.t5_gemma_target_length = t5_gemma_target_length
    self.coords_style = coords_style
    self.text_offset = text_offset
    self.audio_in_channels = audio_in_channels

fastvideo.pipelines.basic.magi_human.stages.MagiHumanReferenceImageStage 

MagiHumanReferenceImageStage(vae: Any, vae_scale_factor: int = 16)

Bases: PipelineStage

Encode a TI2V reference image into the first-frame video latent.

Source code in fastvideo/pipelines/basic/magi_human/stages/reference_image.py 
46
47
48
49
def __init__(self, vae: Any, vae_scale_factor: int = 16) -> None:
    super().__init__()
    self.vae = vae
    self.video_processor = VideoProcessor(vae_scale_factor=vae_scale_factor)

fastvideo.pipelines.basic.magi_human.stages.MagiHumanSRDenoisingStage 

MagiHumanSRDenoisingStage(transformer, scheduler, patch_size: tuple[int, int, int] = (1, 2, 2), video_in_channels: int = 192, audio_in_channels: int = 64, sr_num_inference_steps: int = 5, sr_video_txt_guidance_scale: float = 3.5, use_cfg_trick: bool = True, cfg_trick_start_frame: int = 13, cfg_trick_value: float = 2.0, cfg_number: int = 2, coords_style: str = 'v1')

Bases: PipelineStage

Denoise only the SR video latent; audio passes through unchanged.

Source code in fastvideo/pipelines/basic/magi_human/stages/sr_denoising.py 
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
def __init__(
    self,
    transformer,
    scheduler,
    patch_size: tuple[int, int, int] = (1, 2, 2),
    video_in_channels: int = 192,
    audio_in_channels: int = 64,
    sr_num_inference_steps: int = 5,
    sr_video_txt_guidance_scale: float = 3.5,
    use_cfg_trick: bool = True,
    cfg_trick_start_frame: int = 13,
    cfg_trick_value: float = 2.0,
    cfg_number: int = 2,
    coords_style: str = "v1",
) -> None:
    super().__init__()
    self.transformer = transformer
    self.scheduler = scheduler
    self.patch_size = patch_size
    self.video_in_channels = video_in_channels
    self.audio_in_channels = audio_in_channels
    self.sr_num_inference_steps = sr_num_inference_steps
    self.sr_video_txt_guidance_scale = sr_video_txt_guidance_scale
    self.use_cfg_trick = use_cfg_trick
    self.cfg_trick_start_frame = cfg_trick_start_frame
    self.cfg_trick_value = cfg_trick_value
    self.cfg_number = cfg_number
    self.coords_style = coords_style

fastvideo.pipelines.basic.magi_human.stages.MagiHumanSRLatentPreparationStage 

MagiHumanSRLatentPreparationStage(vae: Any, vae_stride: tuple[int, int, int] = (4, 16, 16), patch_size: tuple[int, int, int] = (1, 2, 2), noise_value: int = 220, sr_audio_noise_scale: float = 0.7, sr_height: int = 512, sr_width: int = 896, vae_scale_factor: int = 16)

Bases: PipelineStage

Upsample base latents, add SR noise, and refresh SR conditioning.

Source code in fastvideo/pipelines/basic/magi_human/stages/sr_latent_preparation.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
def __init__(
    self,
    vae: Any,
    vae_stride: tuple[int, int, int] = (4, 16, 16),
    patch_size: tuple[int, int, int] = (1, 2, 2),
    noise_value: int = 220,
    sr_audio_noise_scale: float = 0.7,
    sr_height: int = 512,
    sr_width: int = 896,
    vae_scale_factor: int = 16,
) -> None:
    super().__init__()
    self.vae = vae
    self.vae_stride = vae_stride
    self.patch_size = patch_size
    self.noise_value = noise_value
    self.sr_audio_noise_scale = sr_audio_noise_scale
    self.sr_height = sr_height
    self.sr_width = sr_width
    self.sigmas = ZeroSNRDDPMDiscretization()(1000, do_append_zero=False, flip=True)
    self.video_processor = VideoProcessor(vae_scale_factor=vae_scale_factor)

Modules

fastvideo.pipelines.basic.magi_human.stages.audio_decoding

Audio decoding stage for daVinci-MagiHuman.

Takes the denoised audio latent that MagiHumanDenoisingStage leaves on batch.audio_latents and decodes it to a waveform using the Stable Audio Open 1.0 VAE. Mirrors the upstream post-process path (see MagiEvaluator.post_process in daVinci-MagiHuman/inference/pipeline/video_generate.py:503):

latent_audio.squeeze(0)                 # (L, C_latent)
audio = self.audio_vae.decode(latent_audio.T)   # (1, audio_ch, samples)
audio = audio.squeeze(0).T.cpu().numpy()        # (samples, audio_ch)
audio = resample_audio_sinc(audio, _UPSTREAM_AUDIO_TIME_STRETCH)

The stage stores the resampled waveform on batch.extra["audio"] (shape [samples, audio_channels]) and the sample rate on batch.extra["audio_sample_rate"]. FastVideo's VideoGenerator._mux_audio then reads those, writes a temp wav, and muxes it into the output mp4 via PyAV — same plumbing LTX-2 and Stable Audio use.

Classes

fastvideo.pipelines.basic.magi_human.stages.audio_decoding.MagiHumanAudioDecodingStage 
MagiHumanAudioDecodingStage(audio_vae, time_stretching: float = _UPSTREAM_AUDIO_TIME_STRETCH)

Bases: PipelineStage

Decode batch.audio_latents to a waveform using Stable Audio's VAE.

The VAE is loaded lazily by SAAudioVAEModel.sa_audio_vae_model — the first call triggers a snapshot_download (requires HF token + accepted terms on stabilityai/stable-audio-open-1.0).

Source code in fastvideo/pipelines/basic/magi_human/stages/audio_decoding.py 
68
69
70
71
72
73
74
75
def __init__(
    self,
    audio_vae,
    time_stretching: float = _UPSTREAM_AUDIO_TIME_STRETCH,
) -> None:
    super().__init__()
    self.audio_vae = audio_vae
    self.time_stretching = time_stretching

fastvideo.pipelines.basic.magi_human.stages.denoising

Joint-modality denoising stage for daVinci-MagiHuman base text-to-AV.

Runs the FlowUniPC denoise loop with CFG=2 over video + audio latents jointly. Text embeddings are already pad-or-trimmed to t5_gemma_target_length by MagiHumanLatentPreparationStage; the original context lengths are stashed on the batch as magi_original_text_lens / magi_original_neg_text_lens.

Classes

fastvideo.pipelines.basic.magi_human.stages.denoising.MagiHumanDenoisingStage 
MagiHumanDenoisingStage(transformer, scheduler, patch_size: tuple[int, int, int] = (1, 2, 2), video_in_channels: int = 192, audio_in_channels: int = 64, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, coords_style: str = 'v2', video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0)

Bases: PipelineStage

UniPC-flow joint denoising with CFG=2 over (video, audio) latents.

Source code in fastvideo/pipelines/basic/magi_human/stages/denoising.py 
 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
def __init__(
    self,
    transformer,
    scheduler,
    patch_size: tuple[int, int, int] = (1, 2, 2),
    video_in_channels: int = 192,
    audio_in_channels: int = 64,
    video_txt_guidance_scale: float = 5.0,
    audio_txt_guidance_scale: float = 5.0,
    cfg_number: int = 2,
    coords_style: str = "v2",
    video_guidance_high_t_threshold: int = 500,
    video_guidance_low_t_value: float = 2.0,
) -> None:
    super().__init__()
    self.transformer = transformer
    self.scheduler = scheduler
    self.patch_size = patch_size
    self.video_in_channels = video_in_channels
    self.audio_in_channels = audio_in_channels
    self.video_txt_guidance_scale = video_txt_guidance_scale
    self.audio_txt_guidance_scale = audio_txt_guidance_scale
    self.cfg_number = cfg_number
    self.coords_style = coords_style
    self.video_guidance_high_t_threshold = video_guidance_high_t_threshold
    self.video_guidance_low_t_value = video_guidance_low_t_value

Functions

fastvideo.pipelines.basic.magi_human.stages.latent_preparation

Latent preparation stage for daVinci-MagiHuman base text-to-AV.

Produces
  • random video latent of shape [1, z_dim, latent_T, latent_H, latent_W],
  • random audio latent of shape [1, num_frames, 64] (the DiT jointly denoises both modalities),
  • padded T5-Gemma text embedding (target length 640) plus the original (pre-pad) context length, which the UniPC + CFG loop needs so the unconditional path sees the same padded length.

Also stakes out the per-token coords / modality map that the DiT consumes (replicates the reference MagiDataProxy.process_input).

Classes

fastvideo.pipelines.basic.magi_human.stages.latent_preparation.MagiHumanLatentPreparationStage 
MagiHumanLatentPreparationStage(vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, patch_size: tuple[int, int, int] = (1, 2, 2), fps: int = 25, t5_gemma_target_length: int = 640, coords_style: Literal['v1', 'v2'] = 'v2', text_offset: int = 0, audio_in_channels: int = 64)

Bases: PipelineStage

Prepare latents, coords, modality maps, and padded text embed.

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
def __init__(
    self,
    vae_stride: tuple[int, int, int] = (4, 16, 16),
    z_dim: int = 48,
    patch_size: tuple[int, int, int] = (1, 2, 2),
    fps: int = 25,
    t5_gemma_target_length: int = 640,
    coords_style: Literal["v1", "v2"] = "v2",
    text_offset: int = 0,
    audio_in_channels: int = 64,
) -> None:
    super().__init__()
    self.vae_stride = vae_stride
    self.z_dim = z_dim
    self.patch_size = patch_size
    self.fps = fps
    self.t5_gemma_target_length = t5_gemma_target_length
    self.coords_style = coords_style
    self.text_offset = text_offset
    self.audio_in_channels = audio_in_channels
fastvideo.pipelines.basic.magi_human.stages.latent_preparation.StaticPackedInputs 
StaticPackedInputs(video_tokens: Tensor, audio_tokens: Tensor, video_coords: Tensor, audio_coords: Tensor, video_mm: Tensor, audio_mm: Tensor, max_ch: int)

Step-invariant packed inputs: video+audio tokens, coords, modality map.

Computed once before the denoise loop; reused for every cond/uncond call. Text tokens are NOT included here because cond/uncond have different lengths.

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
def __init__(
    self,
    video_tokens: torch.Tensor,
    audio_tokens: torch.Tensor,
    video_coords: torch.Tensor,
    audio_coords: torch.Tensor,
    video_mm: torch.Tensor,
    audio_mm: torch.Tensor,
    max_ch: int,
) -> None:
    self.video_tokens = video_tokens
    self.audio_tokens = audio_tokens
    self.video_coords = video_coords
    self.audio_coords = audio_coords
    self.video_mm = video_mm
    self.audio_mm = audio_mm
    self.video_token_num = video_tokens.size(0)
    self.audio_feat_len = audio_tokens.size(0)
    self.max_ch = max_ch
fastvideo.pipelines.basic.magi_human.stages.latent_preparation.StaticPackedLayout 
StaticPackedLayout(video_coords: Tensor, audio_coords: Tensor, video_mm: Tensor, audio_mm: Tensor, max_ch: int, video_token_num: int, audio_feat_len: int)

Step- and value-invariant portion of the static packed inputs.

Coords, modality maps, and the channel-padding width depend only on the latent shape, audio length, channel widths, and patch sizes — all fixed for a single generation. Precompute once before the denoise loop and reuse on every step. Only the per-step token tensors must be rebuilt.

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
def __init__(
    self,
    video_coords: torch.Tensor,
    audio_coords: torch.Tensor,
    video_mm: torch.Tensor,
    audio_mm: torch.Tensor,
    max_ch: int,
    video_token_num: int,
    audio_feat_len: int,
) -> None:
    self.video_coords = video_coords
    self.audio_coords = audio_coords
    self.video_mm = video_mm
    self.audio_mm = audio_mm
    self.max_ch = max_ch
    self.video_token_num = video_token_num
    self.audio_feat_len = audio_feat_len

Functions

fastvideo.pipelines.basic.magi_human.stages.latent_preparation.assemble_packed_inputs 
assemble_packed_inputs(static: StaticPackedInputs, txt_feat: Tensor, txt_feat_len: int, coords_style: Literal['v1', 'v2'] = 'v2', text_offset: int = 0) -> tuple[Tensor, Tensor, Tensor]

Attach per-call text tokens to the precomputed static packed inputs.

Returns (token_seq, coords, modality_map) ready for the DiT.

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
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
520
521
def assemble_packed_inputs(
    static: StaticPackedInputs,
    txt_feat: torch.Tensor,
    txt_feat_len: int,
    coords_style: Literal["v1", "v2"] = "v2",
    text_offset: int = 0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """Attach per-call text tokens to the precomputed static packed inputs.

    Returns (token_seq, coords, modality_map) ready for the DiT.
    """
    text_tokens = txt_feat[0, :txt_feat_len].contiguous()
    max_ch = max(static.max_ch, text_tokens.size(-1))

    video_tokens = F.pad(static.video_tokens, (0, max_ch - static.video_tokens.size(-1)))
    audio_tokens = F.pad(static.audio_tokens, (0, max_ch - static.audio_tokens.size(-1)))
    text_tokens = F.pad(text_tokens, (0, max_ch - text_tokens.size(-1)))
    token_seq = torch.cat([video_tokens, audio_tokens, text_tokens], dim=0)

    device = token_seq.device
    dtype = token_seq.dtype
    text_mm = torch.full((txt_feat_len, ), MODALITY_TEXT, dtype=torch.int64, device=device)
    mm = torch.cat([static.video_mm, static.audio_mm, text_mm], dim=0)

    if coords_style == "v2":
        text_coords = _build_coords(
            shape=(txt_feat_len, 1, 1),
            ref_feat_shape=(1, 1, 1),
            offset_thw=(-txt_feat_len, 0, 0),
            device=device,
            dtype=dtype,
        )
    else:
        text_coords = _build_coords(
            shape=(txt_feat_len, 1, 1),
            ref_feat_shape=_V1_TEXT_REF_SHAPE,
            offset_thw=(text_offset, 0, 0),
            device=device,
            dtype=dtype,
        )

    coords = torch.cat([static.video_coords, static.audio_coords, text_coords], dim=0)
    return token_seq, coords, mm
fastvideo.pipelines.basic.magi_human.stages.latent_preparation.build_packed_inputs 
build_packed_inputs(video_latent: Tensor, audio_latent: Tensor, audio_feat_len: int, txt_feat: Tensor, txt_feat_len: int, patch_size: tuple[int, int, int], coords_style: Literal['v1', 'v2'] = 'v2', text_offset: int = 0) -> tuple[Tensor, Tensor, Tensor]

Build the full packed token stream in one call (backwards-compat wrapper).

Equivalent to assemble_packed_inputs(build_static_packed_inputs(...), ...). Prefer calling the two helpers separately when the static portion can be reused across multiple calls (e.g. cond/uncond in the denoise loop).

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
def build_packed_inputs(
    video_latent: torch.Tensor,
    audio_latent: torch.Tensor,
    audio_feat_len: int,
    txt_feat: torch.Tensor,
    txt_feat_len: int,
    patch_size: tuple[int, int, int],
    coords_style: Literal["v1", "v2"] = "v2",
    text_offset: int = 0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """Build the full packed token stream in one call (backwards-compat wrapper).

    Equivalent to assemble_packed_inputs(build_static_packed_inputs(...), ...).
    Prefer calling the two helpers separately when the static portion can be
    reused across multiple calls (e.g. cond/uncond in the denoise loop).
    """
    static = build_static_packed_inputs(
        video_latent=video_latent,
        audio_latent=audio_latent,
        audio_feat_len=audio_feat_len,
        patch_size=patch_size,
        coords_style=coords_style,
    )
    return assemble_packed_inputs(
        static=static,
        txt_feat=txt_feat,
        txt_feat_len=txt_feat_len,
        coords_style=coords_style,
        text_offset=text_offset,
    )
fastvideo.pipelines.basic.magi_human.stages.latent_preparation.build_static_packed_inputs 
build_static_packed_inputs(video_latent: Tensor, audio_latent: Tensor, audio_feat_len: int, patch_size: tuple[int, int, int], coords_style: Literal['v1', 'v2'] = 'v2', layout: StaticPackedLayout | None = None) -> StaticPackedInputs

Build the step-invariant portion of the packed token stream.

Returns video+audio tokens (padded to a common channel width), their coords, and their modality slices. Text is excluded because cond/uncond differ in length; call assemble_packed_inputs to attach text per call.

Mirrors SingleData.token_sequence / coords_mapping / modality_mapping in inference/pipeline/data_proxy.py, minus the text portion.

When layout is provided, coords / modality maps / max_ch are taken from the precomputed values and only the per-step token tensors are rebuilt; this is the hot-path call from the denoising loop. When layout is None the function recomputes everything from scratch (e.g. for one-shot tests via build_packed_inputs).

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
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
def build_static_packed_inputs(
    video_latent: torch.Tensor,
    audio_latent: torch.Tensor,
    audio_feat_len: int,
    patch_size: tuple[int, int, int],
    coords_style: Literal["v1", "v2"] = "v2",
    layout: StaticPackedLayout | None = None,
) -> StaticPackedInputs:
    """Build the step-invariant portion of the packed token stream.

    Returns video+audio tokens (padded to a common channel width), their
    coords, and their modality slices. Text is excluded because cond/uncond
    differ in length; call assemble_packed_inputs to attach text per call.

    Mirrors SingleData.token_sequence / coords_mapping / modality_mapping in
    inference/pipeline/data_proxy.py, minus the text portion.

    When ``layout`` is provided, coords / modality maps / max_ch are taken
    from the precomputed values and only the per-step token tensors are
    rebuilt; this is the hot-path call from the denoising loop. When
    ``layout`` is None the function recomputes everything from scratch
    (e.g. for one-shot tests via ``build_packed_inputs``).
    """
    pT, pH, pW = patch_size
    assert video_latent.size(0) == 1, "batch size 1 required for MagiHuman base"

    video_tokens = _img2tokens(video_latent, t_patch=pT, patch=pH)[0]
    audio_tokens = audio_latent[0, :audio_feat_len].contiguous()

    if layout is not None:
        max_ch = layout.max_ch
        video_tokens = F.pad(video_tokens, (0, max_ch - video_tokens.size(-1)))
        audio_tokens = F.pad(audio_tokens, (0, max_ch - audio_tokens.size(-1)))
        return StaticPackedInputs(
            video_tokens=video_tokens,
            audio_tokens=audio_tokens,
            video_coords=layout.video_coords,
            audio_coords=layout.audio_coords,
            video_mm=layout.video_mm,
            audio_mm=layout.audio_mm,
            max_ch=max_ch,
        )

    # Slow path: rebuild every invariant from scratch. Kept for the
    # ``build_packed_inputs`` one-shot wrapper used by tests/parity helpers.
    _, z_dim, T, H, W = video_latent.shape

    max_ch = max(video_tokens.size(-1), audio_tokens.size(-1))
    video_tokens = F.pad(video_tokens, (0, max_ch - video_tokens.size(-1)))
    audio_tokens = F.pad(audio_tokens, (0, max_ch - audio_tokens.size(-1)))

    device = video_tokens.device
    dtype = video_tokens.dtype
    video_token_num = video_tokens.size(0)

    video_mm = torch.full((video_token_num, ), MODALITY_VIDEO, dtype=torch.int64, device=device)
    audio_mm = torch.full((audio_feat_len, ), MODALITY_AUDIO, dtype=torch.int64, device=device)

    video_ref_shape = (T // pT, H // pH, W // pW)
    video_coords = _build_coords(
        shape=video_ref_shape,
        ref_feat_shape=video_ref_shape,
        device=device,
        dtype=dtype,
    )

    if coords_style == "v2":
        audio_ref_t = (audio_feat_len - 1) // _AUDIO_TEMPORAL_COMPRESSION + 1
        audio_coords = _build_coords(
            shape=(audio_feat_len, 1, 1),
            ref_feat_shape=(audio_ref_t // pT, 1, 1),
            device=device,
            dtype=dtype,
        )
    else:
        audio_coords = _build_coords(
            shape=(audio_feat_len, 1, 1),
            ref_feat_shape=(T // pT, 1, 1),
            device=device,
            dtype=dtype,
        )

    return StaticPackedInputs(
        video_tokens=video_tokens,
        audio_tokens=audio_tokens,
        video_coords=video_coords,
        audio_coords=audio_coords,
        video_mm=video_mm,
        audio_mm=audio_mm,
        max_ch=max_ch,
    )
fastvideo.pipelines.basic.magi_human.stages.latent_preparation.precompute_static_packed_layout 
precompute_static_packed_layout(latent_shape: tuple[int, int, int, int, int], audio_feat_len: int, z_dim: int, audio_in_channels: int, patch_size: tuple[int, int, int], coords_style: Literal['v1', 'v2'] = 'v2', device: device | None = None, dtype: dtype = float32) -> StaticPackedLayout

Precompute the invariant fields used by build_static_packed_inputs.

Arguments are derived from configs and latent shape — none depend on the current denoising-step values. Call this once in the latent preparation stage (or any pre-loop site) and pass the result via the layout= arg of build_static_packed_inputs to skip the meshgrid/full() work on every step.

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
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
def precompute_static_packed_layout(
    latent_shape: tuple[int, int, int, int, int],
    audio_feat_len: int,
    z_dim: int,
    audio_in_channels: int,
    patch_size: tuple[int, int, int],
    coords_style: Literal["v1", "v2"] = "v2",
    device: torch.device | None = None,
    dtype: torch.dtype = torch.float32,
) -> StaticPackedLayout:
    """Precompute the invariant fields used by ``build_static_packed_inputs``.

    Arguments are derived from configs and latent shape — none depend on
    the current denoising-step values. Call this once in the latent
    preparation stage (or any pre-loop site) and pass the result via the
    ``layout=`` arg of ``build_static_packed_inputs`` to skip the
    meshgrid/full() work on every step.
    """
    pT, pH, pW = patch_size
    _, _, T, H, W = latent_shape
    if device is None:
        device = torch.device("cpu")

    video_token_num = (T // pT) * (H // pH) * (W // pW)
    # `_img2tokens` packs to channel `z_dim * pT * pH * pW`; audio tokens
    # are `audio_in_channels` wide — both are config constants.
    max_ch = max(z_dim * pT * pH * pW, audio_in_channels)

    video_ref_shape = (T // pT, H // pH, W // pW)
    video_coords = _build_coords(
        shape=video_ref_shape,
        ref_feat_shape=video_ref_shape,
        device=device,
        dtype=dtype,
    )

    if coords_style == "v2":
        audio_ref_t = (audio_feat_len - 1) // _AUDIO_TEMPORAL_COMPRESSION + 1
        audio_coords = _build_coords(
            shape=(audio_feat_len, 1, 1),
            ref_feat_shape=(audio_ref_t // pT, 1, 1),
            device=device,
            dtype=dtype,
        )
    else:
        audio_coords = _build_coords(
            shape=(audio_feat_len, 1, 1),
            ref_feat_shape=(T // pT, 1, 1),
            device=device,
            dtype=dtype,
        )

    video_mm = torch.full((video_token_num, ), MODALITY_VIDEO, dtype=torch.int64, device=device)
    audio_mm = torch.full((audio_feat_len, ), MODALITY_AUDIO, dtype=torch.int64, device=device)

    return StaticPackedLayout(
        video_coords=video_coords,
        audio_coords=audio_coords,
        video_mm=video_mm,
        audio_mm=audio_mm,
        max_ch=max_ch,
        video_token_num=video_token_num,
        audio_feat_len=audio_feat_len,
    )
fastvideo.pipelines.basic.magi_human.stages.latent_preparation.unpack_tokens 
unpack_tokens(output: Tensor, video_token_num: int, audio_feat_len: int, video_in_channels: int, audio_in_channels: int, latent_shape: tuple[int, int, int, int, int], patch_size: tuple[int, int, int]) -> tuple[Tensor, Tensor]

Inverse of build_packed_inputs for the DiT output.

Splits the flat output back into a video latent (un-patched into B C T H W) and an audio latent (B, L, 64).

Source code in fastvideo/pipelines/basic/magi_human/stages/latent_preparation.py 
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
def unpack_tokens(
    output: torch.Tensor,  # [L, max(V_ch, A_ch)]
    video_token_num: int,
    audio_feat_len: int,
    video_in_channels: int,
    audio_in_channels: int,
    latent_shape: tuple[int, int, int, int, int],  # [1, z_dim, T, H, W]
    patch_size: tuple[int, int, int],
) -> tuple[torch.Tensor, torch.Tensor]:
    """Inverse of `build_packed_inputs` for the DiT output.

    Splits the flat output back into a video latent (un-patched into
    B C T H W) and an audio latent (B, L, 64).
    """
    pT, pH, pW = patch_size
    _, z_dim, T, H, W = latent_shape
    tH, tW = H // pH, W // pW

    video_flat = output[:video_token_num, :video_in_channels]
    video_latent = rearrange(
        video_flat,
        "(T H W) (pT pH pW C) -> C (T pT) (H pH) (W pW)",
        H=tH,
        W=tW,
        pT=pT,
        pH=pH,
        pW=pW,
    ).contiguous().unsqueeze(0)

    audio_latent = output[
        video_token_num:video_token_num + audio_feat_len,
        :audio_in_channels,
    ].unsqueeze(0)

    return video_latent, audio_latent

fastvideo.pipelines.basic.magi_human.stages.reference_image

Reference-image encoding for MagiHuman TI2V.

The upstream daVinci-MagiHuman TI2V path encodes the user image through the Wan VAE and overwrites the first denoising latent frame with that clean latent at every step. This stage mirrors MagiEvaluator.encode_image and stashes the normalized latent on batch.image_latent for the latent-prep and denoise stages.

Classes

fastvideo.pipelines.basic.magi_human.stages.reference_image.MagiHumanReferenceImageStage 
MagiHumanReferenceImageStage(vae: Any, vae_scale_factor: int = 16)

Bases: PipelineStage

Encode a TI2V reference image into the first-frame video latent.

Source code in fastvideo/pipelines/basic/magi_human/stages/reference_image.py 
46
47
48
49
def __init__(self, vae: Any, vae_scale_factor: int = 16) -> None:
    super().__init__()
    self.vae = vae
    self.video_processor = VideoProcessor(vae_scale_factor=vae_scale_factor)

Functions

fastvideo.pipelines.basic.magi_human.stages.sr_denoising

SR video-only denoising stage for daVinci-MagiHuman SR-540p.

Classes

fastvideo.pipelines.basic.magi_human.stages.sr_denoising.MagiHumanSRDenoisingStage 
MagiHumanSRDenoisingStage(transformer, scheduler, patch_size: tuple[int, int, int] = (1, 2, 2), video_in_channels: int = 192, audio_in_channels: int = 64, sr_num_inference_steps: int = 5, sr_video_txt_guidance_scale: float = 3.5, use_cfg_trick: bool = True, cfg_trick_start_frame: int = 13, cfg_trick_value: float = 2.0, cfg_number: int = 2, coords_style: str = 'v1')

Bases: PipelineStage

Denoise only the SR video latent; audio passes through unchanged.

Source code in fastvideo/pipelines/basic/magi_human/stages/sr_denoising.py 
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
def __init__(
    self,
    transformer,
    scheduler,
    patch_size: tuple[int, int, int] = (1, 2, 2),
    video_in_channels: int = 192,
    audio_in_channels: int = 64,
    sr_num_inference_steps: int = 5,
    sr_video_txt_guidance_scale: float = 3.5,
    use_cfg_trick: bool = True,
    cfg_trick_start_frame: int = 13,
    cfg_trick_value: float = 2.0,
    cfg_number: int = 2,
    coords_style: str = "v1",
) -> None:
    super().__init__()
    self.transformer = transformer
    self.scheduler = scheduler
    self.patch_size = patch_size
    self.video_in_channels = video_in_channels
    self.audio_in_channels = audio_in_channels
    self.sr_num_inference_steps = sr_num_inference_steps
    self.sr_video_txt_guidance_scale = sr_video_txt_guidance_scale
    self.use_cfg_trick = use_cfg_trick
    self.cfg_trick_start_frame = cfg_trick_start_frame
    self.cfg_trick_value = cfg_trick_value
    self.cfg_number = cfg_number
    self.coords_style = coords_style

Functions

fastvideo.pipelines.basic.magi_human.stages.sr_latent_preparation

Super-resolution latent preparation for daVinci-MagiHuman SR-540p.

Classes

fastvideo.pipelines.basic.magi_human.stages.sr_latent_preparation.MagiHumanSRLatentPreparationStage 
MagiHumanSRLatentPreparationStage(vae: Any, vae_stride: tuple[int, int, int] = (4, 16, 16), patch_size: tuple[int, int, int] = (1, 2, 2), noise_value: int = 220, sr_audio_noise_scale: float = 0.7, sr_height: int = 512, sr_width: int = 896, vae_scale_factor: int = 16)

Bases: PipelineStage

Upsample base latents, add SR noise, and refresh SR conditioning.

Source code in fastvideo/pipelines/basic/magi_human/stages/sr_latent_preparation.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
def __init__(
    self,
    vae: Any,
    vae_stride: tuple[int, int, int] = (4, 16, 16),
    patch_size: tuple[int, int, int] = (1, 2, 2),
    noise_value: int = 220,
    sr_audio_noise_scale: float = 0.7,
    sr_height: int = 512,
    sr_width: int = 896,
    vae_scale_factor: int = 16,
) -> None:
    super().__init__()
    self.vae = vae
    self.vae_stride = vae_stride
    self.patch_size = patch_size
    self.noise_value = noise_value
    self.sr_audio_noise_scale = sr_audio_noise_scale
    self.sr_height = sr_height
    self.sr_width = sr_width
    self.sigmas = ZeroSNRDDPMDiscretization()(1000, do_append_zero=False, flip=True)
    self.video_processor = VideoProcessor(vae_scale_factor=vae_scale_factor)
fastvideo.pipelines.basic.magi_human.stages.sr_latent_preparation.ZeroSNRDDPMDiscretization 
ZeroSNRDDPMDiscretization(linear_start: float = 0.00085, linear_end: float = 0.012, num_timesteps: int = 1000, shift_scale: float = 1.0, keep_start: bool = False, post_shift: bool = False)

Upstream ZeroSNR schedule used to corrupt interpolated SR latents.

Source code in fastvideo/pipelines/basic/magi_human/stages/sr_latent_preparation.py 
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
def __init__(
    self,
    linear_start: float = 0.00085,
    linear_end: float = 0.0120,
    num_timesteps: int = 1000,
    shift_scale: float = 1.0,
    keep_start: bool = False,
    post_shift: bool = False,
) -> None:
    if keep_start and not post_shift:
        linear_start = linear_start / (shift_scale + (1 - shift_scale) * linear_start)
    self.num_timesteps = num_timesteps
    betas = torch.linspace(
        linear_start**0.5,
        linear_end**0.5,
        num_timesteps,
        dtype=torch.float64,
    )**2
    alphas = 1.0 - betas.numpy()
    self.alphas_cumprod = np.cumprod(alphas, axis=0)
    self.post_shift = post_shift
    self.shift_scale = shift_scale

    if not post_shift:
        self.alphas_cumprod = self.alphas_cumprod / (shift_scale + (1 - shift_scale) * self.alphas_cumprod)

Functions