Skip to content

magi_human

Modules

fastvideo.pipelines.basic.magi_human.magi_human_pipeline

MagiHuman base text-to-AV pipeline.

Top-level composition for the daVinci-MagiHuman base model. Wires:

InputValidationStage -> TextEncodingStage (T5-Gemma)
    -> MagiHumanLatentPreparationStage
    -> MagiHumanDenoisingStage
    -> DecodingStage (Wan 2.2 TI2V-5B VAE decode for video)
    -> MagiHumanAudioDecodingStage (Stable Audio Open 1.0 VAE decode)

The base checkpoint is a joint audio-visual generator; both the video and audio paths run in the denoising loop and both are decoded.

load_modules is overridden so the four cross-variant shared components (text_encoder, tokenizer, audio_vae, video vae) lazy-load from their canonical upstream HF repos at first build time instead of being bundled inside every converted MagiHuman variant. This keeps each variant's converted repo at ~5-30 GB (transformer + scheduler + model_index.json) instead of ~30-55 GB, and lets all variants share the same ~25 GB of cached upstream weights.

Classes

fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanI2VPipeline 
MagiHumanI2VPipeline(model_path: str, fastvideo_args: FastVideoArgs | TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, Module] | None = None)

Bases: MagiHumanPipeline

MagiHuman text+image-to-AV pipeline using the T2V DiT weights.

Source code in fastvideo/pipelines/composed_pipeline_base.py 
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
def __init__(self,
             model_path: str,
             fastvideo_args: FastVideoArgs | TrainingArgs,
             required_config_modules: list[str] | None = None,
             loaded_modules: dict[str, torch.nn.Module] | None = None):
    """
    Initialize the pipeline. After __init__, the pipeline should be ready to
    use. The pipeline should be stateless and not hold any batch state.
    """
    self.fastvideo_args = fastvideo_args

    self.model_path: str = model_path
    self._stages: list[PipelineStage] = []
    self._stage_name_mapping: dict[str, PipelineStage] = {}
    self._trace_mgr = None

    if required_config_modules is not None:
        self._required_config_modules = required_config_modules

    if self._required_config_modules is None:
        raise NotImplementedError("Subclass must set _required_config_modules")

    maybe_init_distributed_environment_and_model_parallel(fastvideo_args.tp_size, fastvideo_args.sp_size)

    # Torch profiler. Enabled and configured through env vars:
    # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace
    trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR
    self.profiler_controller = get_or_create_profiler(trace_dir)
    self.profiler = self.profiler_controller.profiler

    self.local_rank = get_world_group().local_rank

    # Load modules directly in initialization
    logger.info("Loading pipeline modules...")
    with self.profiler_controller.region("profiler_region_model_loading"):
        self.modules = self.load_modules(fastvideo_args, loaded_modules)
fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanPipeline 
MagiHumanPipeline(model_path: str, fastvideo_args: FastVideoArgs | TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, Module] | None = None)

Bases: ComposedPipelineBase

Base MagiHuman text-to-AV pipeline (no LoRA, no distill, no SR).

Source code in fastvideo/pipelines/composed_pipeline_base.py 
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
def __init__(self,
             model_path: str,
             fastvideo_args: FastVideoArgs | TrainingArgs,
             required_config_modules: list[str] | None = None,
             loaded_modules: dict[str, torch.nn.Module] | None = None):
    """
    Initialize the pipeline. After __init__, the pipeline should be ready to
    use. The pipeline should be stateless and not hold any batch state.
    """
    self.fastvideo_args = fastvideo_args

    self.model_path: str = model_path
    self._stages: list[PipelineStage] = []
    self._stage_name_mapping: dict[str, PipelineStage] = {}
    self._trace_mgr = None

    if required_config_modules is not None:
        self._required_config_modules = required_config_modules

    if self._required_config_modules is None:
        raise NotImplementedError("Subclass must set _required_config_modules")

    maybe_init_distributed_environment_and_model_parallel(fastvideo_args.tp_size, fastvideo_args.sp_size)

    # Torch profiler. Enabled and configured through env vars:
    # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace
    trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR
    self.profiler_controller = get_or_create_profiler(trace_dir)
    self.profiler = self.profiler_controller.profiler

    self.local_rank = get_world_group().local_rank

    # Load modules directly in initialization
    logger.info("Loading pipeline modules...")
    with self.profiler_controller.region("profiler_region_model_loading"):
        self.modules = self.load_modules(fastvideo_args, loaded_modules)
Functions
fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanPipeline.load_modules 
load_modules(fastvideo_args: FastVideoArgs, loaded_modules: dict[str, Any] | None = None) -> dict[str, Any]

Load the variant-specific transformer + scheduler from the converted MagiHuman repo and lazy-load the four cross-variant shared components from their canonical upstream HF repos:

* text_encoder, tokenizer -> google/t5gemma-9b-9b-ul2 (gated, requires HF token with accepted terms of use) * audio_vae -> stabilityai/stable-audio-open-1.0 (gated) * vae -> Wan-AI/Wan2.2-TI2V-5B-Diffusers

Backwards-compatible with bundled converted repos: if any of these subfolders is present locally and listed in model_index.json, the standard component loader picks it up via super(). Otherwise the loader is told to skip the entry and we lazy-load it here.

Source code in fastvideo/pipelines/basic/magi_human/magi_human_pipeline.py 
 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
def load_modules(
    self,
    fastvideo_args: FastVideoArgs,
    loaded_modules: dict[str, Any] | None = None,
) -> dict[str, Any]:
    """Load the variant-specific transformer + scheduler from the
    converted MagiHuman repo and lazy-load the four cross-variant
    shared components from their canonical upstream HF repos:

      * text_encoder, tokenizer -> ``google/t5gemma-9b-9b-ul2``
        (gated, requires HF token with accepted terms of use)
      * audio_vae -> ``stabilityai/stable-audio-open-1.0`` (gated)
      * vae -> ``Wan-AI/Wan2.2-TI2V-5B-Diffusers``

    Backwards-compatible with bundled converted repos: if any of
    these subfolders is present locally and listed in
    ``model_index.json``, the standard component loader picks it up
    via super(). Otherwise the loader is told to skip the entry and
    we lazy-load it here.
    """
    # T5-Gemma is gated: expose `HF_API_KEY` as `HF_TOKEN` if needed.
    _ensure_hf_token_env()

    # Resolve to a local cache path so we can inspect
    # model_index.json before invoking super(). `maybe_download_model`
    # is idempotent for local paths; super() repeats the call cheaply
    # via `_load_config`.
    local_path = maybe_download_model(self.model_path)

    # Identify which cross-variant shared keys are bundled in the
    # converted repo (declared in model_index.json with a non-null
    # spec) versus absent (the umbrella scheme). Bundled keys stay
    # in `required_config_modules` and are loaded normally by super()
    # from `<model_path>/<key>/`. Absent keys are temporarily
    # dropped so super() does not fail the "every required entry
    # must appear in model_index.json" check, then lazy-loaded
    # below.
    model_index: dict[str, Any] = {}
    try:
        with open(Path(local_path) / "model_index.json") as f:
            model_index = json.load(f)
    except (FileNotFoundError, json.JSONDecodeError):
        pass

    def _is_bundled(key: str) -> bool:
        spec = model_index.get(key)
        return (isinstance(spec, list | tuple) and len(spec) >= 1 and spec[0] is not None)

    deferred = []
    for key in ("text_encoder", "tokenizer", "audio_vae", "vae"):
        if key in self.required_config_modules and not _is_bundled(key):
            self.required_config_modules.remove(key)
            deferred.append(key)

    try:
        modules = super().load_modules(fastvideo_args, loaded_modules)
    finally:
        for key in deferred:
            if key not in self.required_config_modules:
                self.required_config_modules.append(key)

    # For each lazy-load key, prefer whatever super() already loaded
    # (a bundled subfolder, or a caller-provided override merged in
    # via `loaded_modules`). Fall back to the caller-provided
    # `loaded_modules` entry for keys absent from model_index.json
    # (super() never iterates those). Otherwise lazy-load from the
    # canonical upstream HF repo.
    def _resolve(key: str) -> bool:
        """Return True if `modules[key]` is already populated."""
        if modules.get(key) is not None:
            return True
        if loaded_modules and key in loaded_modules:
            modules[key] = loaded_modules[key]
            return True
        return False

    if not _resolve("text_encoder"):
        logger.info("Building T5-Gemma text encoder (lazy-load from %s)", _T5GEMMA_HF_ID)
        enc_config = T5GemmaEncoderConfig()
        enc_config.arch_config.t5gemma_model_path = _T5GEMMA_HF_ID
        modules["text_encoder"] = T5GemmaEncoderModel(enc_config)

    if not _resolve("tokenizer"):
        logger.info("Loading T5-Gemma tokenizer from %s", _T5GEMMA_HF_ID)
        modules["tokenizer"] = AutoTokenizer.from_pretrained(_T5GEMMA_HF_ID)

    if not _resolve("audio_vae"):
        logger.info(
            "Building Stable Audio Open 1.0 VAE (lazy-load from %s) — "
            "requires HF terms accepted for gated repo",
            _SA_AUDIO_HF_ID,
        )
        audio_config = OobleckVAEConfig()
        audio_config.pretrained_path = _SA_AUDIO_HF_ID
        modules["audio_vae"] = SAAudioVAEModel(audio_config)

    if not _resolve("vae"):
        modules["vae"] = self._load_video_vae(fastvideo_args)

    return modules
fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanSR1080pI2VPipeline 
MagiHumanSR1080pI2VPipeline(model_path: str, fastvideo_args: FastVideoArgs | TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, Module] | None = None)

Bases: MagiHumanSRI2VPipeline

Two-stage MagiHuman base + SR-1080p text+image-to-AV pipeline.

Source code in fastvideo/pipelines/composed_pipeline_base.py 
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
def __init__(self,
             model_path: str,
             fastvideo_args: FastVideoArgs | TrainingArgs,
             required_config_modules: list[str] | None = None,
             loaded_modules: dict[str, torch.nn.Module] | None = None):
    """
    Initialize the pipeline. After __init__, the pipeline should be ready to
    use. The pipeline should be stateless and not hold any batch state.
    """
    self.fastvideo_args = fastvideo_args

    self.model_path: str = model_path
    self._stages: list[PipelineStage] = []
    self._stage_name_mapping: dict[str, PipelineStage] = {}
    self._trace_mgr = None

    if required_config_modules is not None:
        self._required_config_modules = required_config_modules

    if self._required_config_modules is None:
        raise NotImplementedError("Subclass must set _required_config_modules")

    maybe_init_distributed_environment_and_model_parallel(fastvideo_args.tp_size, fastvideo_args.sp_size)

    # Torch profiler. Enabled and configured through env vars:
    # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace
    trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR
    self.profiler_controller = get_or_create_profiler(trace_dir)
    self.profiler = self.profiler_controller.profiler

    self.local_rank = get_world_group().local_rank

    # Load modules directly in initialization
    logger.info("Loading pipeline modules...")
    with self.profiler_controller.region("profiler_region_model_loading"):
        self.modules = self.load_modules(fastvideo_args, loaded_modules)
fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanSR1080pPipeline 
MagiHumanSR1080pPipeline(model_path: str, fastvideo_args: FastVideoArgs | TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, Module] | None = None)

Bases: MagiHumanSRPipeline

Two-stage MagiHuman base + SR-1080p text-to-AV pipeline.

The stage chain is identical to SR-540p. The paired pipeline config enables block-sparse local-window attention on 32 SR-DiT layers and requests the 1080p latent target.

Source code in fastvideo/pipelines/composed_pipeline_base.py 
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
def __init__(self,
             model_path: str,
             fastvideo_args: FastVideoArgs | TrainingArgs,
             required_config_modules: list[str] | None = None,
             loaded_modules: dict[str, torch.nn.Module] | None = None):
    """
    Initialize the pipeline. After __init__, the pipeline should be ready to
    use. The pipeline should be stateless and not hold any batch state.
    """
    self.fastvideo_args = fastvideo_args

    self.model_path: str = model_path
    self._stages: list[PipelineStage] = []
    self._stage_name_mapping: dict[str, PipelineStage] = {}
    self._trace_mgr = None

    if required_config_modules is not None:
        self._required_config_modules = required_config_modules

    if self._required_config_modules is None:
        raise NotImplementedError("Subclass must set _required_config_modules")

    maybe_init_distributed_environment_and_model_parallel(fastvideo_args.tp_size, fastvideo_args.sp_size)

    # Torch profiler. Enabled and configured through env vars:
    # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace
    trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR
    self.profiler_controller = get_or_create_profiler(trace_dir)
    self.profiler = self.profiler_controller.profiler

    self.local_rank = get_world_group().local_rank

    # Load modules directly in initialization
    logger.info("Loading pipeline modules...")
    with self.profiler_controller.region("profiler_region_model_loading"):
        self.modules = self.load_modules(fastvideo_args, loaded_modules)
fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanSRI2VPipeline 
MagiHumanSRI2VPipeline(model_path: str, fastvideo_args: FastVideoArgs | TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, Module] | None = None)

Bases: MagiHumanSRPipeline

Two-stage MagiHuman base + SR-540p text+image-to-AV pipeline.

Source code in fastvideo/pipelines/composed_pipeline_base.py 
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
def __init__(self,
             model_path: str,
             fastvideo_args: FastVideoArgs | TrainingArgs,
             required_config_modules: list[str] | None = None,
             loaded_modules: dict[str, torch.nn.Module] | None = None):
    """
    Initialize the pipeline. After __init__, the pipeline should be ready to
    use. The pipeline should be stateless and not hold any batch state.
    """
    self.fastvideo_args = fastvideo_args

    self.model_path: str = model_path
    self._stages: list[PipelineStage] = []
    self._stage_name_mapping: dict[str, PipelineStage] = {}
    self._trace_mgr = None

    if required_config_modules is not None:
        self._required_config_modules = required_config_modules

    if self._required_config_modules is None:
        raise NotImplementedError("Subclass must set _required_config_modules")

    maybe_init_distributed_environment_and_model_parallel(fastvideo_args.tp_size, fastvideo_args.sp_size)

    # Torch profiler. Enabled and configured through env vars:
    # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace
    trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR
    self.profiler_controller = get_or_create_profiler(trace_dir)
    self.profiler = self.profiler_controller.profiler

    self.local_rank = get_world_group().local_rank

    # Load modules directly in initialization
    logger.info("Loading pipeline modules...")
    with self.profiler_controller.region("profiler_region_model_loading"):
        self.modules = self.load_modules(fastvideo_args, loaded_modules)
fastvideo.pipelines.basic.magi_human.magi_human_pipeline.MagiHumanSRPipeline 
MagiHumanSRPipeline(model_path: str, fastvideo_args: FastVideoArgs | TrainingArgs, required_config_modules: list[str] | None = None, loaded_modules: dict[str, Module] | None = None)

Bases: MagiHumanPipeline

Two-stage MagiHuman base + SR-540p text-to-AV pipeline.

Source code in fastvideo/pipelines/composed_pipeline_base.py 
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
def __init__(self,
             model_path: str,
             fastvideo_args: FastVideoArgs | TrainingArgs,
             required_config_modules: list[str] | None = None,
             loaded_modules: dict[str, torch.nn.Module] | None = None):
    """
    Initialize the pipeline. After __init__, the pipeline should be ready to
    use. The pipeline should be stateless and not hold any batch state.
    """
    self.fastvideo_args = fastvideo_args

    self.model_path: str = model_path
    self._stages: list[PipelineStage] = []
    self._stage_name_mapping: dict[str, PipelineStage] = {}
    self._trace_mgr = None

    if required_config_modules is not None:
        self._required_config_modules = required_config_modules

    if self._required_config_modules is None:
        raise NotImplementedError("Subclass must set _required_config_modules")

    maybe_init_distributed_environment_and_model_parallel(fastvideo_args.tp_size, fastvideo_args.sp_size)

    # Torch profiler. Enabled and configured through env vars:
    # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace
    trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR
    self.profiler_controller = get_or_create_profiler(trace_dir)
    self.profiler = self.profiler_controller.profiler

    self.local_rank = get_world_group().local_rank

    # Load modules directly in initialization
    logger.info("Loading pipeline modules...")
    with self.profiler_controller.region("profiler_region_model_loading"):
        self.modules = self.load_modules(fastvideo_args, loaded_modules)

Functions

fastvideo.pipelines.basic.magi_human.pipeline_configs

PipelineConfig for the daVinci-MagiHuman base text-to-AV pipeline.

Classes

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanBaseConfig dataclass 
MagiHumanBaseConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 32, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0)

Bases: PipelineConfig

Base MagiHuman text-to-AV pipeline config (prompt → video + audio).

MagiHuman's base model is a joint audio-visual generator. This config wires up both the video VAE (Wan 2.2 TI2V-5B) and the audio VAE (Stable Audio Open 1.0); the pipeline produces an mp4 with a muxed audio track. The framework's WorkloadType enum has no T2AV variant yet, so the registry entry uses WorkloadType.T2V as a placeholder.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanBaseI2VConfig dataclass 
MagiHumanBaseI2VConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 32, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0, image_conditioning: bool = True)

Bases: MagiHumanBaseConfig

Base MagiHuman text+image-to-AV pipeline config.

TI2V reuses the T2V DiT weights; the only pipeline-side difference is that a reference image is encoded with the Wan VAE and reinserted into the first video-latent frame before every denoise step.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanDistillConfig dataclass 
MagiHumanDistillConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 8, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 1, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0)

Bases: MagiHumanBaseConfig

DMD-2 distilled MagiHuman text-to-AV pipeline config.

Same arch as base (identical 331 keys, same shapes, same module tree), but trained via DMD-2 for 8-step inference without classifier-free guidance. Weights are stored in fp32 upstream; the conversion script's --cast-bf16 flag reduces the checkpoint to ~30 GB on disk.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanDistillI2VConfig dataclass 
MagiHumanDistillI2VConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 8, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 1, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0, image_conditioning: bool = True)

Bases: MagiHumanDistillConfig

DMD-2 distilled MagiHuman text+image-to-AV pipeline config.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanSR1080pConfig dataclass 
MagiHumanSR1080pConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 32, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0, noise_value: int = 220, sr_audio_noise_scale: float = 0.7, 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, sr_height: int = 1080, sr_width: int = 1920, sr_local_attn_layers: tuple[int, ...] = _SR_1080P_LOCAL_ATTN_LAYERS)

Bases: MagiHumanSR540pConfig

Two-stage MagiHuman base + SR-1080p text-to-AV pipeline config.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanSR1080pI2VConfig dataclass 
MagiHumanSR1080pI2VConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 32, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0, noise_value: int = 220, sr_audio_noise_scale: float = 0.7, 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, sr_height: int = 1080, sr_width: int = 1920, sr_local_attn_layers: tuple[int, ...] = _SR_1080P_LOCAL_ATTN_LAYERS, image_conditioning: bool = True)

Bases: MagiHumanSR1080pConfig

Two-stage MagiHuman base + SR-1080p text+image-to-AV config.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanSR540pConfig dataclass 
MagiHumanSR540pConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 32, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0, noise_value: int = 220, sr_audio_noise_scale: float = 0.7, 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, sr_height: int = 512, sr_width: int = 896, sr_local_attn_layers: tuple[int, ...] = ())

Bases: MagiHumanBaseConfig

Two-stage MagiHuman base + SR-540p text-to-AV pipeline config.

fastvideo.pipelines.basic.magi_human.pipeline_configs.MagiHumanSR540pI2VConfig dataclass 
MagiHumanSR540pI2VConfig(model_path: str = '', pipeline_config_path: str | None = None, embedded_cfg_scale: float = 6.0, flow_shift: float | None = 5.0, flow_shift_sr: float | None = None, disable_autocast: bool = False, is_causal: bool = False, dit_config: DiTConfig = MagiHumanVideoConfig(), dit_precision: str = 'bf16', upsampler_config: UpsamplerConfig = UpsamplerConfig(), upsampler_precision: str = 'fp32', vae_config: VAEConfig = WanVAEConfig(), vae_precision: str = 'fp32', vae_tiling: bool = False, vae_sp: bool = False, image_encoder_config: EncoderConfig = EncoderConfig(), image_encoder_precision: str = 'fp32', text_encoder_configs: tuple[EncoderConfig, ...] = (lambda: (T5GemmaEncoderConfig(),))(), text_encoder_precisions: tuple[str, ...] = (lambda: ('bf16',))(), preprocess_text_funcs: tuple[Callable[[str], str], ...] = (lambda: (preprocess_text,))(), postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], Tensor], ...] = (lambda: (t5gemma_postprocess_text,))(), dmd_denoising_steps: list[int] | None = None, ti2v_task: bool = False, boundary_ratio: float | None = None, audio_vae_config: VAEConfig = OobleckVAEConfig(), precision: str = 'bf16', t5_gemma_target_length: int = 640, fps: int = 25, num_inference_steps: int = 32, video_txt_guidance_scale: float = 5.0, audio_txt_guidance_scale: float = 5.0, cfg_number: int = 2, vae_stride: tuple[int, int, int] = (4, 16, 16), z_dim: int = 48, frame_receptive_field: int = 11, coords_style: str = 'v2', ref_audio_offset: int = 1000, text_offset: int = 0, video_guidance_high_t_threshold: int = 500, video_guidance_low_t_value: float = 2.0, noise_value: int = 220, sr_audio_noise_scale: float = 0.7, 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, sr_height: int = 512, sr_width: int = 896, sr_local_attn_layers: tuple[int, ...] = (), image_conditioning: bool = True)

Bases: MagiHumanSR540pConfig

Two-stage MagiHuman base + SR-540p text+image-to-AV config.

Functions

fastvideo.pipelines.basic.magi_human.pipeline_configs.t5gemma_postprocess_text 
t5gemma_postprocess_text(outputs: BaseEncoderOutput) -> Tensor

Return per-prompt last_hidden_state as a batched [B, L, D] tensor.

MagiHuman pads/trims the embedding to a fixed length in its own pad_or_trim helper at pipeline time. Here we simply hand through whatever the tokenizer produced — the latent-prep stage is responsible for pad/trim so that the original context length can be preserved.

Source code in fastvideo/pipelines/basic/magi_human/pipeline_configs.py 
21
22
23
24
25
26
27
28
29
30
31
32
33
def t5gemma_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
    """Return per-prompt last_hidden_state as a batched [B, L, D] tensor.

    MagiHuman pads/trims the embedding to a fixed length in its own
    `pad_or_trim` helper at pipeline time. Here we simply hand through
    whatever the tokenizer produced — the latent-prep stage is responsible
    for pad/trim so that the original context length can be preserved.
    """
    hidden = outputs.last_hidden_state
    assert torch.isnan(hidden).sum() == 0
    # Keep the shape the tokenizer emitted; the pipeline stage handles
    # pad-or-trim to t5_gemma_target_length=640.
    return hidden

fastvideo.pipelines.basic.magi_human.presets

Presets for the daVinci-MagiHuman pipelines.

Classes

fastvideo.pipelines.basic.magi_human.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