Skip to content

vllm_omni.diffusion.models.cosmos3.transfer

Cosmos3 transfer inference helpers.

The reference Cosmos Framework transfer path accepts one or more control hints and packs them before the noisy target video. This module keeps the vLLM Omni pipeline-specific pieces small: request parsing, media normalization, and optional on-the-fly edge/blur control generation.

Transfer is a video-output mode. It cannot be combined with T2I, sound generation, or action generation.

BILATERAL_D module-attribute 

BILATERAL_D = 30

BILATERAL_ITERATIONS module-attribute 

BILATERAL_ITERATIONS = 1

BILATERAL_REFERENCE_RESOLUTION module-attribute 

BILATERAL_REFERENCE_RESOLUTION = 720

BILATERAL_SIGMA_COLOR module-attribute 

BILATERAL_SIGMA_COLOR = 150

BILATERAL_SIGMA_SPACE module-attribute 

BILATERAL_SIGMA_SPACE = 100

BLUR_DOWNUP_PRESETS module-attribute 

BLUR_DOWNUP_PRESETS: dict[str, int] = {
    "none": 1,
    "very_low": 4,
    "low": 4,
    "medium": 10,
    "high": 16,
    "very_high": 16,
}

BLUR_PRE_BLUR_DOWNSCALE_PRESETS module-attribute 

BLUR_PRE_BLUR_DOWNSCALE_PRESETS: dict[str, int] = {
    "none": 1,
    "very_low": 1,
    "low": 4,
    "medium": 2,
    "high": 1,
    "very_high": 4,
}

EDGE_PRESETS module-attribute 

EDGE_PRESETS: dict[str, tuple[int, int]] = {
    "none": (20, 50),
    "very_low": (20, 50),
    "low": (50, 100),
    "medium": (100, 200),
    "high": (200, 300),
    "very_high": (300, 400),
}

IMAGE_EXTENSIONS module-attribute 

IMAGE_EXTENSIONS = {
    ".bmp",
    ".gif",
    ".jpg",
    ".jpeg",
    ".png",
    ".tif",
    ".tiff",
    ".webp",
}

TRANSFER_DEFAULTS module-attribute 

TRANSFER_DEFAULTS: dict[str, dict[str, Any]] = {
    "edge": {
        "guidance_scale": 3.0,
        "control_guidance": 1.5,
        "flow_shift": 10.0,
    },
    "blur": {
        "guidance_scale": 3.0,
        "control_guidance": 1.5,
        "flow_shift": 10.0,
    },
    "depth": {
        "guidance_scale": 3.0,
        "control_guidance": 1.5,
        "flow_shift": 10.0,
    },
    "seg": {
        "guidance_scale": 3.0,
        "control_guidance": 2.0,
        "flow_shift": 10.0,
    },
    "wsm": {
        "guidance_scale": 1.0,
        "control_guidance": 3.0,
        "flow_shift": 10.0,
        "num_frames": 101,
        "fps": 10,
        "num_video_frames_per_chunk": 101,
    },
}

TRANSFER_HINT_KEYS module-attribute 

TRANSFER_HINT_KEYS: tuple[str, ...] = (
    "edge",
    "blur",
    "depth",
    "seg",
    "wsm",
)

TRANSFER_SAMPLE_DEFAULTS module-attribute 

TRANSFER_SAMPLE_DEFAULTS: dict[str, Any] = {
    "num_video_frames_per_chunk": 93,
    "num_conditional_frames": 1,
    "max_frames": 5000,
    "show_control_condition": False,
    "show_input": False,
    "num_first_chunk_conditional_frames": 0,
    "share_vision_temporal_positions": True,
}

Cosmos3TransferConfig dataclass

control_guidance class-attribute instance-attribute 

control_guidance: float = 1.0

control_guidance_interval class-attribute instance-attribute 

control_guidance_interval: tuple[float, float] | None = None

flow_shift class-attribute instance-attribute 

flow_shift: float | None = None

fps class-attribute instance-attribute 

fps: float | None = None

guidance_scale class-attribute instance-attribute 

guidance_scale: float | None = None

hints class-attribute instance-attribute 

hints: dict[str, Cosmos3TransferHint] = field(
    default_factory=dict
)

max_frames class-attribute instance-attribute 

max_frames: int = 5000

num_conditional_frames class-attribute instance-attribute 

num_conditional_frames: int = 1

num_first_chunk_conditional_frames class-attribute instance-attribute 

num_first_chunk_conditional_frames: int = 0

num_frames class-attribute instance-attribute 

num_frames: int | None = None

num_video_frames_per_chunk class-attribute instance-attribute 

num_video_frames_per_chunk: int = 93

ordered_hints property 

ordered_hints: list[Cosmos3TransferHint]

share_vision_temporal_positions class-attribute instance-attribute 

share_vision_temporal_positions: bool = True

show_control_condition class-attribute instance-attribute 

show_control_condition: bool = False

show_input class-attribute instance-attribute 

show_input: bool = False

Cosmos3TransferHint dataclass

control class-attribute instance-attribute 

control: Any | None = None

control_path class-attribute instance-attribute 

control_path: str | None = None

key instance-attribute 

key: str

preset_blur_strength class-attribute instance-attribute 

preset_blur_strength: str = 'medium'

preset_edge_threshold class-attribute instance-attribute 

preset_edge_threshold: str = 'medium'

has_transfer_hints 

has_transfer_hints(
    extra_args: Mapping[str, Any] | None,
) -> bool

load_or_compute_control_frames 

load_or_compute_control_frames(
    hint: Cosmos3TransferHint,
    *,
    height: int,
    width: int,
    max_frames: int,
    input_frames: Tensor | None,
) -> Tensor

make_blur_control 

make_blur_control(frames: Tensor, preset: str) -> Tensor

make_edge_control 

make_edge_control(frames: Tensor, preset: str) -> Tensor

media_hw 

media_hw(value: Any) -> tuple[int, int] | None

media_to_uint8_cthw 

media_to_uint8_cthw(
    value: Any,
    *,
    height: int,
    width: int,
    max_frames: int | None = None,
) -> Tensor

normalized_video_to_uint8_cthw 

normalized_video_to_uint8_cthw(video: Tensor) -> Tensor

pad_temporal_frames 

pad_temporal_frames(
    frames: Tensor, target_frames: int
) -> Tensor

resize_center_crop_uint8_cthw 

resize_center_crop_uint8_cthw(
    frames: Tensor, height: int, width: int
) -> Tensor

resolve_transfer_config 

resolve_transfer_config(
    sp: Any, prompt_data: Any = None
) -> Cosmos3TransferConfig | None

transfer_max_frames_from_extra_args 

transfer_max_frames_from_extra_args(
    extra_args: Mapping[str, Any] | None,
) -> int

uint8_cthw_to_normalized_5d 

uint8_cthw_to_normalized_5d(
    frames: Tensor, *, dtype: dtype
) -> Tensor