Skip to content

vllm_omni.diffusion.models.ltx2.vae.decoder

LTX-2.5 Native-weight adapter for Diffusers' diffusion VAE decoder.

LTX25_NATIVE_ARTIFACT_REVISION module-attribute 

LTX25_NATIVE_ARTIFACT_REVISION = (
    "8a4ff96f581e72bedc1b44367581c49d544a05f1"
)

LTX25_NATIVE_DIFFUSION_DECODER_FILENAME module-attribute 

LTX25_NATIVE_DIFFUSION_DECODER_FILENAME = (
    "vae/ltx-2.5-video-vae-bf16.safetensors"
)

LTX25_NATIVE_DIFFUSION_DECODER_REPO_ID module-attribute 

LTX25_NATIVE_DIFFUSION_DECODER_REPO_ID = (
    "Lightricks/LTX-2.5"
)

LTX2VideoDiffusionDecoder3d

Bases: LTX2VideoDiffusionDecoder3d

Diffusers decoder core with the short-clip NATTEN context fix.

forward_stage_4 

forward_stage_4(
    hidden_states: Tensor,
    drop_leading_frame: bool = True,
    crop_trailing_ghost: bool = True,
) -> Tensor

LTX2VideoDiffusionDecoderModel

Bases: LTX2VideoDiffusionDecoderModel

Diffusers 0.40 decoder plus canonical Native loading and short-clip support.

decode 

decode(
    z: Tensor,
    generator: Generator | None = None,
    num_inference_steps: int | None = None,
    return_dict: bool = True,
) -> DecoderOutput | tuple[Tensor]

from_ltx25_native_checkpoint classmethod 

from_ltx25_native_checkpoint(
    checkpoint_path: str,
    config: Mapping[str, Any],
    dtype: dtype,
) -> LTX2VideoDiffusionDecoderModel

Construct this decoder class directly from the canonical Native checkpoint.

LTX2VideoVaeDiffusionNABlock

Bases: LTX2VideoVaeDiffusionNABlock

forward 

forward(
    hidden_states: Tensor,
    latent_context: Tensor,
    modulation: tuple[Tensor, ...],
    block_mask=None,
) -> Tensor

LTX2VideoVaeNeighborhoodAttention

Bases: LTX2VideoVaeNeighborhoodAttention

project_qkv 

project_qkv(
    hidden_states: Tensor,
) -> tuple[Tensor, Tensor, Tensor]

LTX2VideoVaeNeighborhoodNattenProcessor

Bases: LTX2VideoVaeNeighborhoodNattenProcessor

NATTEN dispatch with Omni FNA for eligible stage-5 grids.

LTX2VideoVaeSwiGLU

Bases: LTX2VideoVaeSwiGLU

forward 

forward(hidden_states: Tensor) -> Tensor

convert_ltx25_native_diffusion_decoder_state_dict 

convert_ltx25_native_diffusion_decoder_state_dict(
    native_state_dict: Mapping[str, Tensor],
) -> dict[str, Tensor]

Convert the canonical Native LTX-2.5 DiffVAE state dict to Diffusers.

load_ltx25_native_diffusion_decoder_state_dict 

load_ltx25_native_diffusion_decoder_state_dict(
    path: str,
) -> dict[str, Tensor]

Read and convert only DiffVAE tensors from the canonical Native VAE file.