LTXVideoTransformer3DModel

A Diffusion Transformer model for 3D data from LTX was introduced by Lightricks.

The model can be loaded with the following code snippet.

from mindone.diffusers import LTXVideoTransformer3DModel
import mindspore as ms

transformer = LTXVideoTransformer3DModel.from_pretrained("Lightricks/LTX-Video", subfolder="transformer", mindspore_dtype=ms.bfloat16)

mindone.diffusers.LTXVideoTransformer3DModel

Bases: ModelMixin, ConfigMixin, AttentionMixin, FromOriginalModelMixin, PeftAdapterMixin, CacheMixin

A Transformer model for video-like data used in LTX.

PARAMETER	DESCRIPTION
in_channels	The number of channels in the input. TYPE: `int`, defaults to `128` DEFAULT: 128
out_channels	The number of channels in the output. TYPE: `int`, defaults to `128` DEFAULT: 128
patch_size	The size of the spatial patches to use in the patch embedding layer. TYPE: `int`, defaults to `1` DEFAULT: 1
patch_size_t	The size of the tmeporal patches to use in the patch embedding layer. TYPE: `int`, defaults to `1` DEFAULT: 1
num_attention_heads	The number of heads to use for multi-head attention. TYPE: `int`, defaults to `32` DEFAULT: 32
attention_head_dim	The number of channels in each head. TYPE: `int`, defaults to `64` DEFAULT: 64
cross_attention_dim	The number of channels for cross attention heads. TYPE: `int`, defaults to `2048 ` DEFAULT: 2048
num_layers	The number of layers of Transformer blocks to use. TYPE: `int`, defaults to `28` DEFAULT: 28
activation_fn	Activation function to use in feed-forward. TYPE: `str`, defaults to `"gelu-approximate"` DEFAULT: 'gelu-approximate'
qk_norm	The normalization layer to use. TYPE: `str`, defaults to `"rms_norm_across_heads"` DEFAULT: 'rms_norm_across_heads'

Source code in mindone/diffusers/models/transformers/transformer_ltx.py

class LTXVideoTransformer3DModel(
    ModelMixin, ConfigMixin, AttentionMixin, FromOriginalModelMixin, PeftAdapterMixin, CacheMixin
):
    r"""
    A Transformer model for video-like data used in [LTX](https://huggingface.co/Lightricks/LTX-Video).

    Args:
        in_channels (`int`, defaults to `128`):
            The number of channels in the input.
        out_channels (`int`, defaults to `128`):
            The number of channels in the output.
        patch_size (`int`, defaults to `1`):
            The size of the spatial patches to use in the patch embedding layer.
        patch_size_t (`int`, defaults to `1`):
            The size of the tmeporal patches to use in the patch embedding layer.
        num_attention_heads (`int`, defaults to `32`):
            The number of heads to use for multi-head attention.
        attention_head_dim (`int`, defaults to `64`):
            The number of channels in each head.
        cross_attention_dim (`int`, defaults to `2048 `):
            The number of channels for cross attention heads.
        num_layers (`int`, defaults to `28`):
            The number of layers of Transformer blocks to use.
        activation_fn (`str`, defaults to `"gelu-approximate"`):
            Activation function to use in feed-forward.
        qk_norm (`str`, defaults to `"rms_norm_across_heads"`):
            The normalization layer to use.
    """

    _supports_gradient_checkpointing = True
    _skip_layerwise_casting_patterns = ["norm"]
    _repeated_blocks = ["LTXVideoTransformerBlock"]

    @register_to_config
    def __init__(
        self,
        in_channels: int = 128,
        out_channels: int = 128,
        patch_size: int = 1,
        patch_size_t: int = 1,
        num_attention_heads: int = 32,
        attention_head_dim: int = 64,
        cross_attention_dim: int = 2048,
        num_layers: int = 28,
        activation_fn: str = "gelu-approximate",
        qk_norm: str = "rms_norm_across_heads",
        norm_elementwise_affine: bool = False,
        norm_eps: float = 1e-6,
        caption_channels: int = 4096,
        attention_bias: bool = True,
        attention_out_bias: bool = True,
    ) -> None:
        super().__init__()

        out_channels = out_channels or in_channels
        inner_dim = num_attention_heads * attention_head_dim

        self.proj_in = mint.nn.Linear(in_channels, inner_dim)

        self.scale_shift_table = ms.Parameter(mint.randn(2, inner_dim) / inner_dim**0.5, name="scale_shift_table")
        self.time_embed = AdaLayerNormSingle(inner_dim, use_additional_conditions=False)

        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)

        self.rope = LTXVideoRotaryPosEmbed(
            dim=inner_dim,
            base_num_frames=20,
            base_height=2048,
            base_width=2048,
            patch_size=patch_size,
            patch_size_t=patch_size_t,
            theta=10000.0,
        )

        self.transformer_blocks = nn.CellList(
            [
                LTXVideoTransformerBlock(
                    dim=inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                    cross_attention_dim=cross_attention_dim,
                    qk_norm=qk_norm,
                    activation_fn=activation_fn,
                    attention_bias=attention_bias,
                    attention_out_bias=attention_out_bias,
                    eps=norm_eps,
                    elementwise_affine=norm_elementwise_affine,
                )
                for _ in range(num_layers)
            ]
        )

        self.norm_out = mint.nn.LayerNorm(inner_dim, eps=1e-6)
        self.proj_out = mint.nn.Linear(inner_dim, out_channels)

        self.gradient_checkpointing = False

    def construct(
        self,
        hidden_states: ms.Tensor,
        encoder_hidden_states: ms.Tensor,
        timestep: ms.Tensor,
        encoder_attention_mask: ms.Tensor,
        num_frames: Optional[int] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        rope_interpolation_scale: Optional[Union[Tuple[float, float, float], ms.Tensor]] = None,
        video_coords: Optional[ms.Tensor] = None,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        return_dict: bool = False,
    ) -> ms.Tensor:
        if attention_kwargs is not None:
            attention_kwargs = attention_kwargs.copy()

        image_rotary_emb = self.rope(hidden_states, num_frames, height, width, rope_interpolation_scale, video_coords)

        # convert encoder_attention_mask to a bias the same way we do for attention_mask
        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)

        batch_size = hidden_states.shape[0]
        hidden_states = self.proj_in(hidden_states)

        temb, embedded_timestep = self.time_embed(
            timestep.flatten(),
            batch_size=batch_size,
            hidden_dtype=hidden_states.dtype,
        )

        temb = temb.view(batch_size, -1, temb.shape[-1])
        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1])

        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])

        for block in self.transformer_blocks:
            hidden_states = block(
                hidden_states=hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                temb=temb,
                image_rotary_emb=image_rotary_emb,
                encoder_attention_mask=encoder_attention_mask,
            )

        scale_shift_values = self.scale_shift_table[None, None] + embedded_timestep[:, :, None]
        shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]

        hidden_states = self.norm_out(hidden_states)
        hidden_states = hidden_states * (1 + scale) + shift
        output = self.proj_out(hidden_states)

        if not return_dict:
            return (output,)
        return Transformer2DModelOutput(sample=output)

mindone.diffusers.models.modeling_outputs.Transformer2DModelOutput dataclass

Bases: BaseOutput

The output of [Transformer2DModel].

PARAMETER	DESCRIPTION
`	The hidden states output conditioned on the encoder_hidden_states input. If discrete, returns probability distributions for the unnoised latent pixels. TYPE: batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete

Source code in mindone/diffusers/models/modeling_outputs.py

@dataclass
class Transformer2DModelOutput(BaseOutput):
    """
    The output of [`Transformer2DModel`].

    Args:
        sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` or
        `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
            distributions for the unnoised latent pixels.
    """

    sample: "ms.Tensor"  # noqa: F821