CosmosTransformer3DModel

A Diffusion Transformer model for 3D video-like data was introduced in Cosmos World Foundation Model Platform for Physical AI by NVIDIA.

The model can be loaded with the following code snippet.

from diffusers import CosmosTransformer3DModel

transformer = CosmosTransformer3DModel.from_pretrained("nvidia/Cosmos-1.0-Diffusion-7B-Text2World", subfolder="transformer", torch_dtype=torch.bfloat16)

CosmosTransformer3DModel

class diffusers.CosmosTransformer3DModel

< source >

( in_channels: int = 16 out_channels: int = 16 num_attention_heads: int = 32 attention_head_dim: int = 128 num_layers: int = 28 mlp_ratio: float = 4.0 text_embed_dim: int = 1024 adaln_lora_dim: int = 256 max_size: typing.Tuple[int, int, int] = (128, 240, 240) patch_size: typing.Tuple[int, int, int] = (1, 2, 2) rope_scale: typing.Tuple[float, float, float] = (2.0, 1.0, 1.0) concat_padding_mask: bool = True extra_pos_embed_type: typing.Optional[str] = 'learnable' )

Parameters

in_channels (int, defaults to 16) — The number of channels in the input.
out_channels (int, defaults to 16) — The number of channels in the output.
num_attention_heads (int, defaults to 32) — The number of heads to use for multi-head attention.
attention_head_dim (int, defaults to 128) — The number of channels in each attention head.
num_layers (int, defaults to 28) — The number of layers of transformer blocks to use.
mlp_ratio (float, defaults to 4.0) — The ratio of the hidden layer size to the input size in the feedforward network.
text_embed_dim (int, defaults to 4096) — Input dimension of text embeddings from the text encoder.
adaln_lora_dim (int, defaults to 256) — The hidden dimension of the Adaptive LayerNorm LoRA layer.
max_size (Tuple[int, int, int], defaults to (128, 240, 240)) — The maximum size of the input latent tensors in the temporal, height, and width dimensions.
patch_size (Tuple[int, int, int], defaults to (1, 2, 2)) — The patch size to use for patchifying the input latent tensors in the temporal, height, and width dimensions.
rope_scale (Tuple[float, float, float], defaults to (2.0, 1.0, 1.0)) — The scaling factor to use for RoPE in the temporal, height, and width dimensions.
concat_padding_mask (bool, defaults to True) — Whether to concatenate the padding mask to the input latent tensors.
extra_pos_embed_type (str, optional, defaults to learnable) — The type of extra positional embeddings to use. Can be one of None or learnable.

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

Transformer2DModelOutput

class diffusers.models.modeling_outputs.Transformer2DModelOutput

< source >

( sample: torch.Tensor )

Parameters

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.

The output of Transformer2DModel.

< > Update on GitHub

Diffusers

CosmosTransformer3DModel