SD3 Transformer Model
The Transformer model introduced in Stable Diffusion 3. Its novelty lies in the MMDiT transformer block.
SD3Transformer2DModel
class diffusers.SD3Transformer2DModel
< source >( sample_size: int = 128 patch_size: int = 2 in_channels: int = 16 num_layers: int = 18 attention_head_dim: int = 64 num_attention_heads: int = 18 joint_attention_dim: int = 4096 caption_projection_dim: int = 1152 pooled_projection_dim: int = 2048 out_channels: int = 16 pos_embed_max_size: int = 96 )
Parameters
- sample_size (
int
) — The width of the latent images. This is fixed during training since it is used to learn a number of position embeddings. - patch_size (
int
) — Patch size to turn the input data into small patches. - in_channels (
int
, optional, defaults to 16) — The number of channels in the input. - num_layers (
int
, optional, defaults to 18) — The number of layers of Transformer blocks to use. - attention_head_dim (
int
, optional, defaults to 64) — The number of channels in each head. - num_attention_heads (
int
, optional, defaults to 18) — The number of heads to use for multi-head attention. - cross_attention_dim (
int
, optional) — The number ofencoder_hidden_states
dimensions to use. - caption_projection_dim (
int
) — Number of dimensions to use when projecting theencoder_hidden_states
. - pooled_projection_dim (
int
) — Number of dimensions to use when projecting thepooled_projections
. - out_channels (
int
, defaults to 16) — Number of output channels.
The Transformer model introduced in Stable Diffusion 3.
Reference: https://arxiv.org/abs/2403.03206
enable_forward_chunking
< source >( chunk_size: Optional = None dim: int = 0 )
Parameters
- chunk_size (
int
, optional) — The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually over each tensor of dim=dim
. - dim (
int
, optional, defaults to0
) — The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch) or dim=1 (sequence length).
Sets the attention processor to use feed forward chunking.
forward
< source >( hidden_states: FloatTensor encoder_hidden_states: FloatTensor = None pooled_projections: FloatTensor = None timestep: LongTensor = None block_controlnet_hidden_states: List = None joint_attention_kwargs: Optional = None return_dict: bool = True )
Parameters
- hidden_states (
torch.FloatTensor
of shape(batch size, channel, height, width)
) — Inputhidden_states
. - encoder_hidden_states (
torch.FloatTensor
of shape(batch size, sequence_len, embed_dims)
) — Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. - pooled_projections (
torch.FloatTensor
of shape(batch_size, projection_dim)
) — Embeddings projected from the embeddings of input conditions. - timestep (
torch.LongTensor
) — Used to indicate denoising step. block_controlnet_hidden_states — (list
oftorch.Tensor
): A list of tensors that if specified are added to the residuals of transformer blocks. - joint_attention_kwargs (
dict
, optional) — A kwargs dictionary that if specified is passed along to theAttentionProcessor
as defined underself.processor
in diffusers.models.attention_processor. - return_dict (
bool
, optional, defaults toTrue
) — Whether or not to return a~models.transformer_2d.Transformer2DModelOutput
instead of a plain tuple.
The SD3Transformer2DModel forward method.
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
This API is 🧪 experimental.
set_attn_processor
< source >( processor: Union )
Parameters
- processor (
dict
ofAttentionProcessor
or onlyAttentionProcessor
) — The instantiated processor class or a dictionary of processor classes that will be set as the processor for allAttention
layers.If
processor
is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainable attention processors.
Sets the attention processor to use to compute attention.
Disables the fused QKV projection if enabled.
This API is 🧪 experimental.