Prior Transformer
The Prior Transformer was originally introduced in Hierarchical Text-Conditional Image Generation with CLIP Latents by Ramesh et al. It is used to predict CLIP image embeddings from CLIP text embeddings; image embeddings are predicted through a denoising diffusion process.
The abstract from the paper is:
Contrastive models like CLIP have been shown to learn robust representations of images that capture both semantics and style. To leverage these representations for image generation, we propose a two-stage model: a prior that generates a CLIP image embedding given a text caption, and a decoder that generates an image conditioned on the image embedding. We show that explicitly generating image representations improves image diversity with minimal loss in photorealism and caption similarity. Our decoders conditioned on image representations can also produce variations of an image that preserve both its semantics and style, while varying the non-essential details absent from the image representation. Moreover, the joint embedding space of CLIP enables language-guided image manipulations in a zero-shot fashion. We use diffusion models for the decoder and experiment with both autoregressive and diffusion models for the prior, finding that the latter are computationally more efficient and produce higher-quality samples.
PriorTransformer
class diffusers.PriorTransformer
< source >( num_attention_heads: int = 32 attention_head_dim: int = 64 num_layers: int = 20 embedding_dim: int = 768 num_embeddings = 77 additional_embeddings = 4 dropout: float = 0.0 time_embed_act_fn: str = 'silu' norm_in_type: Optional = None embedding_proj_norm_type: Optional = None encoder_hid_proj_type: Optional = 'linear' added_emb_type: Optional = 'prd' time_embed_dim: Optional = None embedding_proj_dim: Optional = None clip_embed_dim: Optional = None )
Parameters
- num_attention_heads (
int, optional, defaults to 32) — The number of heads to use for multi-head attention. - attention_head_dim (
int, optional, defaults to 64) — The number of channels in each head. - num_layers (
int, optional, defaults to 20) — The number of layers of Transformer blocks to use. - embedding_dim (
int, optional, defaults to 768) — The dimension of the model inputhidden_states - num_embeddings (
int, optional, defaults to 77) — The number of embeddings of the model inputhidden_states - additional_embeddings (
int, optional, defaults to 4) — The number of additional tokens appended to the projectedhidden_states. The actual length of the usedhidden_statesisnum_embeddings + additional_embeddings. - dropout (
float, optional, defaults to 0.0) — The dropout probability to use. - time_embed_act_fn (
str, optional, defaults to ‘silu’) — The activation function to use to create timestep embeddings. - norm_in_type (
str, optional, defaults to None) — The normalization layer to apply on hidden states before passing to Transformer blocks. Set it toNoneif normalization is not needed. - embedding_proj_norm_type (
str, optional, defaults to None) — The normalization layer to apply on the inputproj_embedding. Set it toNoneif normalization is not needed. - encoder_hid_proj_type (
str, optional, defaults tolinear) — The projection layer to apply on the inputencoder_hidden_states. Set it toNoneifencoder_hidden_statesisNone. - added_emb_type (
str, optional, defaults toprd) — Additional embeddings to condition the model. Choose fromprdorNone. if chooseprd, it will prepend a token indicating the (quantized) dot product between the text embedding and image embedding as proposed in the unclip paper https://arxiv.org/abs/2204.06125 If it isNone, no additional embeddings will be prepended. - time_embed_dim (
int, *optional*, defaults to None) -- The dimension of timestep embeddings. If None, will be set tonum_attention_heads * attention_head_dim` - embedding_proj_dim (
int, optional, default to None) — The dimension ofproj_embedding. If None, will be set toembedding_dim. - clip_embed_dim (
int, optional, default to None) — The dimension of the output. If None, will be set toembedding_dim.
A Prior Transformer model.
forward
< source >( hidden_states timestep: Union proj_embedding: FloatTensor encoder_hidden_states: Optional = None attention_mask: Optional = None return_dict: bool = True ) → ~models.prior_transformer.PriorTransformerOutput or tuple
Parameters
- hidden_states (
torch.FloatTensorof shape(batch_size, embedding_dim)) — The currently predicted image embeddings. - timestep (
torch.LongTensor) — Current denoising step. - proj_embedding (
torch.FloatTensorof shape(batch_size, embedding_dim)) — Projected embedding vector the denoising process is conditioned on. - encoder_hidden_states (
torch.FloatTensorof shape(batch_size, num_embeddings, embedding_dim)) — Hidden states of the text embeddings the denoising process is conditioned on. - attention_mask (
torch.BoolTensorof shape(batch_size, num_embeddings)) — Text mask for the text embeddings. - return_dict (
bool, optional, defaults toTrue) — Whether or not to return a~models.prior_transformer.PriorTransformerOutputinstead of a plain tuple.
Returns
~models.prior_transformer.PriorTransformerOutput or tuple
If return_dict is True, a ~models.prior_transformer.PriorTransformerOutput is returned, otherwise a
tuple is returned where the first element is the sample tensor.
The PriorTransformer forward method.
set_attn_processor
< source >( processor: Union )
Parameters
- processor (
dictofAttentionProcessoror onlyAttentionProcessor) — The instantiated processor class or a dictionary of processor classes that will be set as the processor for allAttentionlayers.If
processoris 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 custom attention processors and sets the default attention implementation.
PriorTransformerOutput
class diffusers.models.transformers.prior_transformer.PriorTransformerOutput
< source >( predicted_image_embedding: FloatTensor )
The output of PriorTransformer.