A sequential module that passes timestep embeddings to the children that support it as an extra input.

A residual block that can optionally change the number of channels. :param channels: the number of input channels. :param emb_channels: t

A single linear layer which does a 1x1 conv

Transformer block for image-like data. First, project the input (aka embedding) and reshape to b, t, d. Then apply standard transform

A downsampling layer with an optional convolution. :param channels: channels in the inputs and outputs. :param use_conv: a bool determini

An upsampling layer with an optional convolution. :param channels: channels in the inputs and outputs. :param use_conv: a bool determinin

A module which performs QKV attention and splits in a different order.

Uses the CLIP transformer encoder for text (from huggingface)

Uses the T5 transformer encoder for text

wraps a discretizer, and prunes the sigmas params: strength: float between 0.0 and 1.0. 1.0 means full sampling (all sigmas are retur

wraps a discretizer, and prunes the sigmas params: strength: float between 0.0 and 1.0. 1.0 means full sampling (all sigmas are retur

to match AttnBlock usage

Uses xformers efficient implementation, see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffu

A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping

wraps a discretizer, and prunes the sigmas params: strength: float between 0.0 and 1.0. 0.0 means full sampling (all sigmas are retur

A minimal version of https://github.com/ShieldMnt/invisible-watermark/blob/main/imwatermark/watermark.py to only reconstruct b

This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators, unCLIP models, etc. Hence,

An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted to the N-d case. https

Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py

Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL (we also restore them explicitly as special cases for legacy

embeds each dimension independently and concatenates them