1、目的

        用transformer来替代U-Net backbone，提升生成效果

2、方法

        Diffusion Transformers (DiTs)

        1）结构

                Latent Diffusion Models (LDMs)

                        -> Transformer (Vision Transformer, ViT) based DDPM

                        -> off-the-shelf convolutional VAE

        2）patchify

                        

                -> converts the spatial input into a sequence of T tokens

                -> linearly embedding each patch in the input to dimension d

                -> ViT frequency-based positional embeddings (the sine-cosine version)

        3）conditional information

                -> In-context conditioning

                        negligible Gflops

                -> Cross-attention block

                        most Gflops

                -> Adaptive layer norm (adaLN) block

                        least Gflops

                -> adaLN-Zero block

                        regress  and 

                        regress dimension-wise scaling parameters  that are applied immediately prior to any residual connections within the DiT block

                        initialize the MLP to output the zero-vector for all 

                        negligible Gflops

        4）Transformer decoder

                linearly decode each token into a 

                rearrange the decoded tokens into their original spatial layout

3、实验

        1）model size ↑ (transformer deeper and wider)，patch size ↓，模型表现↑

        2）Gflops, rather than parameter counts, determines the quality of a DiT model

        3）larger DiT models are more compute-efficient