DIT / TRANSFORMER DIFFUSION BASELINE / 2022

DiT-XL/2

TOTAL675Mparameters

ACTIVE675Mper denoise step

DEPTH28repeated blocks

HIDDEN1,152main trunk

CONTEXT256² / 512²max canvas

CHECKPOINT≈2.7GB1 shards · CC-BY-NC

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

1024² · 50 FLOW STEPS · 385.15T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT1024²IMAGE CANVAS

TRANSFORMER STACK28 DIT BLOCKSstandard self-attention Transformer replaces U-Net

AdaLN-Zero Residual

OUTPUT1024²DENOISE STEPS

wraps every repeated block

FULL SELF-ATTENTION · COMPUTE–QUALITY SCALING论文直接展示更深、更宽或更多 token 的 DiT 有更低 FID。

AdaLN-Zero Blocks条件通过 scale、shift 与 residual gate 注入每个 Transformer block。

DEPTH28Transformer

HIDDEN1,152XL

HEADS16self-attn

PATCH2latent

FLOPS525G@512 pass

PARAMS675MDiT