LATENT DIFFUSION / U-NET BASELINE / 2022

Stable Diffusion 1.5

TOTAL≈1.07Bparameters

ACTIVE≈1.07Bper denoise step

DEPTHU-Net 4 scalesrepeated blocks

HIDDEN320→1,280main trunk

CONTEXT512²max canvas

CHECKPOINT≈4.27GB1 shards · CreativeML Open RAIL-M

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

1024² · 50 FLOW STEPS · 1.44P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT1024²IMAGE CANVAS

TRANSFORMER STACKMULTI-SCALE U-NETdownsample → bottleneck → upsample with skip connections

U-Net Skip Connections

OUTPUT1024²DENOISE STEPS

wraps every repeated block

SPATIAL + CROSS ATTENTION · LATENT-SPACE DIFFUSION扩散不在 512² RGB 像素上运行，VAE 将空间 token 降低 64 倍。

860M U-Net同一 U-Net 在每个去噪步重复执行，是总 FLOPs 的主项。

U-NET860Mbackbone

TEXT123MCLIP

VAE8×spatial

LATENT64²×4@512

STEPS50reference

PARAMS≈1.07Bsystem