STABILITY / TWO-STAGE LATENT DIFFUSION / 2023

Stable Diffusion XL 1.0

TOTAL≈3.5B baseparameters

ACTIVE≈3.5B / stepper denoise step

DEPTHU-Net + refinerrepeated blocks

HIDDEN320→1,280main trunk

CONTEXT1024²max canvas

CHECKPOINT≈6.94GB base1 shards · CreativeML Open RAIL++-M

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

1024² · 50 FLOW STEPS · 4.40P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT1024²IMAGE CANVAS

TRANSFORMER STACKBASE U-NET → REFINERcomposition first · high-frequency detail second

Multi-Scale U-Net Skips

OUTPUT1024²DENOISE STEPS

wraps every repeated block

CROSS-ATTENTION U-NET · MICRO-CONDITIONING训练与推理显式输入原图尺寸、裁剪与目标尺寸，降低错误构图。

Base + Refiner SplitRefiner 只处理低噪声阶段，高频细节不必由 base 全程承担。

BASE≈3.5Bsystem

REFINERoptionalsecond model

TEXT2CLIP encoders

LATENT128²@1024

TARGET1024²pixels

STAGES2base/refiner