official Stability AI SD3 paper, SD3.5 model card and configs
STABILITY / MMDIT / 2024
Stable Diffusion 3.5 Large
TOTAL8.1B MMDiTparameters
ACTIVE8.1B / stepper denoise step
DEPTH38repeated blocks
HIDDEN2,432main trunk
CONTEXT1024²max canvas
CHECKPOINT≈16.5GB core2 shards · Stability Community
INPUT1024²IMAGE CANVAS
TRANSFORMER STACK38 MMDIT BLOCKSseparate image/text parameters · shared joint attention
Dual-Stream Residual
++
OUTPUT1024²DENOISE STEPS
wraps every repeated block
JOINT ATTENTION · MMDIT它不是简单 concat 后完全共享参数,而是在融合与专用容量之间取中间点。
8.1B Diffusion Transformer大部分计算在每一步重复执行的 MMDiT 主干。
DEPTH38MMDiT
CORE8.1Bdenoiser
TEXT3encoders
FUSIONjointattention
TARGET1024²pixels
LATENT16chVAE