NVIDIA / LINEAR DIT / 2025

SANA 1.5 · 4.8B

TOTAL4.8Bparameters

ACTIVE4.8B / stepper denoise step

DEPTH40repeated blocks

HIDDEN2,240main trunk

CONTEXT1024² / 4Kmax canvas

CHECKPOINT≈9.6GB2 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

1024² · 50 FLOW STEPS · 638.92T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT1024²IMAGE CANVAS

TRANSFORMER STACK40 LINEAR DIT BLOCKSlinear attention + convolutional Mix-FFN

AdaLN Residual

OUTPUT1024²DENOISE STEPS

wraps every repeated block

LINEAR ATTENTION · 32× DEEP COMPRESSION比常见 8× VAE 再减少 16 倍 latent 空间位置，是效率的第一来源。

Linear DiT + Mix-FFN局部卷积补足线性 attention 的空间归纳偏置。

DEPTH40Linear DiT

HIDDEN2,2404.8B

COMPRESSION32×DC-AE

ATTENTIONlinearO(n)

TARGET4Ksupported

PARAMS4.8Blargest