WAN / VIDEO FLOW DIT / 2025

Wan2.1 T2V · 14B

TOTAL14Bparameters

ACTIVE14B / stepper denoise step

DEPTH40repeated blocks

HIDDEN5,120main trunk

CONTEXT480P / 720Platent canvas

CHECKPOINT≈28GB7 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

704² · 81 FRAMES · 50 STEPS · 127.38P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT704²VIDEO CANVAS

TRANSFORMER STACK40 VIDEO DIT BLOCKS3D latent self-attention + text cross-attention

AdaLN Video Residual

OUTPUT81FFRAMES

wraps every repeated block

SPATIOTEMPORAL ATTENTION · WAN 3D CAUSAL VAE因果解码避免未来帧依赖，使超长视频可分块还原。

14B Flow DiTT5 文本条件通过每层 cross-attention 注入视频 latent。

DEPTH40Video DiT

HIDDEN5,120main

HEADS40attention

FFN13,824dim

TARGET720Psupported

PARAMS14BT2V