TENCENT / DUAL-TO-SINGLE VIDEO / 2024

HunyuanVideo · 13B

TOTAL≈13Bparameters

ACTIVE≈13B / stepper denoise step

DEPTH20 + 40repeated blocks

HIDDEN3,072main trunk

CONTEXT720P videolatent canvas

CHECKPOINT≈26GB4 shards · Tencent Hunyuan Community

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

704² · 129 FRAMES · 50 STEPS · 237.54P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT704²VIDEO CANVAS

TRANSFORMER STACK20 DOUBLE + 40 SINGLEmodality-specific streams → unified video/text stream

Modulated Flow Residual

OUTPUT129FFRAMES

wraps every repeated block

FULL SPATIOTEMPORAL ATTENTION · DUAL-TO-SINGLE FUSION融合时机显式编码在深度结构中，而不是所有层都做相同操作。

13B Video Transformer视频生成主干在每个采样步处理完整时空 latent。

DOUBLE20Ltwo streams

SINGLE40Ljoint

HIDDEN3,072main

VAEcausal 3Dcodec

TARGET720Pvideo

PARAMS≈13Bcore