main
official SWivid/F5-TTS repository, F5-TTS paper and v1 base config
F5 / FLOW-MATCHING TTS / 2024

F5-TTS v1 Base

TOTAL≈336Mparameters
ACTIVE≈336M / stepruntime path
DEPTH22repeated blocks
HIDDEN1,024main trunk
CONTEXTutterancestream / sequence
CHECKPOINT≈1.35GB1 shards · MIT
INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

F5-TTS v1 Base · Architecture Explorer

128 tokens → 10 seconds · 20.33T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER
INPUT128TOKENS
TRANSFORMER STACK22 FLOW DIT BLOCKSmasked mel infilling conditioned on text and reference audio
AdaLN DiT Residual
++
OUTPUT10SECONDS
wraps every repeated block
FULL MEL ATTENTION · NO EXPLICIT ALIGNMENT时长与对齐由条件 flow 内部学习,拓扑比传统 TTS pipeline 更短。
Conditional Flow Matcher同一 DiT 在 ODE 采样中重复调用,NFE 是最关键的成本旋钮。
DEPTH22DiT
HIDDEN1,024main
HEADS16attention
MEL100bins
NFE32default
PARAMS≈336Mbase