META / PROMPTABLE VIDEO SEGMENTATION / 2024

SAM 2.1 · Hiera Large

TOTAL224Mparameters

ACTIVE224Mper token

DEPTHHiera + 4 memoryrepeated blocks

HIDDEN256 memorymain trunk

CONTEXTstreaming framesnative tokens

CHECKPOINT≈898MB1 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

3,096 INPUT + 1,000 OUTPUT · 1.87T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT3.1KIMAGE · VIDEO + PROMPT → MASK

TRANSFORMER STACKFRAME ENCODER + MEMORYcurrent frame attends to stored mask memories

Hierarchical Residual

OUTPUT1.0KTOKENS

wraps every repeated block

MEMORY ATTENTION · STREAMING MEMORY BANK每帧预测可编码回 memory，后续帧无需重跑完整历史视频。

Prompt + Mask Decoder用户提示被编码为稀疏/稠密条件并与图像特征解码成 mask。

BACKBONEHiera-Lvision

MEM ATTN4Ltransformer

MEM DIM256hidden

PROMPTS3points/box/mask

MODEstreamvideo

PARAMS224Mcheckpoint