OPENAI / CONTRASTIVE DUAL ENCODER / 2021

CLIP ViT-L/14

TOTAL≈428Mparameters

ACTIVE≈428Mper token

DEPTH24 + 12repeated blocks

HIDDEN1,024 / 768main trunk

CONTEXT257 patches / 77 textnative tokens

CHECKPOINT≈1.7GB1 shards · MIT code · model card terms

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

1 INPUT + 333 OUTPUT · 294.15G arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT1IMAGE ↔ TEXT EMBEDDING

TRANSFORMER STACK24L VISION + 12L TEXTindependent encoders meet only at normalized embeddings

Independent Residual Towers

OUTPUT333TOKENS

wraps every repeated block

BIDIRECTIONAL SELF-ATTENTION · CONTRASTIVE ALIGNMENT训练目标把匹配图文拉近、非匹配拉远，无需生成 decoder。

Projection Heads两个塔的最终表示归一化后做温度缩放的点积。

VISION24L1,024

TEXT12L768

PATCH14pixels

TEXT LEN77tokens

FUSIONnonedual tower

PARAMS≈428MViT-L/14