Core Architecture · Fast-moving · Advanced
Multimodal Rotary Position Embeddings
Also known as: M-RoPE, Multimodal Position Encoding
An extension of RoPE that encodes 1D text positions alongside 2D image spatial coordinates and 3D video temporal frames.
What Multimodal Rotary Position Embeddings is
M-RoPE allows transformers to process visual tokens with explicit positional awareness of spatial grids and frame sequences.
How it works
Decomposes rotary embedding channels into dedicated text, height, width, and temporal rotation components.
Why it matters
Adopted in cutting-edge vision-language architectures like Qwen2-VL.
Common uses
- →Vision-language model position encoding
- →Video understanding transformers
- →Spatial reasoning AI
More in this collection
Browse all AI Concepts