Cross-Attention

What It Is

An attention mechanism where queries come from one sequence (the target) and keys/values come from a different sequence (the source) — allowing one modality or representation to selectively attend to information from another.

Why It Matters

Cross-attention is how conditioning works in generative models. In Latent Diffusion Models, the denoising U-Net queries the noisy image latent while attending to a CLIP text embedding — this is how text prompts guide image generation. In the original Transformer, the decoder uses cross-attention to attend to encoder outputs during translation.

How It Works

Given queries $Q$ from the target sequence and keys/values $K,V$ from the source sequence:

$\text{CrossAttn}(Q,K,V)=\text{softmax}\left(\frac{Q{K}^{\top }}{\sqrt{d_k}}\right)V$

This is identical to self-attention in form, but $Q$ and $K,V$ are projected from different inputs. In LDMs, $Q=W_Q\cdot z_t$ (noisy latent at timestep $t$) and $K,V=W_{K/V}\cdot {\tau }_{\theta }(y)$ (projected text embedding). Each spatial position in the image can attend to any token in the text prompt.

Key Sources

• latent-diffusion-models-high-resolution-image-synthesis

• attention-is-all-you-need

• bert-pre-training-of-deep-bidirectional-transformers-for-language-understanding

• flash-attention-2

• flash-attention-fast-and-memory-efficient-exact-attention

• gqa-grouped-query-attention

• mistral-7b

• numina-counting-text-to-video

• rope-rotary-position-embedding

Related Concepts

• attention
• transformer
• latent-space