AReA: Adaptive Receptive-Field Attention Cuts Cost for High-Resolution Diffusion Transformers

Researchers have introduced AReA (Adaptive Receptive-Field Attention), an attention mechanism designed to make high-resolution diffusion transformers significantly more efficient. Rather than computing full attention across every token, AReA pools keys and values into spatial blocks — one token per f×f block — so that each query attends to a much smaller set of keys at roughly 1/f² of the standard computational cost, while outputs remain at full resolution.

The approach targets a core bottleneck in diffusion transformer scaling: the quadratic cost of self-attention grows rapidly with image resolution. By preserving full-resolution outputs while compressing the key-value space, AReA offers a path toward more practical high-resolution image and video generation without sacrificing output fidelity.