Rearranging residual math turns DiT into U-Net-style model, 2.3x faster

ostrisai · x · 2026-09-29

Developer @LodestoneRock shows that with residual math rearranging, you can convert a residual model like DiT — or any transformer — into a more efficient U-Net-style architecture where the middle layers process only 1/4 of tokens, yielding 2.3x faster inference than the base model. With quick calibration, most of the original output quality is restored.

Original post →

More from Research

Research channel →