A popular trick for speeding up AI image and video generators may not be doing what its name promises.
Diffusion Transformers, the architecture behind tools like FLUX and PixArt, are slow because they recompute the same internal features at every generation step. Feature caching methods try to fix that by reusing calculations from previous steps, and a newer variant called token-wise caching goes further, letting the model skip so-called unimportant tokens while still recalculating the important ones. Researchers tested that assumption directly and found the selection of important tokens is often ineffective, sometimes performing worse than picking tokens at random. In response, they built DuCa, a dual caching method that alternates between aggressive and conservative caching and picks which tokens to recompute randomly rather than by importance score.
Generation speed is a real cost line for anyone running image or video models at scale, so any credible speedup matters. More interesting is what this says about a common assumption in efficiency research: that models can reliably flag which of their own computations matter most. Tested across DiT, PixArt, FLUX, and OpenSora, DuCa outperformed the token-wise caching it replaced, at least by the paper's own benchmarks.
Turns out a coin flip can out-engineer a heuristic, which is a humbling result for anyone who assumed their model knew what it was doing.