A new sampling trick lets AI diffusion models branch their guesses into trees instead of a single line, and generate faster because of it.
Diffusion models create images and other data by repeatedly refining random noise, a process that requires many expensive passes through a large neural network. Speculative sampling speeds this up by drafting cheap, rough guesses and then correcting them against the real model, using a mathematical coupling that still guarantees exact results. Previous versions of this trick, including ones based on reflection maximal coupling, could only draft guesses in a single chain, one after another, which limits how many get accepted per expensive check. This paper draws a link between speculative sampling and relative entropy coding, a compression technique, and uses it to justify draft trees: multiple branching guesses per round, evaluated with a coupling method called greedy rejection sampling.
The result matters because diffusion models are slow for a structural reason, not a hardware one: they need many sequential evaluations of a big network, and most fixes attack that by cutting steps at the cost of quality. This method instead squeezes more useful information out of each step by widening what gets guessed, without weakening the guarantee that outputs still match the real model's distribution.
The reported gain, up to 8.3% faster than the reflection coupling baseline, is a solid engineering result and a modest one; it will not make diffusion models fast, just measurably less slow.