Research · Updated 16 Oct, 09:30 am IST
VISTA selects sparse SMC guidance schedules to speed discrete diffusion
Why it matters for readers: Explains how selective guidance can keep results strong while cutting the amount of expensive computation.
- VISTA frames sparse guidance as approximating the ideal full-guidance probability path and derives an upper bound on the ℓ1 path deviation to guide selection.1
- The method collects value statistics from a small number of full-guidance warmup generations and then optimizes the empirical surrogate exactly via dynamic programming.1
- After schedule selection, VISTA has the same per-generation guidance complexity as other sparse schedules with the same budget.1
- On tasks across text, biological sequences, and images, VISTA discovers task-dependent schedules and typically outperforms hand-designed sparse schedules, while approaching or exceeding full-guidance results on several metrics.1
Get a brief like this every morning
Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.
Get started