The data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity
Wainwright proposes a way to certify masking-diffusion schedules against a target KL error.
The paper introduces unmasking growth complexity, a path-based measure meant to track how data geometry affects discretization error. In log-reveal-odds coordinates, it derives optimized single-block and multi-block unmasking schedules. The abstract says the needed increments can be estimated from samples, yielding samplers that are high-probability near-oracle within a constant factor. Examples claim dimension-dependent gains over coarse schedules, including roughly square-root-in-d improvements with a constant number of adaptive blocks. ArXiv · AI/CL/LG's note
The paper introduces unmasking growth complexity, a path-based measure meant to track how data geometry affects discretization error. In log-reveal-odds coordinates, it derives optimized single-block and multi-block unmasking schedules. The abstract says the needed increments can be estimated from samples, yielding samplers that are high-probability near-oracle within a constant factor. Examples claim dimension-dependent gains over coarse schedules, including roughly square-root-in-d improvements with a constant number of adaptive blocks. ArXiv · AI/CL/LG's note
score 5