The authors formalize a query model for parallel diffusion sampling where an algorithm can make polynomially many score evaluations per round but must wait for all answers before choosing the next batch of queries.
First polynomial parallel-round lower bounds for diffusion sampling with approximate scores.
Authors prove that even with many parallel queries per round, sequential rounds are unavoidable for certain distributions.
Top University
Yiwen Kou · Yimeng Wang
UCLA
Research Digest··3 min read
Yiwen Kou and Yimeng Wang establish the first polynomial lower bounds on the number of sequential rounds required for diffusion sampling when only approximate scores are available.
Why this paper
From UCLA
In one line
Sequential rounds of diffusion sampling cannot be reduced below polynomial in dimension even with unlimited parallel score queries.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§