ConceptDiffusion Greedy Rejection Sampling
D-GRS
A tree-structured drafting method for speculative sampling in diffusion models. At each node it generates multiple conditionally independent candidates and verifies them sequentially in generation order. It serves as a baseline for comparison with the RASS method.
Timeline (last 90 days)
Oct 5
RASS outperforms D-GRS (Diffusion Greedy Rejection Sampling) across all evaluated settings.
Unverified50%
Oct 5
D-GRS generates K conditionally independent candidates at each node and tests them sequentially in generation order, with the first candidate that passes the test being selected.
Unverified50%