RASS
A rank-aware speculative sampling method for diffusion models that sorts draft candidates by mean-shift ranking and optimizes selection weights. This approach reduces the number of target model evaluations while preserving exact sampling, thereby accelerating inference.
Timeline (last 90 days)
The paper 'Rank-Aware Speculative Sampling for Diffusion Draft Trees' (RASS) was published on arXiv, proposing an improved method for speculative sampling draft trees in diffusion models.
RASS achieves approximately 20% speedup on CIFAR-10 under matched computational budget conditions.
RASS outperforms D-GRS (Diffusion Greedy Rejection Sampling) across all evaluated settings.
The core mechanism of RASS is based on rank-aware list coupling, comprising three steps: sorting candidates by proposal-target mean shift, sampling ranks with optimized weights, and maximum coupling with residual correction.
The optimization objective of RASS's ranking weights is to minimize the total variation distance between the selected proposal distribution and the target distribution.
RASS guarantees that output samples strictly follow the target distribution through maximum coupling and residual correction, achieving exact sampling without introducing bias.
The paper evaluated RASS on four types of tasks: Gaussian mixture target distributions, FFHQ unconditional pixel-space generation, CIFAR-10 conditional generation, and latent-space diffusion based on Stable Diffusion 3.5.
The RASS evaluation uses COCO2014 prompts interfaced with the Stable Diffusion 3.5 latent-space diffusion model.
RASS is an acceleration method that modifies only the verification rules without modifying the model, and can be applied on top of already-trained large diffusion models without retraining.
All Facts (9)
RASS is an acceleration method that modifies only the verification rules without modifying the model, and can be applied on top of already-trained large diffusion models without retraining.
50%UnverifiedRASS achieves approximately 20% speedup on CIFAR-10 under matched computational budget conditions.
50%UnverifiedRASS outperforms D-GRS (Diffusion Greedy Rejection Sampling) across all evaluated settings.
50%UnverifiedThe core mechanism of RASS is based on rank-aware list coupling, comprising three steps: sorting candidates by proposal-target mean shift, sampling ranks with optimized weights, and maximum coupling with residual correction.
50%UnverifiedThe optimization objective of RASS's ranking weights is to minimize the total variation distance between the selected proposal distribution and the target distribution.
50%UnverifiedRASS guarantees that output samples strictly follow the target distribution through maximum coupling and residual correction, achieving exact sampling without introducing bias.
50%UnverifiedThe paper evaluated RASS on four types of tasks: Gaussian mixture target distributions, FFHQ unconditional pixel-space generation, CIFAR-10 conditional generation, and latent-space diffusion based on Stable Diffusion 3.5.
50%UnverifiedThe RASS evaluation uses COCO2014 prompts interfaced with the Stable Diffusion 3.5 latent-space diffusion model.
50%UnverifiedThe paper 'Rank-Aware Speculative Sampling for Diffusion Draft Trees' (RASS) was published on arXiv, proposing an improved method for speculative sampling draft trees in diffusion models.
50%