# Accelerated Diffusion Models via Speculative Sampling

**Authors:** Valentin De Bortoli, Alexandre Galashov, Arthur Gretton, Arnaud Doucet  
**arXiv:** [2501.05370](https://arxiv.org/abs/2501.05370)  
**Date:** 2025-01-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.05370v2/

## Key points

- Authors: Valentin De Bortoli, Alexandre Galashov, Arthur Gretton, Arnaud Doucet
- Technique: Speculative sampling method for accelerating inference in large language models
- Approach: Generating candidate tokens using a fast draft model and accepting/rejecting based on target model's distribution
- Extension: Applied to diffusion models generating samples through continuous, vector-valued Markov chains
- Simple and effective approach applicable to any diffusion model without training a draft model
- Significant speedup in generation by halving number of function evaluations while ensuring exact sample generation from target model
- Metrics tracked: Wasserstein-2 distance, FID, IS, reward
- Number of Function Evaluations metric reported as calls to target model with batch data
- Low-dimensional experiments with mixture of Gaussians target distribution dimensions ranging from 2 to 32 components
- Two drafting strategies considered: INDEPENDENT and FROZEN
- Analysis of effects of stochasticity ε and window size L on algorithm performance

## Summary

In their paper titled "Accelerated Diffusion Models via Speculative Sampling," authors Valentin De Bortoli, Alexandre Galashov, Arthur Gretton, and Arnaud Doucet introduce a novel technique for accelerating inference in large language models. The speculative sampling method involves generating candidate tokens using a fast draft model and accepting or rejecting them based on the target model's distribution. This approach has been extended to diffusion models that generate samples through continuous, vector-valued Markov chains. The target model in this context is a high-quality but computationally expensive diffusion model. To address this challenge, the authors propose various drafting strategies, including a simple and effective approach that does not require training a draft model and can be applied to any diffusion model out of the box. Their experiments demonstrate significant speedup in generation on various diffusion models by halving the number of function evaluations while ensuring exact sample generation from the target model. The study includes experiments tracking metrics such as Wasserstein-2 distance, FID (Fréchet Inception Distance), IS (Inception Score), and reward in different settings to evaluate the quality of output distributions obtained through speculative sampling. Additionally, the Number of Function Evaluations metric is reported, defining each evaluation as a call to the target model with a batch of data. In low-dimensional experiments investigating Algorithm 3 with key hyperparameters variation in mixture of Gaussians target distribution dimensions ranging from 2 to 32 components, two drafting strategies - INDEPENDENT and FROZEN - are considered. The effects of stochasticity ε in the sampler and window size L on algorithm performance are analyzed. Overall, the proposed speculative sampling technique shows promise for accelerating inference in diffusion models by improving generation speed while maintaining sample accuracy from high-quality but computationally expensive models. Further details on accelerating Langevin diffusions can be found in Appendix K of the paper.