Research map / Image generation / Fast diffusion sampling
Accelerated Diffusion Models via Speculative Sampling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts speculative sampling to diffusion models using reflection maximal coupling for exact sampling, with a frozen-target draft model that needs no extra training, achieving about 2-3x fewer target function evaluations on CIFAR10/LSUN and robot policy generation.
Problem
Diffusion model sampling requires many sequential neural network evaluations; existing speculative sampling works only for discrete token sequences, and continuous-state adjusted rejection sampling is inefficient.
Approach
Replace the adjusted rejection step of speculative sampling with a reflection maximal coupling between Gaussian transition kernels, so rejected states are computed deterministically with one target evaluation; and draft states are obtained from a frozen target model (reusing the drift at the window start) so no draft model needs training.
Key results
- On CIFAR10, the speculative sampler reaches FID 2.34 with 35 target NFE versus 2.45 FID with 100 NFE for standard sampling (about 65% fewer target calls).
- On PushT with L=10, speculative sampling achieves reward 0.903 ± 0.007 using 5.05 target NFE, versus 0.876 ± 0.008 for a 5-step target run.
- The method reports around halving the NFE on LSUN while keeping quality of the target model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add speculative sampling (ε=0.25, τ=2.0) to the target sampler | CIFAR-10 target diffusion model · CIFAR-10 | FID: 2.45 -> 2.34 | Helped |
| add speculative sampling (ε=0.25, τ=2.0) to the target sampler | CIFAR-10 target diffusion model · CIFAR-10 | NFE: 100 -> 35.40 | Helped |
| reduce target sampling to 30 steps | CIFAR-10 target diffusion model · CIFAR-10 | FID: 2.45 -> 4.32 | Hurt |
| add speculative sampling (L=10, K=10) | PushT target diffusion policy · PushT | NFE: 10 -> 5.053 ± 0.001 | Helped |
| add speculative sampling (L=10, K=10) | PushT target diffusion policy · PushT | Reward: 0.901 ± 0.008 -> 0.903 ± 0.007 | No clear effect |
| add speculative sampling (L=5, K=5) | PushT target diffusion policy · PushT | Reward: 0.876 ± 0.008 -> 0.870 ± 0.009 | No clear effect |
| increase τ from 0.5 to 2.0 (ablation) | CIFAR-10 speculative sampler (ε=0.25) · CIFAR-10 | NFE: 42.44 -> 35.40 | Helped |
| increase τ from 0.5 to 2.0 (ablation) | CIFAR-10 speculative sampler (ε=0.25) · CIFAR-10 | FID: 2.42 -> 2.34 | Helped |
| increase τ from 1.0 to 2.0 (ablation) | CIFAR-10 speculative sampler (ε=0.25) · CIFAR-10 | NFE: 39.31 -> 35.40 | Helped |
| use ε=0.01 instead of ε=0.25 (τ=2.0) (ablation) | CIFAR-10 speculative sampler · CIFAR-10 | FID: 2.34 -> 2.83 | Hurt |
| use ε=0.01 instead of ε=0.25 (τ=2.0) (ablation) | CIFAR-10 speculative sampler · CIFAR-10 | NFE: 35.40 -> 57.47 | Hurt |
Surprising findings
- Naive adjusted rejection sampling for continuous diffusion costs on average as many target samples as direct sampling, so the reflection coupling is needed.
- Acceptance probability is very low for Gaussians with different covariances in high dimension, motivating the same-covariance design.
Limitations
- Not directly applicable to deterministic samplers without adding noise.
- Increases memory overhead due to parallel calls during verification.
- Diffusion models already benefit from parallelism, so speedups in serving are less clear than for memory-bound LLMs.
Experimental setup
Datasets: CIFAR10 LSUN PushT
Benchmarks: CIFAR10 LSUN PushT
Metrics: FID IS NFE reward Wasserstein-2
Baselines: standard diffusion sampling with 100 steps standard diffusion sampling with 30 steps ParaDiGMS Masked autoregressive speculative sampling with rejection sampling
Builds on: speculative sampling maximal coupling reflection maximal coupling ParaDiGMS Medusa
Closest papers by meaning
- Accelerated Speculative Sampling Based on Tree Monte Carlo (ICML 2024 · similarity 0.92)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- Fast Sampling via Discrete Non-Markov Diffusion Models with Predetermined Transition Time (NeurIPS 2024 · similarity 0.90)
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.89)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.89)
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.