Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, yet they require simulating a Markov chain for many steps to produce a sample. To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from. We empirically demonstrate that DDIMs can produce high quality samples $10 \times$ to $50 \times$ faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.
核心贡献 · Key contributions
提出去噪扩散隐式模型（DDIM）以实现更快的采样。 Proposes denoising diffusion implicit models (DDIMs) for faster sampling.
构建了与 DDPM 具有相同训练目标的非马尔可夫扩散过程。 Constructs non-Markovian diffusion processes with same training objective as DDPMs.
在高质量图像生成上实现 10 倍到 50 倍的加速。 Achieves 10× to 50× speedup in wall-clock time for high-quality image generation.
通过调整步数实现计算与样本质量之间的权衡。 Enables trade-off between computation and sample quality via step count adjustment.
展示了直接在潜在空间中进行语义有意义的图像插值。 Demonstrates semantically meaningful image interpolation directly in latent space.
展示了确定性生成过程允许对观测进行编码和重建。 Shows deterministic generation process allows encoding and reconstruction of observations.
局限 · Limitations
与全步数 DDPM 相比，极少步数下样本质量可能下降。 Sample quality may degrade with very few steps compared to full-step DDPM.
确定性过程可能比随机 DDPM 降低多样性。 Deterministic process may reduce diversity compared to stochastic DDPM.
编码-解码中步数越少重建误差越大。 Reconstruction error increases with fewer steps in encoding-decoding.
方法假设与 DDPM 相同的训练目标，限制了架构变化。 Method assumes same training objective as DDPM, limiting architectural changes.
实证验证仅限于图像生成任务。 Empirical validation limited to image generation tasks.
论文章节 · Sections（共 18）
摘要Abstract
1 引言1 Introduction
2 背景2 Background
3 非马尔可夫前向过程的变分推断3 Variational Inference for non-Markovian Forward Processes
3.1 非马尔可夫前向过程3.1 Non-Markovian forward processes
3.2 生成过程与统一变分推断目标3.2 Generative process and unified variational inference objective
4 从广义生成过程中采样4 Sampling from Generalized Generative Processes