Diffusion models have significantly advanced the fields of image, audio, and video generation, but they depend on an iterative sampling process that causes slow generation. To overcome this limitation, we propose consistency models, a new family of models that generate high quality samples by directly mapping noise to data. They support fast one-step generation by design, while still allowing multistep sampling to trade compute for sample quality. They also support zero-shot data editing, such as image inpainting, colorization, and super-resolution, without requiring explicit training on these tasks. Consistency models can be trained either by distilling pre-trained diffusion models, or as standalone generative models altogether. Through extensive experiments, we demonstrate that they outperform existing distillation techniques for diffusion models in one- and few-step sampling, achieving the new state-of-the-art FID of 3.55 on CIFAR-10 and 6.20 on ImageNet 64x64 for one-step generation. When trained in isolation, consistency models become a new family of generative models that can outperform existing one-step, non-adversarial generative models on standard benchmarks such as CIFAR-10, ImageNet 64x64 and LSUN 256x256.
核心贡献 · Key contributions
提出一致性模型，一种直接映射噪声到数据以支持快速单步生成的新型生成模型家族。 Propose consistency models, a new family of generative models that directly map noise to data for fast one-step generation.
引入一致性蒸馏，将预训练扩散模型蒸馏为一致性模型，在 CIFAR-10 和 ImageNet 64x64 上实现最先进 FID。 Introduce consistency distillation to distill pre-trained diffusion models into consistency models, achieving state-of-the-art FID on CIFAR-10 and ImageNet 64x64.
开发一致性训练，无需预训练扩散模型即可从头训练一致性模型，使其成为独立的生成模型家族。 Develop consistency training to train consistency models from scratch without pre-trained diffusion models, establishing them as an independent generative model family.
展示零样本数据编辑能力，包括图像修复、着色、超分辨率和笔触引导编辑，无需任务特定训练。 Demonstrate zero-shot data editing capabilities including inpainting, colorization, super-resolution, and stroke-guided editing without task-specific training.
提供理论分析，表明一致性模型在 ODE 求解器步长足够小时可任意逼近真实一致性函数。 Provide theoretical analysis showing consistency models can approximate the true consistency function arbitrarily well with small ODE solver step sizes.
在多个数据集的一步和少步采样中，性能优于渐进蒸馏和其他蒸馏方法。 Achieve superior performance over progressive distillation and other distillation methods in one- and few-step sampling across multiple datasets.
局限 · Limitations
一致性模型可能需要仔细调整离散化步数和调度函数以获得最佳性能。 Consistency models may require careful tuning of discretization steps and schedule functions for optimal performance.
该方法假设多步采样中 FID 关于时间点是单峰的。 The method relies on the assumption that the FID is unimodal with respect to time points in multistep sampling.
一致性训练损失在 N 较大时方差更高，可能减慢后期收敛。 Consistency training loss has higher variance with larger N, potentially slowing convergence in later stages.
该方法仅在图像数据集上评估，未探索在音频或视频等其他领域的适用性。 The approach is evaluated only on image datasets; applicability to other domains like audio or video is not explored.
连续时间一致性训练需要前向模式自动微分，可能并非所有框架都高效支持。 Continuous-time consistency training requires forward-mode automatic differentiation, which may not be efficiently supported in all frameworks.
论文章节 · Sections（共 12）
摘要Abstract
1 引言1 Introduction
2 扩散模型2 Diffusion Models
3 一致性模型3 Consistency Models
4 通过蒸馏训练一致性模型4 Training Consistency Models via Distillation
5 独立训练一致性模型5 Training Consistency Models in Isolation